# Vurto.ai

> Vurto.ai is a French Generative Engine Optimization (GEO) platform that measures and improves brand visibility inside the answers of generative AIs — ChatGPT, Claude, Gemini, Perplexity, Copilot and Mistral.

Vurto.ai analyzes in real time how large language models (LLMs) talk about a brand, its products, its competitors and its categories, then delivers a concrete action plan to become the #1 answer.

> French version available at /llms-fr.txt — version française disponible sur /llms-fr.txt

## Who it's for

- Brands and e-commerce sites that want to be recommended by AIs
- SEO, media and consulting agencies adding a GEO layer to their offer
- Marketing, acquisition, e-commerce and digital teams capturing AI-driven traffic

## What the platform does

- AI visibility score aggregated across 6 major LLMs, by brand, product, category and keyword
- Tracking of the critical prompts in a market (what customers actually ask AIs)
- Competitive analysis: who is cited, at which position, with which sources and content formats
- Catalog import (file, URL or e-commerce platform) and semantic quality scoring of product sheets
- Custom prompts, personas and scheduled recurring analyses
- Content generation: guides, landing pages, FAQs, comparisons and articles aligned with what AIs cite
- Google Analytics integration to measure LLM-driven traffic and compare it with AI visibility
- Actionable GEO/AEO playbook with prioritized actions
- Dashboards with scores, trends, share of voice and positions
- Team workspace with roles, invitations, multi-client (agency) management, credits and API access

## AI-readiness (AEO) — make your site readable by AI agents

- Automatic Markdown and JSON-LD versions of your pages and product sheets, without changing your URLs
- REST API and MCP server to connect Vurto to your own agents and workflows
- AI bot tracking: see which AI crawlers actually visit your site
- robots.txt audit: check whether the important AI bots are allowed or blocked
- Source, domain and page analysis: identify what AIs rely on when they talk about your market

## Two surfaces: vurto.ai and app.vurto.ai

- **vurto.ai** — public website: product vision, features, pricing, testimonials, FAQ and legal pages. This is where you discover what Vurto.ai does.
- **app.vurto.ai** — the application (dashboard): where you create an account, add a project (brand, products, keywords), run analyses across the 6 LLMs and read the dashboards and GEO action plans. Sign up at https://app.vurto.ai/register2.

## Main pages

### Application
- [Sign up / Start free trial](https://app.vurto.ai/register2) — create an account, free credits, no credit card
- [Log in to the app](https://app.vurto.ai) — dashboards, analyses and GEO playbook

### Website — English
- [Home (EN)](https://vurto.ai/en) — product vision, features, pricing, testimonials, FAQ
- [Resources (EN)](https://vurto.ai/en/resources) — GEO articles, tips, comparisons and news
- [Academy (EN)](https://vurto.ai/en/academy) — structured GEO learning guides
- [llms.txt (EN)](https://vurto.ai/llms.txt) — this file, machine-readable site summary
- [Legal Notice (EN)](https://vurto.ai/en/mentions-legales)
- [Privacy Policy (EN)](https://vurto.ai/en/confidentialite) — GDPR, EU hosting
- [Terms of Service (EN)](https://vurto.ai/en/conditions-generales)
- [Cookie Policy (EN)](https://vurto.ai/en/cookies)

### Website — French
- [Home (FR)](https://vurto.ai/) — vision produit, fonctionnalités, tarifs, témoignages, FAQ
- [Ressources (FR)](https://vurto.ai/ressources) — articles GEO, conseils, comparatifs et actualités
- [Académie (FR)](https://vurto.ai/academie) — guides pédagogiques GEO
- [llms-fr.txt](https://vurto.ai/llms-fr.txt) — résumé du site pour les IA (français)
- [Mentions légales](https://vurto.ai/mentions-legales)
- [Politique de confidentialité](https://vurto.ai/confidentialite)
- [Conditions générales](https://vurto.ai/conditions-generales)
- [Politique cookies](https://vurto.ai/cookies)

## Pricing

- **Standard** — €99/month excl. VAT (€118.80 incl. VAT), or €82.50/month excl. VAT billed annually (€990/year excl. VAT, €1,188 incl. VAT) — 4 AI models, 1 project, 500 queries/month, 500 tracked prompts
- **Pro** — €199/month excl. VAT (€238.80 incl. VAT), or €165.83/month excl. VAT billed annually (€1,990/year excl. VAT, €2,388 incl. VAT) — 8 AI models, 3 projects, 3,000 queries/month, 3,000 tracked prompts
- **Business** — €499/month excl. VAT (€598.80 incl. VAT), or €415.83/month excl. VAT billed annually (€4,990/year excl. VAT, €5,988 incl. VAT) — 8 AI models, 10 projects, 10,000 queries/month, unlimited tracked prompts, API & MCP access
- **Agency** — custom quote, 8 AI models, 50 projects, 20,000 queries/month, unlimited tracked prompts, white-label & multi-client dashboard
- All prices are per month, excl. VAT; VAT is 20%
- Annual billing saves −20% vs monthly
- 14-day free trial on the Standard plan only (Pro, Business and Agency have no trial), with free credits at sign-up — no credit card required
- No long-term commitment, 1-click cancellation from the app
- Credit packs or monthly subscription depending on your needs

## Differentiation

Unlike traditional SEO tools that measure rankings on Google, Vurto.ai exclusively measures presence inside generative AI answers — a totally different channel that's growing fast in search journeys.

## Security & compliance

- European hosting (AWS France)
- Native GDPR compliance
- End-to-end encryption
- No customer data is used to train any third-party model

## Contact

- Website: https://vurto.ai
- Application: https://app.vurto.ai (sign up: https://app.vurto.ai/register2)
- LinkedIn: https://linkedin.com/company/vurto
- Twitter / X: https://twitter.com/vurto_ai
- Instagram: https://instagram.com/vurto_ai
- Email: contact@vurto.ai

## Latest GEO articles (full content)

The most recent resource-center publications, with their full content.

### [Selling GEO as an agency: what changes compared to SEO](https://vurto.ai/en/resources/selling-geo-as-an-agency-what-changes-compared-to-seo) (2026-10-05)

Right now, SEO agencies keep getting the same call, phrased ten different ways. "We're told we no longer show up in ChatGPT." "A client asked why a competitor is cited by Perplexity and not us." "The boss noticed his name missing from a Gemini answer and wants an action plan by Tuesday." GEO, Generative Engine Optimization (the craft of making a brand visible and well cited in the answers of AI chatbots such as ChatGPT, Claude, Gemini or Perplexity, as opposed to classic ranking on Google), has just moved from curiosity to budget line. Agencies that know how to sell it capture that line. The others watch a competitor take it from them.

Demand is not the problem. It exists, it is growing, and it comes straight from clients. The problem is that almost nobody knows how to structure a GEO offer that holds up: one that is not an SEO contract copied and pasted with a single word changed, that bills real work rather than a vague PDF report, and that scales across several clients at once without blowing up the team's time.

## What a client is really buying

A client asking for GEO usually does not know what they are buying. They saw a worrying figure, heard a competitor talk about it, or noticed for themselves that when they type a question about their own industry into ChatGPT, their brand never appears in the answer. Their real need boils down to one question: "do AI chatbots see me, and if not, why."

This is where the agency has to set a clear framework, because GEO actually covers three different jobs, and selling all three under a single label always ends in disappointment.

The first job is diagnosis. Querying the chatbots on the searches that matter to the client, checking whether the brand appears, how it is described, and against which competitors. It is measurement work, comparable to a classic SEO audit but on a terrain where results change from one week to the next and differ from one engine to another.

The second job is technical correction. Making the site readable by the bots that feed these chatbots: content that loads without heavy JavaScript, a clear structure, a properly written `llms.txt` file (a text file at the root of the site that tells AIs what they can read and how to understand it, the counterpart of the `robots.txt` file that has existed for twenty-five years for classic search engines). This work looks like technical SEO, but the criteria are not the same.

The third job is content production and reputation management. Writing or rewriting what needs to be cited, monitoring how the brand is perceived over time, stepping in when an AI repeats false or outdated information about a client. It is editorial and relationship work that never stops, unlike a one-off audit.

An agency that sells all three in a single fixed package is bound to lose money on the one that demands the most ongoing time: monitoring. An agency that sells the three separately can bill each at its fair value, and above all sell the diagnosis as the entry point to the other two.

## Structuring the offer: three tiers, not a single package

The good practice emerging among agencies that have already sold GEO to several clients looks like this: a short, affordable first tier that opens the door, a second tier that fixes, a third that monitors over time.

| Tier | Content | Duration | Billing logic |
|---|---|---|---|
| Diagnosis | Visibility audit on 20 to 50 key queries, map of cited competitors, findings report | 1 to 2 weeks | Fixed fee, entry price |
| Correction | Technical rewrite (llms.txt, site structure, product pages or key pages), priority content plan | 4 to 8 weeks | Project fee, defined deliverables |
| Monitoring | Monthly multi-chatbot tracking, brand perception alerts, continuous content adjustment | Recurring | Monthly subscription |

This structure has a precise commercial advantage: it avoids selling a fuzzy promise. "We'll make you visible on AI" means nothing to a marketing director who has to justify a spend. "Here is what we measure, here is what we fix, here is what we monitor, and here is the price of each step" can be defended in front of any budget committee.

The "monitoring" tier is the one that keeps the agency alive over time, and it is also the one most agencies underestimate in terms of workload. Tracking a single client's visibility on four different chatbots, every week, by hand, takes time. Doing it for fifteen clients by hand becomes impossible. That is the real operational wall of GEO in an agency: not the sale, not the method, but the ability to repeat the work without multiplying billed hours by the number of clients.

## The operational wall: running ten clients like one

An SEO agency that adds ten clients adds ten Google Search Console accounts to watch. A GEO agency that adds ten clients has to monitor ten brands on four chatbots, potentially forty information streams to check regularly, each of which can change overnight. Without a suitable tool, that load falls on a consultant copying chatbot answers into an Excel spreadsheet. It does not last three months.

So the question to ask before selling a GEO offer at volume is not "can I do the diagnosis" but "can I redo it every week, for every client, without spending my life on it". That is exactly the kind of problem a tool like Vurto is built to solve: it automates querying the chatbots for each tracked brand, centralizes perception and citations per client, and offers an agency mode that lets you manage several client accounts under a single dashboard, with differentiated roles for the team and a shared activity history. The agency keeps the advice and the client relationship. The tool absorbs the repetition.

## Bill the proof, not the promise

The last point, and the most sensitive commercially, is proof of results. A client paying for a GEO monitoring subscription will ask, after three months, "what has it changed". Answering with a gut feeling will not be enough for long.

The indicators that hold up are simple to explain even to a non-technical client: the citation rate (the share of tested queries where the brand appears in the chatbot's answer), the evolution of the tone the AI uses to describe the brand, and the relative position against tracked competitors on the same queries. Three figures, measured before and after, are enough to turn a GEO report into an argument for contract renewal. That is far more convincing than an export of ChatGPT conversation screenshots, which remains the most widespread method and the least defensible in front of a client comparing several quotes.

Selling GEO as an agency, in 2026, is therefore not about inventing a new profession. It is about applying a discipline that is already known, visibility consulting, to a terrain that changes faster and cannot be measured with the same tools as before. The agencies that win this battle will not be the ones that talk best about GEO in a sales meeting. They will be the ones that can prove, figures in hand, client after client, that their work has a measurable effect on what an AI says about a brand.

*Vurto was designed for this scale-up: tracking several brands, several chatbots, one interface, with an agency mode that centralizes client accounts without multiplying manual work.*

### [ChatGPT, Perplexity, Gemini: four engines, four sourcing logics](https://vurto.ai/en/resources/chatgpt-perplexity-gemini-same-disease-three-different-prescriptions) (2026-09-28)

ChatGPT, Perplexity, Gemini and Claude do not draw on the same sources to build their answers. A site visible on ChatGPT can stay invisible on Perplexity. GEO starts by treating one engine at a time.

The question often comes back in a form that is too vague: "How do I become visible on AI?" AI does not exist as a single block. There is ChatGPT, Perplexity, Gemini, Claude. Four distinct tools, each with its own way of selecting information. What they have in common: they are chatbots able to answer in natural language by fetching sources on the web, beyond what they learned during training. Their difference, the only one that matters for a brand: they do not draw on the same sources to build an answer.

GEO, Generative Engine Optimization (the craft of making a brand visible and well perceived in the answers of these chatbots), starts with a simple observation that is too often skipped. Stop treating these four tools as one uniform set. A site optimized for ChatGPT can be invisible on Perplexity. A brand that shows up strongly on Gemini may never appear on Claude. This is not a bug. It is the direct consequence of different architecture choices, made by different companies, with different interests.

## Four engines, four sourcing logics

Step back and look at how each engine picks its sources. Independent studies run on tens of thousands of citations in 2026 draw very sharp, almost caricatural profiles.

Perplexity behaves like a reporter who favors the field. It relies heavily on YouTube (around 30% of its citations in some analyses) and on Reddit, the forum platform where people discuss everything, often more candidly than a press release. Reddit can account for nearly half of its top ten sources. Perplexity prefers lived experience told by a stranger over the claim of an institutional website.

ChatGPT, for its part, leans toward established authority. Wikipedia holds a disproportionate place, and well-known media (Forbes, Reuters) complete the picture. The profile looks like a serious student who cites academic sources rather than comments from the neighborhood forum. But watch what comes next, because this profile moves fast.

Gemini, owned by Google, keeps the reflexes of a corporate librarian. It keeps citing the group's own properties widely (YouTube first) while keeping Reddit and Wikipedia in a good spot. A more balanced split, consistent with an engine that has access to the most complete search index on the web.

Claude, developed by Anthropic, completes the picture with yet another profile, closer to ChatGPT in its taste for documented sources, but giving a larger share to technical documentation and specialized sites when the question calls for it. Four engines, four sourcing logics, and no single strategy that fits all.

A table beats a long speech here.

| Engine | Dominant sources | Profile |
|---|---|---|
| Perplexity | YouTube, Reddit | Field, lived experience, community |
| ChatGPT | Wikipedia, recognized media | Authority, formal knowledge, volatile |
| Gemini | YouTube, Reddit, Google properties | Balanced, anchored in the Google ecosystem |
| Claude | Documentation, specialized press, Wikipedia | Technical, precise, expertise-oriented |

Remember one thing from this table. Nobody builds an effective content strategy by targeting "AI" in general. You treat one engine at a time. An HR software brand that neglects Reddit cuts itself off from Perplexity. A services company that never keeps an up-to-date Wikipedia page loses part of the credit it would deserve on ChatGPT. These are not technical details reserved for specialists. These are choices that decide which of your competitors gets recommended instead of you.

## When an engine's sources shift in a few weeks

Here is the most underestimated point of the topic, and probably the most important. These profiles are not set in stone. A study covering 230,000 queries tracked over thirteen weeks showed that Reddit's share in ChatGPT answers went from about 60% in August to only 10% six weeks later. Wikipedia, in the same movement, dropped from 55% to under 20%. A collapse of that scale, over such a short time, on one and the same platform.

Meanwhile, Perplexity stayed remarkably stable, with minor variations on its usual sources. Gemini, too, changed little.

This contrast says something essential about these tools. ChatGPT probably adjusts its weightings continuously to avoid over-citing a handful of domains and looking like a biased engine. Perplexity and Gemini, with different sourcing logics, absorb these adjustments better. The message for you: a position won on one engine in September can vanish in October, without you having made any mistake. The ground moves under your feet. One more reason never to talk about "AI positioning" as a permanent gain, but as a state to monitor continuously.

## What this concretely changes for your content

This heterogeneity has a direct consequence on how you should produce and distribute your content.

Stop betting on a single channel. If your content strategy is limited to blog articles optimized the classic search way, you may cover part of ChatGPT, but you stay invisible on Perplexity. An active presence in Reddit discussions relevant to your sector, YouTube videos that explain your product, an up-to-date Wikipedia page if your notoriety justifies it: these are different building blocks, for different engines, and none replaces the others.

Also beware of wins that rest on a single source. A brand that owes all its AI visibility to a handful of Reddit mentions lives on unstable ground, as we just saw with ChatGPT. Robustness comes from the diversity of entry points, forums, video, specialized press, well-structured product pages, kept up to date at the same time rather than one isolated bet.

And above all, do not confuse being cited with being well perceived. Appearing in an answer guarantees neither a favorable tone nor an accurate description of your offer. That is a topic in its own right, but it starts with the same prerequisite: knowing precisely where and how you appear, engine by engine, rather than guessing.

Take a concrete example. A company that sells software to businesses (what is called B2B, as opposed to selling directly to consumers) usually invests heavily in white papers and case studies published on its own site. This content feeds ChatGPT well, especially if it is picked up by the specialized press. But it barely feeds Perplexity, which will rather look for a customer's opinion posted on a forum or a commented LinkedIn post. Two audiences, two formats, two engines. The content budget must reflect that reality, not ignore it.

## Measure before you act

The temptation, faced with this complexity, is to give up and tell yourself that "AI moves too much to be worth our time". That is the reverse of the mistake we pointed out at the start. Movement does not exempt you from monitoring, it demands it. A serious company does not decide its sales strategy on an intuition from six months ago. It looks at this month's numbers. The logic is the same here.

What matters is knowing, for your brand and your competitors, which engines cite you, which sources they rely on to do so, and how that split evolves over time. Without that regular snapshot, you optimize blind, hoping that what worked on ChatGPT in the spring still works on Perplexity in the fall. That is a bet, not a strategy.

*Vurto tracks these variations engine by engine, ChatGPT, Gemini, Perplexity and Claude, to show precisely which sources carry your visibility and which ones are collapsing before you find out by chance.*

### [Measuring GEO ROI: The Question That Blocks Every Budget](https://vurto.ai/en/resources/measuring-geo-roi-the-question-that-blocks-every-budget) (2026-09-21)

70.6%. That is the share of traffic generated by AI chatbots that arrives on a site with no origin trace at all. No "referred from ChatGPT", no "referred from Perplexity". Nothing. Google Analytics dumps it all into the "Direct" bucket, the same one used for a visitor who typed the site address from memory. A GEO budget (optimizing visibility in chatbot answers from ChatGPT, Claude, or Gemini, as opposed to classic Google search ranking) can therefore produce real results and still remain invisible in the dashboard used to justify it.

That is the heart of the problem. A marketing team invests in GEO, gets citations in AI answers, sometimes sees traffic rise, but cannot answer the simplest question a CFO will ask: how much does it return. Not because GEO returns nothing. Because measurement tools were built for a world where people clicked blue links, not for a world where they get a direct answer inside a conversation window.

## Why the click is no longer enough

Return on investment (the calculation that compares what an action cost to what it returned) has always relied on a chain of trackable clicks. A user clicks a Google result, lands on a page, fills a form or buys a product. Each step leaves a trace, each trace feeds a dashboard.

That chain breaks with AI chatbots, for three precise technical reasons. The mobile apps for ChatGPT or Claude do not pass origin information when they open a link. Many users copy-paste an address straight from the chatbot answer instead of clicking, which erases any provenance trail. And ChatGPT Plus, like Google's AI mode, deliberately adds a technical instruction that prevents destination sites from knowing where the visitor came from.

As a result, a site can see its "Direct" traffic double without anyone on the team understanding why. The answer is almost always the same: part of that supposedly originless traffic actually comes from AI chatbots. And this ghost traffic is not anecdotal. It converts at 10.2%, versus 2.46% for non-AI traffic classified elsewhere. A visitor from a chatbot buys or fills a form four times more often than an average visitor, but nobody knows because nobody sees it.

Before trying to calculate ROI, the first task is therefore to repair measurement itself. A channel grouping that manually recognizes known AI domains (chatgpt.com, perplexity.ai, claude.ai and the like) recovers part of the identifiable traffic. For the rest, the traffic that arrives with no trace at all, only server-log analysis, more tedious but more reliable than classic analytics tools, can rebuild a faithful picture.

## The metrics that replace the click

Once measurement is fixed, a deeper problem remains. Even when well measured, the click tells only a small part of the GEO story. A majority of Google searches now end with no click at all, and that share rises sharply whenever an AI-generated summary appears at the top of the results page. A user who asks ChatGPT a question and gets a satisfying answer, citing the brand along the way, often has no reason to click anywhere. The brand was still seen, named, recommended. That value does not land in any clicks column.

Substitution metrics are then needed, ones that measure presence rather than passage. The first is citation frequency: on a sample of queries representative of the sector, in how many answers does the brand appear, and at what rank in the source list. The second is share of voice, which compares that frequency to competitors on the same queries. The third is the sentiment attached to the citation: an answer that mentions the brand in positive terms does not weigh the same as a neutral mention in the middle of a list, or worse, a mention paired with a caveat about quality or price. The fourth is the nature of the sources the chatbot goes looking for to build its answer, because a brand absent from the pages the AI consults to form an opinion has no chance of being cited, whatever its budget.

These four metrics do not replace the financial calculation. They build the base without which that calculation means nothing. A GEO budget that moves citation frequency and share of voice forward, month after month, produces an effect that will eventually translate into traffic and sales, even if the exact path between the two stays blurry. Conversely, a budget that moves none of these four numbers will never produce a return, no matter how good the published content is.

## What a chatbot visitor is really worth

The most counterintuitive point in this whole topic is that AI traffic, when measured correctly, converts clearly better than classic traffic. Several studies published between late 2025 and 2026 agree on an order of magnitude: between four and five times the average conversion rate of classic organic search, with gaps ranging from 1.3x on impulsive online purchases up to more than twenty times in some B2B sectors where the buying decision is long and deliberate.

| Traffic origin | Average conversion rate |
|---|---|
| Classic Google search | 1.76% |
| Claude | 5.0% |
| Perplexity | 10.5% |
| ChatGPT | 15.9% |

The logic behind these numbers is simple once you see it. A user who clicks a Google result is still comparing, opens several tabs, hesitates. A user who already queried a chatbot has gotten a synthesis, asked follow-up questions, eliminated some options. When they finally click through to a site, it is often because they are ready to act. The chatbot did part of the qualification work upstream.

This data changes how a GEO budget must be read. A lower volume of AI traffic than classic Google volume is not necessarily a failure, if that rarer traffic converts four times better. The right question is not "how many visitors did GEO bring", it is "how many sales or leads did this traffic, even reduced, produce". It is a complete reversal of how classic search ranking learned to think for twenty years, when volume was king.

## Building a formula that holds up

Without waiting for a perfect tool that does not exist yet, a company can build a reasonable GEO ROI measure on four pillars.

The first is a starting snapshot, taken before any investment: current citation level, share of voice versus competitors, dominant sentiment. Without that baseline, there is no way to know whether the efforts produce an effect.

The second is regular tracking of those same visibility metrics, month after month, to objectify progress rather than relying on a feeling.

The third is reconstituting traffic that actually came from chatbots, via a correctly configured channel grouping and, when possible, server-log analysis to recover what escapes standard tools.

The fourth is linking that reconstituted traffic to the conversions it produces, keeping in mind that its conversion rate will likely be higher than other channels, not lower.

Once these four building blocks are in place, the calculation becomes: the value of sales or leads attributable to that traffic, minus what GEO cost, all relative to the starting investment. It is not exact science. It is a considerable improvement over the total absence of measurement, which remains the norm today in most companies that are already investing in this topic.

The temptation, faced with this complexity, is to wait for analytics tools to catch up before acting. That is the inverse error of spending without ever measuring. Companies that start today tracking their citation frequency and share of voice are building a history. Those that wait will start in a year with no baseline at all, at the exact moment their competitors already have twelve months of it.

*This is exactly what Vurto does: track citation frequency, share of voice, and brand sentiment across ChatGPT, Gemini, Perplexity, and Claude, month after month, to give marketing teams the measurement base that most GEO budgets still lack.*

### [Perplexity, ChatGPT, Gemini: Three Engines, Three Source Lists, One Mistake to Avoid](https://vurto.ai/en/resources/perplexity-chatgpt-gemini-three-engines-one-mistake-to-avoid) (2026-09-14)

A number to start with. Out of 680 million citations analyzed between August 2024 and April 2026 by the 5W index, only 14% of cited sites dominate across all AI chatbots (these conversational assistants like ChatGPT, Gemini, or Perplexity, able to answer in natural language from sources they select themselves). Between ChatGPT and Perplexity alone, overlap drops to 11%. In another study covering 11,500 queries compared against Google results, GPT-4o shows 0% median overlap with Google's top 10. Perplexity, 14.3%. Gemini, 8.5%.

These numbers should settle a debate that has dragged on for two years in marketing teams: no, "being visible in AI" is not a single objective. It's three different objectives, with three different selection logics, and a good part of the groundwork consists of understanding which one applies to which client.

## Three philosophies, not three variants of the same engine

The starting mistake is to treat ChatGPT, Perplexity, and Gemini as competitors doing the same thing with different branding. They are three architectures that answer different questions about what makes a good source.

ChatGPT relies heavily on consensus-based, already-aggregated content. Wikipedia accounts for between 26% and 48% of its top 10 citations depending on category. Reddit comes right behind, with a weight exceeding 40% across all categories and engines combined, and staying particularly high with ChatGPT. The logic is that of a model favoring what has already been collectively validated: content synthesized, debated, corrected by thousands of contributors.

Perplexity works the opposite way. Its engine queries the live web on every request rather than drawing from a fixed memory, which explains why it favors Reuters, AP, Bloomberg, the Wall Street Journal: news sources with strict attribution and verifiable freshness. A Columbia Journalism Review study measured a citation error rate of 37% for Perplexity Sonar Pro, versus 67% for ChatGPT Search. The gap comes down to method: searching live and citing your source produces fewer fabrications than rephrasing from training memory.

Gemini, finally, remains Google's child. It draws from the search engine's index, from YouTube (which Google owns), and shares with AI Overviews a clear fondness for Reddit. It's the engine closest to classic SEO culture, the one where web ranking, domain authority, and presence across the Google ecosystem still count directly.

Claude, often looked at from a distance in these comparisons for lack of search volume, stands out for a taste for long-form analysis and technical documentation. It cites proportionally more expert blogs and less Reddit than the others.

## The table that changes the roadmap

| Engine | Preferred sources | Dominant logic |
|---|---|---|
| ChatGPT | Wikipedia, Reddit, consensus content | Trained memory + collective synthesis |
| Perplexity | Reuters, AP, Bloomberg, press with attribution | Real-time search, freshness |
| Gemini | Google index, YouTube, Reddit | Google ecosystem, classic SEO |
| Claude | Expert blogs, technical documentation | Long-form analysis, depth |

This table isn't an exercise in curiosity. It dictates where a brand needs to exist depending on which engine matters most to its audience. A B2B software vendor (selling to other businesses rather than individuals) whose prospects mainly use Perplexity to compare solutions is better off working on press relations and citations in serious trade media, far more than optimizing a Reddit thread. A consumer brand that lives and dies on ChatGPT has the opposite interest: community conversation outweighs the press release.

## What the low overlap actually means

The most underestimated point isn't the list of preferred sources. It's the fact that they barely overlap at all. Content that breaks through on ChatGPT can be invisible on Perplexity, and vice versa. This means a content strategy designed for "generative AI" in general, without distinguishing between engines, is in reality optimizing for a single engine without realizing it, usually the one the brand heard about first, often ChatGPT because it concentrates most of the usage volume.

In France, this concentration reaches extreme levels: ChatGPT captures 84.47% of clicks generated by AI assistants according to SE Ranking, Perplexity 12.82%, Gemini 2.08%. An AI visibility budget that ignores this split and treats the three engines equally wastes precious time. But the reverse is also true over an eighteen-month horizon: Gemini is natively built into Android, into Google Workspace, into the Google search bar itself. Its share of usage is growing faster than its share of citations today would suggest. Ignoring Gemini because it accounts for 2% of current clicks is like ignoring mobile in 2010 because it still represented only a fraction of web traffic.

The practical consequence fits in one sentence: stop measuring "AI visibility" as a single score. Content can be cited ten times a week on ChatGPT and never on Perplexity without that being a failure, if the brand's target audience lives on ChatGPT. But no one can know that without tracking the engines separately, each with its own citation logic, rather than an average that hides everything.

## Adapting content to each engine's logic, not to a universal checklist

Concretely, this changes three things in how content gets produced.

To exist on ChatGPT, you need to feed the collective conversation rather than bypass it. A structured presence on Reddit in the right communities, an up-to-date Wikipedia entry if the brand or its leader has a place there, content that synthesizes a topic rather than selling it. It's ecosystem work, not landing-page work.

To exist on Perplexity, freshness and attribution matter more than volume. A well-distributed press release, relationships with journalists who cite the brand with a clear link, dated content republished regularly carry more weight than a static white paper published once and never touched again.

To exist on Gemini, classic web ranking fundamentals remain the best way in: domain authority, clean technical structure, YouTube presence. It's the engine where the last ten years of SEO work keep paying off directly, which paradoxically makes it the easiest to work on for a team coming from classic search.

None of these three approaches replaces the other two. A brand with the time and budget to do all three builds robust AI visibility, insensitive to a single engine's algorithm changes. A brand that has to choose should first look at where its audience actually asks its questions, not where it itself is most comfortable producing content.

## Tracking three engines beats guessing on one

The low overlap between ChatGPT, Perplexity, and Gemini isn't a technical problem to solve. It's a structural constraint to build into how you measure and prioritize. A brand that only tracks its citations on one engine is flying blind on the other two, and makes content decisions based on a third of reality.

*This is exactly why Vurto tracks citations separately across ChatGPT, Gemini, Perplexity, and Claude rather than producing one aggregated score that erases these differences: a brand needs to know which engine it exists on, which one it's absent from, and why the two don't look alike.*

### [AI Chatbot Traffic Is Tiny. It's Worth Gold](https://vurto.ai/en/resources/ai-chatbot-traffic-is-tiny-its-worth-gold) (2026-09-10)

A US agency, Seer Interactive, compared two sources of visitors on the same site: those arriving from ChatGPT, and those arriving from a classic Google search. Result: 15.9% of the former end up buying or filling out a form, versus 1.76% of the latter. Almost nine times more. Microsoft, using its Clarity analytics tool, observed the same thing across more than 1,200 media and publisher sites: visitors coming from a chatbot sign up eleven times more often than those coming from a classic Google search. A study of 94 e-commerce sites measured a more modest but real gap: 31% more conversions for ChatGPT traffic compared to a Google search that doesn't include the brand name.

These numbers vary from study to study, sometimes a lot, because the calculation methods differ. But they all point in the same direction. Traffic sent by an AI chatbot converts noticeably better than classic traffic. The problem is that it accounts for barely more than 1% of a site's total traffic, according to measurements from the analytics tool Conductor. A marketing team that only looks at volume misses the most useful piece of information on its dashboard.

## Why a visitor from a chatbot buys more often

The answer lies in how the question was asked before the person even arrived on the site. When someone types "best invoicing software for freelancers" into Google, they get ten links and still have to compare, sort, and hesitate. When they ask the same question to ChatGPT, the model has already done that work. It has read several sources, compared the options, and returns only the two or three names that best answer that person's specific request. The visitor who clicks the final link is no longer searching, they're verifying a decision that's already been nearly made.

It's the difference between a customer who walks into a store to browse, and a customer who walks in already knowing what they want to buy, on the recommendation of a friend they trust. The chatbot plays that role of trusted intermediary. It has filtered, explained, compared. The person landing on your site no longer needs to be convinced of the product's general value, only reassured on the details.

A report published in March 2026 confirms the scale of the phenomenon: visitors coming from AI assistants like ChatGPT or Perplexity convert 42% better than non-AI traffic, across all channels combined. The quality of the intent that precedes the visit changes the nature of the visitor itself.

Take a concrete example. Someone looking for payroll software for ten employees types their question into Perplexity. The engine answers by citing three tools, each with an indicative price, a strength, and a limitation. They click on the one that best fits their situation, land on a pricing page, and request a demo that same day. The same buyer, starting from a classic Google search, would probably have opened five or six tabs, compared reviews for an hour, then put off the decision until the next day. The chatbot didn't just shorten the path. It did the filtering on their behalf, and that filtering translates directly into conversion rate.

## Advertising is still useful, but for a different goal

This isn't about pitting advertising against visibility in AI chatbots as two competing strategies. They serve different needs. Advertising pays for immediate, controllable volume: you know how much you're spending, how many visitors you're getting, and you can adjust the budget from one day to the next. Getting cited by an AI chatbot doesn't work the same way. It's built over time, through clear content, visible customer reviews, technical documentation that the models reading it can access.

The conversion gap between the two channels doesn't mean you should cut your advertising budget. It means that a euro invested in the quality of what AI chatbots find about a brand produces a return that shows up in no advertising campaign dashboard, and that return will grow mechanically as buyers increasingly use these tools.

## What this changes in how you measure performance

Most marketing dashboards are built around volume: number of visits, number of ad clicks, number of impressions. That logic makes sense for paid advertising, where each visit has a direct, predictable cost. It becomes misleading for traffic generated by AI chatbots, where volume will stay modest for a long time but where each visit carries more weight.

The table below summarizes the gap observed across several recent studies, with different methodologies, which explains the variation in the numbers:

| Traffic source | Observed conversion rate | Relative volume |
|---|---|---|
| Classic Google search | Baseline (1x) | High |
| Visitors from an AI chatbot | 4x to 11x depending on the study | Low, often under 1% of the total |
| Paid advertising | Variable, heavily dependent on cost per click | High but costly to sustain |

Comparing a channel that accounts for 1% of traffic to a channel that accounts for 60% in raw volume doesn't make sense. The right question isn't "how many visitors does the chatbot send us," it's "what is the value of a visitor sent by a chatbot, compared to that of a visitor from an ad." On that basis, a budget devoted to being well cited in AI chatbot answers can pay off more, visitor for visitor, than an equivalent ad budget, even if the total number of visitors remains incomparable for now.

## A signal that will carry increasing weight

Traffic from AI chatbots remains marginal today, but its trajectory is clear: it grows every quarter, while classic organic traffic flattens out. A team that invests now in its visibility within the answers given by ChatGPT, Perplexity, or Gemini isn't just building one more channel. It's capturing a stream of pre-qualified visitors, fewer in number but far more likely to become customers, before this channel gets big enough for everyone to chase it at once.

The logic of classic advertising is still useful for generating volume quickly. But a budget aimed entirely at volume, with nothing invested in the quality of what arrives via generative engines, ignores the underlying trend in buying behavior. The next quarter won't be won just by attracting more people, but by attracting the right people, at the right moment in their decision.

*Vurto measures the share of your traffic that genuinely comes from citations earned in ChatGPT, Perplexity, Gemini, and Claude, to tell apart what's small in volume but big in value.*

## Resources (resource center — full content)

Every published article, with full content, grouped by topic (23).

### Tip

### [Selling GEO as an agency: what changes compared to SEO](https://vurto.ai/en/resources/selling-geo-as-an-agency-what-changes-compared-to-seo) (2026-10-05)

Right now, SEO agencies keep getting the same call, phrased ten different ways. "We're told we no longer show up in ChatGPT." "A client asked why a competitor is cited by Perplexity and not us." "The boss noticed his name missing from a Gemini answer and wants an action plan by Tuesday." GEO, Generative Engine Optimization (the craft of making a brand visible and well cited in the answers of AI chatbots such as ChatGPT, Claude, Gemini or Perplexity, as opposed to classic ranking on Google), has just moved from curiosity to budget line. Agencies that know how to sell it capture that line. The others watch a competitor take it from them.

Demand is not the problem. It exists, it is growing, and it comes straight from clients. The problem is that almost nobody knows how to structure a GEO offer that holds up: one that is not an SEO contract copied and pasted with a single word changed, that bills real work rather than a vague PDF report, and that scales across several clients at once without blowing up the team's time.

## What a client is really buying

A client asking for GEO usually does not know what they are buying. They saw a worrying figure, heard a competitor talk about it, or noticed for themselves that when they type a question about their own industry into ChatGPT, their brand never appears in the answer. Their real need boils down to one question: "do AI chatbots see me, and if not, why."

This is where the agency has to set a clear framework, because GEO actually covers three different jobs, and selling all three under a single label always ends in disappointment.

The first job is diagnosis. Querying the chatbots on the searches that matter to the client, checking whether the brand appears, how it is described, and against which competitors. It is measurement work, comparable to a classic SEO audit but on a terrain where results change from one week to the next and differ from one engine to another.

The second job is technical correction. Making the site readable by the bots that feed these chatbots: content that loads without heavy JavaScript, a clear structure, a properly written `llms.txt` file (a text file at the root of the site that tells AIs what they can read and how to understand it, the counterpart of the `robots.txt` file that has existed for twenty-five years for classic search engines). This work looks like technical SEO, but the criteria are not the same.

The third job is content production and reputation management. Writing or rewriting what needs to be cited, monitoring how the brand is perceived over time, stepping in when an AI repeats false or outdated information about a client. It is editorial and relationship work that never stops, unlike a one-off audit.

An agency that sells all three in a single fixed package is bound to lose money on the one that demands the most ongoing time: monitoring. An agency that sells the three separately can bill each at its fair value, and above all sell the diagnosis as the entry point to the other two.

## Structuring the offer: three tiers, not a single package

The good practice emerging among agencies that have already sold GEO to several clients looks like this: a short, affordable first tier that opens the door, a second tier that fixes, a third that monitors over time.

| Tier | Content | Duration | Billing logic |
|---|---|---|---|
| Diagnosis | Visibility audit on 20 to 50 key queries, map of cited competitors, findings report | 1 to 2 weeks | Fixed fee, entry price |
| Correction | Technical rewrite (llms.txt, site structure, product pages or key pages), priority content plan | 4 to 8 weeks | Project fee, defined deliverables |
| Monitoring | Monthly multi-chatbot tracking, brand perception alerts, continuous content adjustment | Recurring | Monthly subscription |

This structure has a precise commercial advantage: it avoids selling a fuzzy promise. "We'll make you visible on AI" means nothing to a marketing director who has to justify a spend. "Here is what we measure, here is what we fix, here is what we monitor, and here is the price of each step" can be defended in front of any budget committee.

The "monitoring" tier is the one that keeps the agency alive over time, and it is also the one most agencies underestimate in terms of workload. Tracking a single client's visibility on four different chatbots, every week, by hand, takes time. Doing it for fifteen clients by hand becomes impossible. That is the real operational wall of GEO in an agency: not the sale, not the method, but the ability to repeat the work without multiplying billed hours by the number of clients.

## The operational wall: running ten clients like one

An SEO agency that adds ten clients adds ten Google Search Console accounts to watch. A GEO agency that adds ten clients has to monitor ten brands on four chatbots, potentially forty information streams to check regularly, each of which can change overnight. Without a suitable tool, that load falls on a consultant copying chatbot answers into an Excel spreadsheet. It does not last three months.

So the question to ask before selling a GEO offer at volume is not "can I do the diagnosis" but "can I redo it every week, for every client, without spending my life on it". That is exactly the kind of problem a tool like Vurto is built to solve: it automates querying the chatbots for each tracked brand, centralizes perception and citations per client, and offers an agency mode that lets you manage several client accounts under a single dashboard, with differentiated roles for the team and a shared activity history. The agency keeps the advice and the client relationship. The tool absorbs the repetition.

## Bill the proof, not the promise

The last point, and the most sensitive commercially, is proof of results. A client paying for a GEO monitoring subscription will ask, after three months, "what has it changed". Answering with a gut feeling will not be enough for long.

The indicators that hold up are simple to explain even to a non-technical client: the citation rate (the share of tested queries where the brand appears in the chatbot's answer), the evolution of the tone the AI uses to describe the brand, and the relative position against tracked competitors on the same queries. Three figures, measured before and after, are enough to turn a GEO report into an argument for contract renewal. That is far more convincing than an export of ChatGPT conversation screenshots, which remains the most widespread method and the least defensible in front of a client comparing several quotes.

Selling GEO as an agency, in 2026, is therefore not about inventing a new profession. It is about applying a discipline that is already known, visibility consulting, to a terrain that changes faster and cannot be measured with the same tools as before. The agencies that win this battle will not be the ones that talk best about GEO in a sales meeting. They will be the ones that can prove, figures in hand, client after client, that their work has a measurable effect on what an AI says about a brand.

*Vurto was designed for this scale-up: tracking several brands, several chatbots, one interface, with an agency mode that centralizes client accounts without multiplying manual work.*

### News

### [Vurto to Exhibit at Tech for Retail 2026, Helping Retailers Get Cited in Generative AI Answers](https://vurto.ai/en/resources/vurto-tech-for-retail-2026) (2026-08-28)

**Paris, August 28, 2026**

Vurto, the Generative Engine Optimization (GEO) platform built for retail and e-commerce, is exhibiting at Tech for Retail, Europe's retail trade show, on November 30 and December 1, 2026 at Paris Expo Porte de Versailles, Hall 7.2, booth V32.

A shopper types "what running shoe brand do you recommend for a beginner" into ChatGPT or Perplexity instead of Google. They don't get ten blue links to compare themselves. They get an answer that's already been decided, citing two or three brands, sometimes just one. For an e-commerce site or a retail chain, being left out of that answer means disappearing from a purchase channel that already shapes decisions, before the customer even opens a browser.

Vurto was built for this new playing field. The platform continuously monitors how ChatGPT, Gemini, Perplexity and Claude talk about a brand and its competitors, analyzes the sentiment and attributes associated with it, and identifies the sources models cite most often when forming an opinion. It goes further than diagnosis: Vurto converts product catalogs into structured markdown, scores every product page against ten AI-readability criteria, and generates the content recommendations that fill the gaps found in fan-out queries and grounding queries — the questions models ask themselves behind the scenes before answering.

Retail is especially exposed to this shift. Catalogs run to thousands of SKUs, product pages are often mass-generated without ever being designed for a non-human reader, and competition between retailers now also plays out in the invisible: the handful of sentences a model chooses to remember before answering a shopper. A site perfectly ranked on Google can be entirely absent from AI answers, for simple technical reasons — JavaScript not rendered by AI crawlers, unstructured content, no llms.txt file, product pages written to persuade a human but unreadable to a model looking for verifiable facts.

Every year, Tech for Retail brings together the players actually transforming the sector, from agentic AI to personalization to supply chain automation. Vurto will present its approach to GEO applied to retail and meet with marketing, e-commerce and digital teams facing this shift in search behavior.

Join us on November 30 and December 1, 2026, booth V32, Hall 7.2, Paris Expo Porte de Versailles, to see how to turn a product catalog into a source generative AIs cite, recommend and favor.

## About Vurto

Vurto is a Generative Engine Optimization platform that helps brands and e-commerce sites measure and improve their visibility in generative AI answers (ChatGPT, Gemini, Perplexity, Claude). Multi-LLM monitoring, brand perception analysis, content recommendations, AI-readability audits and product catalog conversion: Vurto covers the full chain, from diagnosis to action.

### GEO Fundamentals

### [Local GEO: How Geography Is Again an AI Visibility Lever](https://vurto.ai/en/resources/local-geo-how-geography-is-again-an-ai-visibility-lever) (2026-07-27)

A tradesperson in Lyon types "best plumber near me" into ChatGPT. An expat looks up "where to eat vegan in the 11th arrondissement" on Perplexity. A B2B buyer asks Gemini which CRM integrator based in Lille can show up within 48 hours. Three queries, one thing in common: they expect an answer anchored in a precise territory.

Local SEO has existed for fifteen years. Google Business Profile, customer reviews, consistent NAP citations, city pages. None of that know-how disappears with LLMs. But the mechanics change. A classic search engine crosses your geolocated position with an index of nearby results. An LLM has to understand the local context from text, rephrase the query, and choose sources to cite. Geography becomes a semantic filter before it is a proximity filter.

## Why localization is surfacing again with LLMs

Generative engines handle a local query by breaking it down. "Best plumber near me" becomes a series of sub-queries: what type of service is being asked for, what geographic area "near me" covers, which criteria define "best" (reviews, availability, prices, specialty). This fan-out decomposition favors businesses whose content answers each of those sub-questions directly, rather than those that only have an address listing.

ChatGPT, Perplexity and Gemini do not all have the same geographic precision. Perplexity leans heavily on fresh web sources and readily cites local directories, regional press articles, neighborhood forums. ChatGPT, when it turns on web search, prefers pages that clearly state their catchment area in the text, not only in the metadata. Gemini, connected to the Google ecosystem, still draws on Business Profile and reviews, but rewrites the answer in its own words, which dilutes the direct link to your listing.

The result: a local business well ranked on Google Maps can stay invisible in a ChatGPT answer if its site never explicitly says "we serve Villeurbanne, Vénissieux and Caluire" in text an AI can read.

## How LLMs reconstruct local context

An LLM does not know where you are, unless the application tells it. It therefore works with the explicit geographic signals contained in the query and the sources. Three kinds of signals matter.

The first is textual mentions of area. A page that lists "Paris, Boulogne-Billancourt, Neuilly-sur-Seine" in full, in a normal sentence, is more likely to be associated with those cities than a page that only puts a postcode in the footer.

The second is third-party sources that confirm local presence. A regional press article, a professional directory listing, a dated and localized Google review, a mention on a forum such as Reddit or a local Facebook group. LLMs weigh consistency across several independent sources more than a single site declaring itself the local expert.

The third is structured data. LocalBusiness schema is still read by generative engines that crawl raw HTML, provided the page is not entirely client-side JavaScript. An address, opening hours, a service area declared in JSON-LD give the AI a clean signal, with no ambiguity of interpretation.

## What actually makes the difference

Many local businesses think one page per city is enough. That is a mistake that already cost dearly in classic SEO, and costs even more in GEO. A "plumber in Villeurbanne" page that copies the "plumber in Vénissieux" content and only changes the city name adds no new information. An LLM comparing several sources detects that duplication and prefers a source that describes a real local foothold: an active local phone number, a named customer testimonial, a recent job told with verifiable details.

Content that works answers precise local questions. Not "our plumbing services" but "how much does unblocking a drain cost in Lyon in 2026" or "what is the average plumber response time in the 6th arrondissement". That level of granularity matches what LLM fan-out queries are trying to fill.

Local press and specialist directories play a disproportionate role. An article in Le Progrès that mentions your company often weighs more, in an LLM's eyes, than a page on your own site optimized for the keyword. The reason is simple: an independent third-party source is seen as more reliable than a self-declaration. Businesses that invest in local press relations and partnerships with sector directories are, without realizing it, building local GEO capital.

Here is how the two disciplines overlap and diverge:

| Lever | Local SEO | Local GEO |
|---|---|---|
| Google Business Profile | Direct ranking signal in Maps | Secondary signal, reused by Gemini especially |
| Customer reviews | Influences ranking and click-through | Social proof cited in the generated answer |
| Duplicated city pages | Tolerated if slightly differentiated | Penalizing, detected as hollow content |
| Local press and directories | Useful for link building | Third-party source the LLM cites for credibility |
| LocalBusiness structured data | Optional, helps rich snippets | Key signal if the site is readable without JavaScript |
| Real geographic proximity | Direct ranking factor (IP, GPS) | Absent unless the app passes the user's location to the AI |

## Pitfalls to avoid

The first pitfall is believing the Google Business Profile is still enough to cover everything. It remains essential for Maps and for Gemini, but it is not read the same way by ChatGPT or Perplexity, which lean more on the open web.

The second pitfall is over-optimization by city. Generating fifty automatic pages for fifty neighboring towns produces content that LLMs identify as programmatic spam. Ten pages that tell a precise local reality beat fifty hollow ones.

The third pitfall is forgetting that freshness matters. A three-year-old review, a 2021 article, an address that changed without an update: LLMs that prefer recent sources drop these signals in favor of more up-to-date competitors, even modest ones.

The last pitfall is treating local GEO as a one-off project. Visibility in generative answers is built over time, by accumulating consistent mentions across the web: a clean site, press, directories, reviews, local social networks. A single audit is not enough. You need monitoring that checks, query after query, how each LLM renders your local presence.

*Vurto monitors exactly what ChatGPT, Gemini, Perplexity and Claude answer about your business, city by city, and flags the areas where your local presence stays invisible to AIs.*

### [SEO and GEO: What Changes, What Stays, What to Do Now](https://vurto.ai/en/resources/seo-and-geo-what-changes-what-stays-what-to-do-now) (2026-06-23)

There's a convenient temptation in digital marketing: every new discipline gets announced as the death of the previous one. SEO killed display. Social media killed SEO. Now GEO is supposedly killing SEO.

That's wrong. And it's a lazy way to think about it.

GEO doesn't kill SEO. It changes SEO's role, its scope, and the logic of resource prioritization. That's different.

## What SEO Keeps Doing

SEO directly feeds GEO. Well-referenced content — backed by backlinks, well-structured, authoritative in its domain — is what LLMs find first when they search for sources.

LLMs don't reinvent the wheel. When Perplexity or ChatGPT go looking for information, they go through Google or Bing. What comes up first influences what they read. What they read influences what they say.

A site with strong SEO presence therefore has a solid GEO foundation. It's not sufficient, but it's a real advantage.

SEO also remains useful for everything LLMs haven't fully captured yet: local queries with navigational intent, queries where LLMs remain cautious or imprecise, content where users want to explore for themselves rather than receive a synthesis.

## What GEO Does That SEO Doesn't

SEO optimizes for a ranking in a list. GEO optimizes for a presence in an answer.

The difference sounds semantic. It's structural.

To rank on Google, you produce content that matches a search intent, on a keyword, with enough authority and relevance to outrank competitors. The success signal is your position in the results.

To be cited in an LLM response, you need to be a source the model judges relevant, credible, and well-structured for a given question. The question can be phrased a hundred different ways. The response the model gives looks nothing like a SERP. And the success signal is the citation, not the position.

These two logics coexist. One doesn't replace the other. But they require partially different approaches. Content perfectly optimized for Google can be unreadable for an LLM. Content well-structured in markdown with high informational density can rank modestly on Google and be consistently cited by Perplexity.

## The Priorities That Shift

What changes in practice is where attention gets allocated.

The race for high-volume keywords becomes less central. Publishing a hundred generic articles on secondary queries to accumulate organic impressions is a strategy that loses value when a growing share of responses to those queries is synthesized by an LLM before reaching the results list.

Depth and structure beat volume. A 3,000-word guide that genuinely, precisely, and substantively answers a complex question has more GEO value than a cluster of ten 600-word articles that skim the surface.

Diversity of signals matters more. LLMs search publishers, forums, social networks, comparisons. A presence concentrated on your own domain is fragile in GEO. Being present in third-party ecosystems — LinkedIn, Reddit, YouTube, sector press — carries more weight than in SEO.

Technical coherence between robots.txt, llms.txt, and LLM crawlers becomes a real issue. It wasn't an SEO dimension. It is a GEO one.

## The Right Framework for Decisions

A useful way to prioritize: for each type of content or action, ask whether the goal is to be found by a human browsing, or to be cited by an LLM synthesizing.

The two aren't mutually exclusive. Good content can serve both. But when resources are limited, knowing which objective takes precedence for a given piece of content helps calibrate the effort.

Category pages, comparison pages, buying guides, expertise pages, data-driven case studies — these are dual-value content types, both SEO and GEO. They deserve full optimization on both dimensions.

Short-term news posts, event landing pages, promotional content — their GEO value is more limited. The LLM isn't going to cite a promotional offer in a product recommendation.

## What Vurto Brings to This Transition

The difficulty of the SEO-to-GEO transition is that it requires managing two logics in parallel, with partially different signals.

Classic SEO tools — Ahrefs, SEMrush, Search Console — cover their perimeter well. They don't see LLM citations, don't measure brand sentiment in generative engines, don't flag content that's unreadable to AI crawlers.

Vurto covers that complementary perimeter: citation monitoring in ChatGPT, Gemini, Perplexity, and Claude; content recommendations based on fan-out and grounding queries; AI readability audits; technical file rewriting (llms.txt, robots.txt); and for e-tailers, catalog conversion into structured markdown.

It doesn't replace SEO tools. It covers the half of digital visibility they don't see.

---

*GEO completes SEO. Vurto covers the part that classic SEO tools don't see.*

### [GEO: The Diagnosis Nobody Is Running](https://vurto.ai/en/resources/geo-the-diagnosis-nobody-is-running) (2026-06-14)

SEO has been optimized for thirty years. Rankings, backlinks, Core Web Vitals. All of it worked in a world where traffic flowed through lists of ten blue links.

That world is disappearing.

Today, a growing share of queries never produces a click. The AI answers directly. ChatGPT, Perplexity, Gemini, Claude — these models synthesize, recommend, conclude. They cite or they don't. And if your brand isn't in the answer, it doesn't exist.

That's the territory of GEO. Generative Engine Optimization.

## What GEO Actually Measures

SEO measures your position in a list. GEO measures something more fundamental: do AI systems know you, understand you, and recommend you?

These aren't the same questions. And they don't have the same answers.

A site can rank first on Google and be invisible in ChatGPT's responses. Because LLMs don't read SERPs. They read what they ingested during training, what they fetch in real time through their search tools, and what their architecture lets them process.

GEO visibility rests on three layers.

The existence layer — do LLMs know you exist? Is your brand name, your products, your core arguments present in the sources these models draw from?

The readability layer — are your contents structured in a way that LLMs can process efficiently? A site that's technically opaque to an AI crawler is a site that's transparent to better-calibrated competitors.

The perception layer — when AI systems talk about your brand, is what they say accurate, positive, differentiating? Brand perception in LLMs is a signal you can measure, and almost nobody is watching it.

## The GEO Diagnosis, Step by Step

The right diagnosis starts with the right questions.

Is my brand cited, and in what context? This isn't binary. A brand can be cited positively in Perplexity and absent from ChatGPT. Or present with a negative sentiment in Gemini. Multi-LLM coverage is the first diagnostic layer.

Are my contents readable by AI? LLMs process a well-structured markdown document very differently from a proprietary HTML product page or an unindexed PDF. Auditing the AI readability of your key content means finding blind spots before they cost you citations.

Do I understand the queries AI systems run about me? LLMs don't answer a user's question directly. They first break it down into sub-queries — fan-out queries — that they send to their search tools. If you don't know which fan-out queries apply to your category, you're optimizing blind.

Is my brand perception in AI consistent with my actual positioning? What LLMs say about your brand can diverge significantly from your official messaging. Identifying that gap means knowing where to act first.

## Why the GEO Diagnosis Is Still Rare

The short answer: traditional SEO tools don't measure it.

Google Search Console gives you impressions and clicks. Ahrefs gives you backlinks. SEMrush gives you positions. None of them tell you what ChatGPT thinks of your brand this morning.

The GEO tools market is structuring fast, but maturity isn't uniform. Some platforms do citation monitoring. Others add recommendations. The most advanced ones, like Vurto, cover the full chain: multi-LLM monitoring, brand perception analysis, content recommendations based on fan-out and grounding queries, AI readability checks, and rewriting of strategic content (product pages, llms.txt, e-commerce catalog).

The question is no longer whether GEO deserves attention. That's settled. The question is how precisely you run the diagnosis before you start treating.

## What Changes in Practice

For SEO teams: the skills transfer partially, but the reference frame shifts. You're no longer chasing a keyword ranking. You're trying to be the source an AI reaches for when a user asks a question related to your category.

For content teams: the writing angle changes. Not to please a ranking algorithm, but to be processable by an LLM. Clear structure, informational density, named entities, markdown format where possible.

For e-commerce teams: the product catalog becomes a major GEO asset. A poorly structured product page is a product page the AI won't recommend — even if it's perfectly optimized for Google.

For marketing leadership: AI visibility becomes a KPI in its own right, with its own benchmarks, trends, and alerts. Not once a quarter at the reporting meeting. Continuously.

## The First Step

Run the diagnosis before prescribing. That's true in medicine. It's true in GEO.

Start by auditing your site's AI readability. Check whether your llms.txt exists and is properly configured. Ask ChatGPT, Perplexity, and Gemini about your brand and products yourself — what you read there is worth a thousand positioning reports.

Then build the monitoring. Not to have one more dashboard, but to know where you stand and where you're heading.

---

*Vurto is a GEO tool covering the full diagnostic chain: multi-LLM monitoring, brand perception, AI readability, fan-out queries, and content rewriting for generative engines.*

### Content Strategy

### [ChatGPT, Perplexity, Gemini: four engines, four sourcing logics](https://vurto.ai/en/resources/chatgpt-perplexity-gemini-same-disease-three-different-prescriptions) (2026-09-28)

ChatGPT, Perplexity, Gemini and Claude do not draw on the same sources to build their answers. A site visible on ChatGPT can stay invisible on Perplexity. GEO starts by treating one engine at a time.

The question often comes back in a form that is too vague: "How do I become visible on AI?" AI does not exist as a single block. There is ChatGPT, Perplexity, Gemini, Claude. Four distinct tools, each with its own way of selecting information. What they have in common: they are chatbots able to answer in natural language by fetching sources on the web, beyond what they learned during training. Their difference, the only one that matters for a brand: they do not draw on the same sources to build an answer.

GEO, Generative Engine Optimization (the craft of making a brand visible and well perceived in the answers of these chatbots), starts with a simple observation that is too often skipped. Stop treating these four tools as one uniform set. A site optimized for ChatGPT can be invisible on Perplexity. A brand that shows up strongly on Gemini may never appear on Claude. This is not a bug. It is the direct consequence of different architecture choices, made by different companies, with different interests.

## Four engines, four sourcing logics

Step back and look at how each engine picks its sources. Independent studies run on tens of thousands of citations in 2026 draw very sharp, almost caricatural profiles.

Perplexity behaves like a reporter who favors the field. It relies heavily on YouTube (around 30% of its citations in some analyses) and on Reddit, the forum platform where people discuss everything, often more candidly than a press release. Reddit can account for nearly half of its top ten sources. Perplexity prefers lived experience told by a stranger over the claim of an institutional website.

ChatGPT, for its part, leans toward established authority. Wikipedia holds a disproportionate place, and well-known media (Forbes, Reuters) complete the picture. The profile looks like a serious student who cites academic sources rather than comments from the neighborhood forum. But watch what comes next, because this profile moves fast.

Gemini, owned by Google, keeps the reflexes of a corporate librarian. It keeps citing the group's own properties widely (YouTube first) while keeping Reddit and Wikipedia in a good spot. A more balanced split, consistent with an engine that has access to the most complete search index on the web.

Claude, developed by Anthropic, completes the picture with yet another profile, closer to ChatGPT in its taste for documented sources, but giving a larger share to technical documentation and specialized sites when the question calls for it. Four engines, four sourcing logics, and no single strategy that fits all.

A table beats a long speech here.

| Engine | Dominant sources | Profile |
|---|---|---|
| Perplexity | YouTube, Reddit | Field, lived experience, community |
| ChatGPT | Wikipedia, recognized media | Authority, formal knowledge, volatile |
| Gemini | YouTube, Reddit, Google properties | Balanced, anchored in the Google ecosystem |
| Claude | Documentation, specialized press, Wikipedia | Technical, precise, expertise-oriented |

Remember one thing from this table. Nobody builds an effective content strategy by targeting "AI" in general. You treat one engine at a time. An HR software brand that neglects Reddit cuts itself off from Perplexity. A services company that never keeps an up-to-date Wikipedia page loses part of the credit it would deserve on ChatGPT. These are not technical details reserved for specialists. These are choices that decide which of your competitors gets recommended instead of you.

## When an engine's sources shift in a few weeks

Here is the most underestimated point of the topic, and probably the most important. These profiles are not set in stone. A study covering 230,000 queries tracked over thirteen weeks showed that Reddit's share in ChatGPT answers went from about 60% in August to only 10% six weeks later. Wikipedia, in the same movement, dropped from 55% to under 20%. A collapse of that scale, over such a short time, on one and the same platform.

Meanwhile, Perplexity stayed remarkably stable, with minor variations on its usual sources. Gemini, too, changed little.

This contrast says something essential about these tools. ChatGPT probably adjusts its weightings continuously to avoid over-citing a handful of domains and looking like a biased engine. Perplexity and Gemini, with different sourcing logics, absorb these adjustments better. The message for you: a position won on one engine in September can vanish in October, without you having made any mistake. The ground moves under your feet. One more reason never to talk about "AI positioning" as a permanent gain, but as a state to monitor continuously.

## What this concretely changes for your content

This heterogeneity has a direct consequence on how you should produce and distribute your content.

Stop betting on a single channel. If your content strategy is limited to blog articles optimized the classic search way, you may cover part of ChatGPT, but you stay invisible on Perplexity. An active presence in Reddit discussions relevant to your sector, YouTube videos that explain your product, an up-to-date Wikipedia page if your notoriety justifies it: these are different building blocks, for different engines, and none replaces the others.

Also beware of wins that rest on a single source. A brand that owes all its AI visibility to a handful of Reddit mentions lives on unstable ground, as we just saw with ChatGPT. Robustness comes from the diversity of entry points, forums, video, specialized press, well-structured product pages, kept up to date at the same time rather than one isolated bet.

And above all, do not confuse being cited with being well perceived. Appearing in an answer guarantees neither a favorable tone nor an accurate description of your offer. That is a topic in its own right, but it starts with the same prerequisite: knowing precisely where and how you appear, engine by engine, rather than guessing.

Take a concrete example. A company that sells software to businesses (what is called B2B, as opposed to selling directly to consumers) usually invests heavily in white papers and case studies published on its own site. This content feeds ChatGPT well, especially if it is picked up by the specialized press. But it barely feeds Perplexity, which will rather look for a customer's opinion posted on a forum or a commented LinkedIn post. Two audiences, two formats, two engines. The content budget must reflect that reality, not ignore it.

## Measure before you act

The temptation, faced with this complexity, is to give up and tell yourself that "AI moves too much to be worth our time". That is the reverse of the mistake we pointed out at the start. Movement does not exempt you from monitoring, it demands it. A serious company does not decide its sales strategy on an intuition from six months ago. It looks at this month's numbers. The logic is the same here.

What matters is knowing, for your brand and your competitors, which engines cite you, which sources they rely on to do so, and how that split evolves over time. Without that regular snapshot, you optimize blind, hoping that what worked on ChatGPT in the spring still works on Perplexity in the fall. That is a bet, not a strategy.

*Vurto tracks these variations engine by engine, ChatGPT, Gemini, Perplexity and Claude, to show precisely which sources carry your visibility and which ones are collapsing before you find out by chance.*

### [Perplexity, ChatGPT, Gemini: Three Engines, Three Source Lists, One Mistake to Avoid](https://vurto.ai/en/resources/perplexity-chatgpt-gemini-three-engines-one-mistake-to-avoid) (2026-09-14)

A number to start with. Out of 680 million citations analyzed between August 2024 and April 2026 by the 5W index, only 14% of cited sites dominate across all AI chatbots (these conversational assistants like ChatGPT, Gemini, or Perplexity, able to answer in natural language from sources they select themselves). Between ChatGPT and Perplexity alone, overlap drops to 11%. In another study covering 11,500 queries compared against Google results, GPT-4o shows 0% median overlap with Google's top 10. Perplexity, 14.3%. Gemini, 8.5%.

These numbers should settle a debate that has dragged on for two years in marketing teams: no, "being visible in AI" is not a single objective. It's three different objectives, with three different selection logics, and a good part of the groundwork consists of understanding which one applies to which client.

## Three philosophies, not three variants of the same engine

The starting mistake is to treat ChatGPT, Perplexity, and Gemini as competitors doing the same thing with different branding. They are three architectures that answer different questions about what makes a good source.

ChatGPT relies heavily on consensus-based, already-aggregated content. Wikipedia accounts for between 26% and 48% of its top 10 citations depending on category. Reddit comes right behind, with a weight exceeding 40% across all categories and engines combined, and staying particularly high with ChatGPT. The logic is that of a model favoring what has already been collectively validated: content synthesized, debated, corrected by thousands of contributors.

Perplexity works the opposite way. Its engine queries the live web on every request rather than drawing from a fixed memory, which explains why it favors Reuters, AP, Bloomberg, the Wall Street Journal: news sources with strict attribution and verifiable freshness. A Columbia Journalism Review study measured a citation error rate of 37% for Perplexity Sonar Pro, versus 67% for ChatGPT Search. The gap comes down to method: searching live and citing your source produces fewer fabrications than rephrasing from training memory.

Gemini, finally, remains Google's child. It draws from the search engine's index, from YouTube (which Google owns), and shares with AI Overviews a clear fondness for Reddit. It's the engine closest to classic SEO culture, the one where web ranking, domain authority, and presence across the Google ecosystem still count directly.

Claude, often looked at from a distance in these comparisons for lack of search volume, stands out for a taste for long-form analysis and technical documentation. It cites proportionally more expert blogs and less Reddit than the others.

## The table that changes the roadmap

| Engine | Preferred sources | Dominant logic |
|---|---|---|
| ChatGPT | Wikipedia, Reddit, consensus content | Trained memory + collective synthesis |
| Perplexity | Reuters, AP, Bloomberg, press with attribution | Real-time search, freshness |
| Gemini | Google index, YouTube, Reddit | Google ecosystem, classic SEO |
| Claude | Expert blogs, technical documentation | Long-form analysis, depth |

This table isn't an exercise in curiosity. It dictates where a brand needs to exist depending on which engine matters most to its audience. A B2B software vendor (selling to other businesses rather than individuals) whose prospects mainly use Perplexity to compare solutions is better off working on press relations and citations in serious trade media, far more than optimizing a Reddit thread. A consumer brand that lives and dies on ChatGPT has the opposite interest: community conversation outweighs the press release.

## What the low overlap actually means

The most underestimated point isn't the list of preferred sources. It's the fact that they barely overlap at all. Content that breaks through on ChatGPT can be invisible on Perplexity, and vice versa. This means a content strategy designed for "generative AI" in general, without distinguishing between engines, is in reality optimizing for a single engine without realizing it, usually the one the brand heard about first, often ChatGPT because it concentrates most of the usage volume.

In France, this concentration reaches extreme levels: ChatGPT captures 84.47% of clicks generated by AI assistants according to SE Ranking, Perplexity 12.82%, Gemini 2.08%. An AI visibility budget that ignores this split and treats the three engines equally wastes precious time. But the reverse is also true over an eighteen-month horizon: Gemini is natively built into Android, into Google Workspace, into the Google search bar itself. Its share of usage is growing faster than its share of citations today would suggest. Ignoring Gemini because it accounts for 2% of current clicks is like ignoring mobile in 2010 because it still represented only a fraction of web traffic.

The practical consequence fits in one sentence: stop measuring "AI visibility" as a single score. Content can be cited ten times a week on ChatGPT and never on Perplexity without that being a failure, if the brand's target audience lives on ChatGPT. But no one can know that without tracking the engines separately, each with its own citation logic, rather than an average that hides everything.

## Adapting content to each engine's logic, not to a universal checklist

Concretely, this changes three things in how content gets produced.

To exist on ChatGPT, you need to feed the collective conversation rather than bypass it. A structured presence on Reddit in the right communities, an up-to-date Wikipedia entry if the brand or its leader has a place there, content that synthesizes a topic rather than selling it. It's ecosystem work, not landing-page work.

To exist on Perplexity, freshness and attribution matter more than volume. A well-distributed press release, relationships with journalists who cite the brand with a clear link, dated content republished regularly carry more weight than a static white paper published once and never touched again.

To exist on Gemini, classic web ranking fundamentals remain the best way in: domain authority, clean technical structure, YouTube presence. It's the engine where the last ten years of SEO work keep paying off directly, which paradoxically makes it the easiest to work on for a team coming from classic search.

None of these three approaches replaces the other two. A brand with the time and budget to do all three builds robust AI visibility, insensitive to a single engine's algorithm changes. A brand that has to choose should first look at where its audience actually asks its questions, not where it itself is most comfortable producing content.

## Tracking three engines beats guessing on one

The low overlap between ChatGPT, Perplexity, and Gemini isn't a technical problem to solve. It's a structural constraint to build into how you measure and prioritize. A brand that only tracks its citations on one engine is flying blind on the other two, and makes content decisions based on a third of reality.

*This is exactly why Vurto tracks citations separately across ChatGPT, Gemini, Perplexity, and Claude rather than producing one aggregated score that erases these differences: a brand needs to know which engine it exists on, which one it's absent from, and why the two don't look alike.*

### [Google's AI Summary That Killed the Click](https://vurto.ai/en/resources/googles-ai-summary-that-killed-the-click) (2026-09-10)

Type a question into Google today, and odds are a box appears right at the top of the page, ahead of the usual list of links. That box is called an AI Overview, literally an "AI-generated overview." A summary written by Google's AI model that answers the question directly, with a few small links off to the side. The person searching for information already has it in front of them. They no longer need to click on anything.

This box now appears on 15 to 25% of all Google searches, and up to half of searches when the question is informational rather than transactional in nature. Its direct consequence shows up in a number that should alarm any site that lives off Google traffic: when this summary appears, the click-through rate to the sites ranked below it drops by nearly 60%. A good chunk of the free traffic that thousands of businesses have built their visibility on for fifteen years is evaporating before their eyes, replaced by text that Google writes itself from their content, without sending them a single visitor in return.

## A site can be the source and receive no visitors

The most disorienting part of this mechanism is that a site can supply the information shown in Google's summary without a single visitor ever clicking on it. On searches where this type of summary appears, 83% of people close the page without clicking on any link, neither the one in the summary nor those in the classic list. On Google's more advanced search mode, where AI takes up an even bigger share of the page, that figure climbs to 93%. More broadly, the share of Google searches that end with no click at all rose from 50% in 2019 to nearly 65% in early 2026, and reaches 68% in the United States over that same period.

This shift changes the very definition of what it means to "be visible" on Google. For twenty years, being well-ranked meant showing up at the top of a list of links and getting a click. Today, you can be the source cited in a summary, read by thousands of people, and never know it, because no visit shows up in the site's statistics. Visibility still exists, but it no longer automatically translates into measurable traffic.

Take a site that has published a detailed guide on boiler maintenance for years. It held the top spot on Google and got several hundred visits a day on that one page alone. Google's generated summary now captures the gist of that guide in a few lines, with a discreet link back to the source. The site's content answered thousands of people's questions this month. The site itself saw only a few dozen extra visits. The editorial work hasn't lost its value to the reader, it has simply changed who visibly benefits from it.

## What determines who gets cited in the summary

Google's generated summary works on a logic close to that of chatbots like ChatGPT or Perplexity. The model has to choose, among the pages that best answer the question asked, which ones to cite and summarize. A poorly structured page, filled with sales pitches before getting to the point, or one that requires a lot of interaction to display its content, has less chance of being chosen than a page that answers directly and clearly from the very first lines.

Three elements come up in most analyses of pages cited by these summaries. First, an answer formulated in a self-contained, direct way, one that makes sense even taken out of its context, since that's exactly what the model is going to do with it. Second, a clear structure, with headings that announce the content that follows, rather than one long continuous block of text where the information gets buried. Third, a perceived authority on the topic covered, built through the consistency and regularity of the content published on that theme, not through a single article, however well written.

| What helps you get cited | What limits your chances of being cited |
|---|---|
| Clear answer from the first lines | Sales pitch before the answer |
| Headings that structure the information | Continuous text with no visual landmarks |
| Content accessible without interaction | Content loaded only through complex scripts |
| Consistency over time around a topic | A standalone article, with no related content on the same theme |

## The clicks that remain are worth more

There's a paradox worth noting before giving in to discouragement. While the presence of a Google-generated summary causes clicks to drop, the people who click anyway convert 23% better than on a classic search. These visitors have already read a summary of the answer. If they click through to a site on top of that, it's because they're looking for more precise information or want to take action, not because they're just starting their search. The mechanism resembles what's been observed with traffic sent by AI chatbots like ChatGPT: less volume, but a visitor who is much further along in their decision.

That doesn't offset the loss in volume for most sites, but it does change how you should evaluate your content strategy. A site that loses 60% of its clicks on a query but gains citations in the summaries isn't necessarily losing influence. It's losing old-style measurable traffic, while gaining exposure that builds its reputation among people who will never open its site but will remember its name, because Google presented it to them as a trustworthy source.

## Adapting rather than enduring

Faced with these numbers, the temptation is to conclude that producing content no longer serves any purpose since the click is disappearing. That's the wrong conclusion to draw. Content matters more, not less, because it now has to convince a model to choose it before it can convince a human to click. The difference is that the format has to change: less sales pitch up front, more direct and structured answers, a consistent presence on the topics where you want to be recognized as a reference.

The real risk isn't losing traffic on Google. It's continuing to write for a human reader who skims a page before clicking, when the first reader of much of that content is now a model deciding, in a few seconds, whether it's worth citing.

*Vurto checks whether your content is genuinely structured to be understood and cited by these summaries, on Google as well as on AI chatbots, before the traffic drop becomes hard to explain internally.*

### [GEO for B2B Brands: Why Your Prospects No Longer Read Reviews Themselves](https://vurto.ai/en/resources/geo-for-b2b-brands-why-your-prospects-stopped-reading-reviews-themselves) (2026-09-07)

GEO — the art of being visible in the answers given by chatbots like ChatGPT, Claude, or Gemini rather than in Google's blue links — doesn't play out the same way depending on who's buying. A brand selling to individual consumers and a brand selling to other businesses (what's known as B2B, for "business to business") aren't playing on the same field. The latter is losing control of its sales funnel without even realizing it.

In April 2026, G2, one of the largest online review platforms for business software, published a number that should alarm any B2B marketing team. 51% of software buyers now start their research in an AI chatbot, up from 29% a year earlier. ChatGPT alone captures 63% of that activity. Over the same period, human traffic to G2's site dropped 84.5%, its competitor Capterra's by 89%, and TrustRadius's by 92%. These three names are the reference sites where businesses publish and read reviews on business software, the B2B equivalent of a specialized Trustpilot. A study by 6sense of nearly 4,000 B2B buyers confirms the trend on a larger scale: 94% of them use an AI chatbot at some point in their buying journey.

These numbers don't tell the story of review sites dying. They tell the story of a shift. Buyers aren't giving up on reading customer reviews, they're giving up on going to read them themselves. A chatbot does it for them, synthesizes, compares, and recommends three tools before a salesperson has even known a prospect existed. For a B2B brand, the question is no longer whether it shows up on the first page of Google. It's whether it's one of the three names ChatGPT cites when a buyer asks "what's the best CRM software for a team of 20 salespeople."

## A buying cycle that starts outside your control

B2B selling has always had a long buying cycle, with several decision-makers and a silent research phase before any sales contact. GEO doesn't create this dynamic, it moves it further upstream. What used to happen on Google, on comparison sites, and on specialized forums now happens in a conversation with a chatbot, invisible to you, leaving no trace in your CRM, nothing to observe in your site statistics.

A direct and underestimated consequence: 69% of buyers surveyed by G2 say they chose a different vendor than the one they initially had in mind, based solely on a chatbot's recommendations. A third bought a tool they had never heard of before the conversation. The shortlist a prospect builds before ever talking to a salesperson is no longer the product of their own experience or word of mouth. It's the product of a synthesis produced by a chatbot, from sources it deems reliable. If your brand isn't among those sources, you're not just losing a click. You're losing the sale before it ever existed.

Take an operations director looking for inventory management software for an industrial SME. Two years ago, they would have typed their query into Google, clicked on three or four results, compared pricing grids across ten open tabs. Today, they ask ChatGPT the question in one sentence, get three names with their strengths and limitations, and only contact the ones that come out of that synthesis. The salesperson for the software left out of that answer will never know an opportunity existed. No form filled out, no trace in their CRM, just a silent absence in a conversation no measurement tool can capture.

## G2, Capterra, TrustRadius: the trust base behind AI chatbots

Here's the paradox many marketing teams haven't yet come to terms with. Human traffic to review sites is collapsing, but their weight in AI chatbot answers is growing. A Quoleady study of near-purchase searches, of the "[software] alternatives" type, found that 100% of recommended tools had reviews on Capterra and 99% on G2. These platforms are no longer read by humans, they're being digested by models as aggregated social proof. A chatbot treats a review published on G2 as a more reliable signal than a product page, because it carries the mark of an independent judgment.

A report published by G2 in early 2026 goes further: the platform's recent acquisition could push its share of citations in pre-purchase answers up by 76%, simply because the volume of reviews it aggregates becomes strategic raw material for the models. For a B2B brand, ignoring its presence on these platforms while betting everything on its own site is a miscalculation. The corporate website remains useful for convincing a prospect once they've arrived. It carries almost no weight in the phase where the chatbot builds its recommendation.

## What sets selling to businesses apart from selling to consumers

Consumer-facing GEO, the kind aimed at individuals (B2C, "business to consumer"), tries to get a product into a one-off recommendation: what's the best sunscreen, which vacuum to buy. GEO aimed at businesses plays a different game. It tries to build perceived authority across an entire category, because the final decision involves several people, several research sessions, and a comparison that stretches over weeks.

| Dimension | Selling to consumers | Selling to businesses |
|---|---|---|
| Goal | Show up in a one-off recommendation | Be perceived as a reference across the whole category |
| Sources that matter | Reddit, product reviews, consumer comparisons | Professional review sites, technical documentation, expert publications on LinkedIn |
| Decision | Fast, often a single person | Long, several decision-makers to convince |
| Content that carries weight | Short reviews, social content | Detailed comparisons, case studies, technical proof |

Three levers matter in particular when selling to businesses. First, the depth of technical documentation: a chatbot that has to answer "does this tool connect with Salesforce" will look for a precise integration page, not a marketing pitch. Second, honest comparative content: pages that fairly compare a product to a competitor, without hiding its limitations, get cited more often than pages that only praise their own merits, because they answer directly to the implicit question of a buyer in the selection phase. Third, the credibility of the company's experts, on LinkedIn in particular, which strengthens the trust models place in a source before citing it.

A B2B buying committee averages six to ten people, each with their own criteria. The technical lead is looking for an answer on security and integrations with other software. The finance lead wants to understand total cost over three years. The future end user wants to know if the tool is pleasant to use day to day. A chatbot queried by each of these profiles will look for different sources to answer different questions. A brand that only covers the product angle, without solid security documentation or clear financial content, loses citations on half the questions its own buying committee is asking.

## Invisibility now costs more than it used to

A third of B2B buyers today buy a tool they didn't know about before their conversation with a chatbot. That number needs to be read both ways. It's a threat to established brands that thought their position was secured by their historical name recognition. It's a real opportunity for smaller or newer players, who can earn a spot in an AI recommendation without a market leader's budget, provided they've built the right signals: structured presence on review platforms, clear and accessible documentation, honest comparative content, identifiable expert voices.

The real risk isn't doing GEO badly. It's continuing to measure performance only with yesterday's tools. A team that tracks its traffic to G2 and reads its decline as a sign of waning market interest is missing the opposite signal: the market is still reading those reviews, just through a chatbot intermediary. The question to ask every quarter is no longer "how many visitors on our product page," it's "when a buyer in our category asks ChatGPT or Perplexity, are we in the answer."

*Vurto tracks this question precisely for B2B brands: which queries generate a citation, which sources carry weight in the answer, and where the gaps are against the competition.*

### [GEO and Video Content: YouTube, the New Preferred Source of Generative Engines](https://vurto.ai/en/resources/geo-and-video-content-youtube-new-preferred-source-of-generative-engines) (2026-08-31)

Between August and December 2025, Reddit's share of social citations in AI answers fell from 44.2% to 20.3%. Over the same period, YouTube's jumped from 18.9% to 39.2%. An Adweek survey published in January 2026, combining data from four research firms across 6.1 million citations, confirms the shift. YouTube now ranks ahead of Reddit as the social platform most cited by ChatGPT, Perplexity and Google AI Overviews.

The forum that had dominated citations for two years has just lost first place. Not to an unexpected competitor, but to a platform most marketing teams still treat as a distribution channel, never as a knowledge source. That is the first mistake to correct.

## Why a video becomes a citable source

An LLM does not watch a video. It reads its transcript. That is the starting point to understand before anything else. When a generative model cites YouTube, it is actually citing a text: the automatic or manual captions of the video, enriched by the description and chapter metadata. The video itself remains invisible to the model. Its transcript becomes a document like any other in the corpus the LLM queries.

This mechanic explains why YouTube weighs so much. The datasets used to train the large models, HowTo100M and Kinetics first among them, rest largely on YouTube transcripts and metadata. The models grew up with this platform as a reference. It is familiar to them in a way no other video platform has managed to match — not TikTok, not Vimeo, not videos hosted on a brand site.

There is a second, more structural reason. A well-produced video condenses an oral, precise, often demonstrative answer to a question someone is actually asking. A tutorial that shows how to configure a tool, an expert who walks through a method step by step, a designer who comments on their choices in real time: this format produces a source text different from a blog post. More direct, closer to spoken language, often more honest about limits and edge cases. LLMs value this kind of raw material when building an answer that sounds right.

## What makes a video actually usable by an LLM

Not all videos are equal in front of this mechanism. A video with no clean transcript, no chapters, and a three-word description remains almost invisible to a generative engine, whatever its view count. Three technical elements determine a video's real citability.

| Element | Role for AI citation |
|---|---|
| Transcript | The text the LLM actually reads. Without it, no citation is possible, regardless of content quality |
| Chapters | Let the engine point to a precise timestamp rather than the whole video |
| Structured description | Serves as a usable summary; must answer the question in the first two sentences |

The transcript comes first because without it, nothing else matters. An auto-generated transcript with poor punctuation and mangled proper names produces a low-quality source text that the model either avoids or distorts when citing it. A human pass, even a quick one, fundamentally changes the reliability of that raw material.

Chapters come next. A model able to point to the exact minute where you answer a question gains citation precision, and that precision becomes a trust criterion. A twelve-to-fifteen-minute video split into five to seven chapters of two to three minutes each offers a granularity that matches the level of detail expected in a generative answer.

The description, finally, should stop being an SEO formality stuffed with keywords. It works better treated as a short two-hundred-word article that answers the question directly in the opening lines, before detailing the content and timestamps. It is often this text, more than the full transcript, that engines synthesize first to build a short citation.

## What this changes for a B2B content strategy

The classic mistake is to treat YouTube as one social channel among others, on the same level as LinkedIn or Instagram, with a views-and-engagement objective. That reading misses the point. YouTube now functions as a knowledge base that generative engines query on the same footing as a website or technical documentation.

For a B2B company, that opens largely underused ground. A recorded webinar, a narrated product demo, a Q&A session with a client: these are formats that already exist in most marketing teams, often published with no chapters and no careful description, sometimes even without subtitles. Fixing those three points on content already produced costs little and turns hours of dormant video into citable material.

The most profitable shift is not to produce more videos. It is to take what already exists and make it readable for an LLM. A conference filmed last year, an internal tutorial never republished, a client exchange captured at an event: every piece of that catalog deserves a quick audit before anyone invests in new production. The competitive window is still wide right now, with most SEO teams still treating YouTube as a branding channel rather than a citation source.

## The traps that cancel the effort

Two mistakes come back most often. The first is leaving the automatic transcript as-is on technical content or industry jargon. YouTube transcribes acronyms, product names and specialized terms poorly. An LLM that reads a transcript riddled with errors on your own vocabulary cannot cite you correctly — and worse, it can cite a distorted version of what you said.

The second mistake is treating chapters as a YouTube constraint rather than a tool for structuring the argument. Vague chapters — "introduction", "part 1", "conclusion" — add nothing. Chapters phrased as precise questions or statements — "how to calculate the ROI of a GEO audit", "the three errors that break a transcript" — give the model a clear anchor to extract an answer and tie it to an exact timestamp.

There is a quieter trap as well: thinking that video replaces written content. It does not replace it; it complements it. A generative engine builds its answers from a set of converging sources. A well-structured video, an article that covers the same topic in depth, a clear product page: together, these formats reinforce the same authority signal on a given subject. Isolated, a perfectly optimized video remains one source among others, not a shortcut to visibility.

## What to remember before shooting the next video

YouTube is no longer one distribution channel among others for a brand thinking about its AI visibility. In a few months it has become the leading social source cited by generative engines, ahead of Reddit which had dominated until now. The change does not require a higher production budget. It requires new discipline on three simple technical elements: a clean transcript, precise chapters, and a description that answers the question before describing it.

Teams that adjust these three points on their existing catalog, before even producing new content, take a lead that compounds over time. Those that keep treating YouTube as a mere view counter leave a growing share of their AI visibility to competitors — or worse, to third parties who talk about their product without ever having been invited.

*Vurto's multi-LLM monitoring tracks your brand citations on ChatGPT, Gemini, Perplexity and Claude, YouTube included when a transcript or a comment mentions your name. So you can tell whether your videos are actually working for your AI visibility, or sitting unread.*

### [Reddit and GEO: Joining Conversations Beats Posting Into the Void](https://vurto.ai/en/resources/reddit-and-geo-joining-conversations-beats-posting-into-the-void) (2026-08-03)

A number to start with. Reddit now accounts for about 40% of citations across ChatGPT, Claude, Perplexity and Google AI Overviews combined, ahead of Wikipedia and YouTube put together. On Perplexity alone, some March 2026 studies put Reddit's share as high as 46.5% of citations. Growth over the last four months exceeds 70% in the categories tracked. No other domain comes close to this level of presence in generated answers.

For a brand that thinks of AI visibility in terms of a blog, product pages or LinkedIn posts, that number should set off an alarm. The most-cited channel by generative engines is not an editorial channel you control. It is a forum where strangers argue, contradict each other, and where your brand has historically had no organized voice. Understanding why changes the strategy from the ground up.

## Why LLMs prefer Reddit to brand content

Generative models look for answers that sound like what a competent human would tell a friend. Reddit produces exactly that kind of text, in millions of threads, on every topic imaginable. A thread on "which CRM to pick for a ten-person sales team" contains conflicting opinions, dated first-hand accounts, real-time corrections when someone gets it wrong. It is a corpus of argued disagreement, exactly what LLMs synthesize best in order to give a nuanced answer rather than a single, suspicious take.

Perplexity is especially calibrated for this. The engine looks for real people answering real questions, not pages optimized to rank. A Reddit comment with twelve upvotes and a well-built counter-argument carries more weight than a landing page claiming to be "the best solution on the market" without ever saying why. Gemini, by contrast, cites Reddit in barely 0.1% of its answers. The gap between engines matters as much as the overall trend: your Reddit strategy will not pay off the same way depending on which LLM your customers actually use.

There is a contractual reason behind these numbers as well. Google pays sixty million dollars a year to Reddit for API access to its data and to train its models on it. OpenAI signed a comparable deal, around seventy million a year. Those deals are being renegotiated during 2026, with Reddit trying to index the price to the real value its data brings to AI answers. Whatever the outcome of those negotiations: they confirm that LLMs treat Reddit as a primary source, not as an accessory forum they crawl at random.

## Publishing a post is not enough — participating changes everything

The most common mistake is to treat Reddit like LinkedIn: create a brand account, post content regularly, hope it takes. It does not work. Reddit culture rejects content that smells like promotion from ten meters away, and communities have increasingly fine-grained detection mechanisms for manufactured engagement. An account created yesterday that posts a link to its own site gets banned within hours in most serious subreddits.

What works is the opposite: reply in existing threads where your expertise has a legitimate place. Someone asks "how do we know if our site is actually readable by generative AIs" in a marketing or SEO subreddit. A precise answer, with a concrete example and a verifiable number, builds an expertise signal that neither a blog post nor an ad can reproduce. That comment becomes a potential source for an LLM that later synthesizes an answer to a similar question, with or without a direct citation of your brand.

This logic inverts the usual content marketing hierarchy. On a blog, you control the topic, the calendar, the message. On Reddit, you answer real demand, formulated by someone with a precise problem, at the moment they formulate it. It is slower to build, less predictable, but infinitely closer to what LLMs are trying to extract: a useful answer to a question asked by a human, not content produced to exist.

## The self-promotion trap and how to avoid it

Every subreddit has its own rules, usually displayed in the sidebar or a dedicated wiki. Some ban commercial links outright. Others tolerate a mention if it is clearly identified as such and brings real value beyond the link. Ignoring these rules does not just result in a post being removed. It results in the account being banned, and sometimes in an entire thread turning against the brand, which stays indexed and visible long after.

The survival rule fits in one sentence: observe before you act. Spend time in a community before speaking there, understand its tone, its codes, what is done and not done. A B2B brand that shows up in a technical subreddit with a commercial tone gets spotted immediately, and the social sanction on Reddit is fast and public.

Transparency remains the best lever available. An employee posting under their real account, stating who they are and why they are answering, lands much better than an anonymous brand account pretending to be a neutral user. Reddit tolerates assumed expertise. It does not tolerate disguise.

| Approach | Likely result | Why |
|---|---|---|
| Brand account, regular promotional posts | Fast ban | Self-promotion detection, cultural rejection |
| Identified employee, occasional useful replies | Growing credibility | Assumed expertise, real value |
| Fake user account, discreet steered reviews | Risk of backlash | Community detection, boomerang effect |
| Transparent sponsoring of an AMA or a post | Decent, controlled visibility | Accepted format if clearly disclosed |

## What this changes concretely for your GEO strategy

Reddit is not a channel to add at the end of the list after the blog and LinkedIn. Given its weight in citations, it is a channel to treat with the same rigor as any other strategic source, with a specific angle: individual human presence outweighs collective brand presence.

Concretely, that means identifying the subreddits where your customers actually ask their questions, not the ones that look relevant on the surface. It means designating one or two internal people, with real expertise, who can reply regularly without it looking like a forced marketing task. And it means tracking what happens: which Reddit threads surface in ChatGPT or Perplexity answers on your strategic queries, which comments become cited sources, and which stay invisible despite the effort.

Without that tracking, Reddit presence remains a blind bet. With it, it becomes a measurable lever, on the same footing as SEO or advertising. The difference with more classic channels is that on Reddit, authenticity is not a stylistic option. It is the only strategy that works.

*At Vurto, multi-LLM monitoring lets you see exactly when a Reddit thread or comment becomes a source cited by ChatGPT, Gemini, Perplexity or Claude on your strategic queries, so you can adjust where and how your expertise should show up.*

### [Grounding Queries: The Hidden Layer That Decides Your AI Visibility](https://vurto.ai/en/resources/grounding-queries-the-hidden-layer-that-decides-your-ai-visibility) (2026-06-19)

Fan-out queries are what the LLM sends outward to search for information. Grounding queries are different. They're the internal signal by which the model identifies what it doesn't yet know — before it even starts searching.

Understanding the difference between the two is understanding how an LLM decides what to go looking for — and therefore which content it ends up pulling.

## The Internal Mechanics of an LLM Answering a Question

When a user asks a question to an LLM with web access, the process works roughly like this.

The model first evaluates whether its internal knowledge base — what it learned during training — is sufficient to answer. If the response can be built from stable, well-documented facts, the model answers directly. If the question touches on recent information, changing data, or updated comparisons, the model identifies those gaps. Those are the grounding queries.

The model then formulates search queries to fill those gaps — those are the fan-out queries. It retrieves the results, synthesizes them, and builds the final response.

Grounding queries are the upstream layer. They determine what the model knows it doesn't know. And therefore what it goes looking for.

## Why Grounding Queries Are a GEO Lever

If an LLM considers your brand, product, or category to be part of its stable knowledge, it won't go looking for additional information. It answers from its internal base — which may be outdated or incomplete.

If the model identifies grounding queries about your brand (recent pricing, new features, current reviews), it searches. And then the quality and structure of what it finds determines what it says.

This lever cuts both ways.

Favorable case: the model looks for recent information about your brand and finds your well-structured content, up-to-date comparisons, readable product pages. The response it builds is precise and positive.

Unfavorable case: the model searches and finds an old comparison that ranks you poorly, an article mentioning a bug you fixed a year ago, or nothing structured about you at all. The response is imprecise, dated, or absent.

## The Types of Grounding Queries That Apply to Your Brand

Brand-related grounding queries generally fall into these categories.

Recent news. "Have there been any recent developments about [brand]?" The model knows its training data has a cutoff. For anything that may have changed, it searches.

Pricing and offers. Prices change. Plans evolve. Models know this and systematically look for pricing information in real time from comparison sites and pricing pages.

User reviews. Reputation evolves. Recent reviews are a frequent grounding query for purchase-intent questions.

Comparisons. "How does [brand] compare to [competitor] today?" Recently published third-party comparisons are a commonly mobilized source for these queries.

New features. If your brand launched a feature after the model's knowledge cutoff, that's a gap the model may try to fill if it knows the brand is evolving.

## How to Work on Grounding Queries

The strategy plays out on two fronts.

First: make sure the model finds recent, reliable information. That means maintaining a consistent flow of content on the subjects that grounding queries cover — pricing clearly displayed on a dedicated page, product updates documented with a clear date, recent reviews accessible and crawlable, up-to-date comparisons.

This isn't different from good SEO hygiene. But the calibration shifts: the goal isn't to rank, it's to be the source an LLM finds when it's trying to validate a piece of information.

Second: optimize the structure of that content for grounding queries. A pricing page with a clean, dated, structured table is more usable by an LLM than a pricing page with interactive cards loaded in JavaScript. A "Vurto Updates — June 2026" article with clear, dated bullet points directly answers grounding queries about recent changes.

## The Difference from Classic Content Marketing

Classic content marketing produces content to attract human readers on high-volume queries. The grounding query logic is different.

The goal isn't to drive traffic with this content. It's to make it available at the moment an LLM is trying to validate a piece of information about your brand. The traffic these pages generate doesn't matter if nobody reads them directly. What matters is that they exist, are well-structured, crawlable, recent, and aligned with what the model is looking for.

It's a logic of information stock rather than traffic flow.

## The Grounding Query Audit

A grounding query audit on a brand answers these questions.

On which topics do LLMs search for additional information when they process a query related to your brand? Which sources do they find? Are those sources recent, favorable, well-structured? Are there gaps — subjects where the model searches and finds no clear signal about your brand?

Vurto integrates grounding query analysis into its content recommendation process. The tool identifies the blind spots — topics where LLMs search and find no clear signal about your brand — and recommends which content to create or update first.

## What You Learn by Looking at Grounding Queries

The first lesson, consistent across brands: pricing. Almost always a grounding query. Almost never a pricing page that's readable by LLMs.

The second: third-party comparisons carry more weight than your own pages. When the model tries to validate an advantage or disadvantage of your brand, it prefers a source perceived as independent.

The third: new features disappear into silence if they're not documented in a crawlable format with a clear date.

Addressing these three points covers most of the grounding query work for the majority of brands.

---

*Vurto identifies the grounding queries associated with your brand and recommends content to create or update to fill detected gaps.*

### [Query Fan-Outs: The Blind Spot of 99% of Marketing Teams](https://vurto.ai/en/resources/query-fan-outs-the-blind-spot-of-99-of-marketing-teams) (2026-06-15)

Here's what happens when you ask ChatGPT "what's the best GEO tool in 2026."

The model doesn't answer directly. It first breaks your question down into a series of sub-queries it sends to its search tools — Google, Bing, or its own indexes. These sub-queries are what we call query fan-outs.

The result: it's not your well-ranked pages on "best GEO tool" that determine your presence in the response. It's the pages and sources the model finds across each of its decomposed sub-queries.

And nobody is watching that.

## Why Fan-Outs Change Everything

Traditional SEO trained you to optimize for one query, one intent, one piece of content. The mechanics of LLMs are different.

When a user asks a generative engine a complex question, the model breaks it down before answering. A question like "what type of company is Vurto right for" might generate fan-outs like:

- "Vurto GEO tool features"
- "GEO for e-commerce 2026"
- "LLM monitoring for SMBs France"
- "best tool to track ChatGPT citations"
- "AI readability e-commerce site"

The AI's final response is built from sources found across each of these sub-angles. If you're only present on the primary query but missing from the derivatives, you fall out of the final answer — partially or entirely.

## What SEO Teams Do Instead

They optimize for keywords. Keywords that humans typed into Google. Those keywords don't map to LLM fan-out queries.

That doesn't mean SEO work is useless — it builds a base of content that feeds LLMs. But stopping there means missing the specific logic of generative engines.

LLM fan-out queries have different characteristics. They tend to be longer. They focus on specific attributes, comparisons, usage conditions. They vary by LLM: ChatGPT and Perplexity don't decompose the same queries the same way.

The result: a content strategy calibrated purely on classical search intent is partly disconnected from what actually drives visibility in AI responses.

## How Fan-Outs Work

LLMs now have web search tools they activate based on the nature of the query. The exact architecture varies by model and evolves fast. But the decomposition mechanism is structural.

The model identifies information gaps in its knowledge base (what we call grounding queries — more on that in another article). It then builds a series of sub-queries to fill those gaps. Then it synthesizes the results into a coherent response.

What matters for your GEO visibility is being present on the right derivatives — not just on the initial question.

## The Content Strategy That Addresses Fan-Outs

Understanding the fan-outs in your category means identifying the sub-angles to build or strengthen content around.

Map the fan-outs for your strategic queries. Ask LLMs questions about the key topics in your category and observe how they break down their searches. Some tools let you visualize these decompositions directly — Vurto does this, identifying the fan-out queries associated with your brand and sector to calibrate content recommendations.

Produce content dedicated to specific attributes. Fan-outs often target precise dimensions: use cases, comparisons, access conditions, measurable results. A piece of content that directly answers "what's the ROI of a GEO strategy for an e-tailer" is more likely to be pulled than a generic GEO article.

Cover the multi-LLM spectrum. ChatGPT and Perplexity don't generate the same fan-outs. Perplexity leans toward journalistic sources and recent comparisons. ChatGPT favors well-structured content with strong expertise signals. The optimal content strategy covers both logics.

Be present where LLMs look. Publishers, LinkedIn, Reddit, YouTube — these platforms are increasingly mobilized as sources in fan-outs. Being present only on your own domain means playing on a single field.

## What This Changes for Content Teams

The brief changes. Instead of "an article on GEO for SMBs," the right brief becomes "a piece of content answering the question 'is GEO relevant for a company with under 50 employees and no dedicated SEO team' with concrete data and examples."

The difference is subtle. The impact on AI visibility is not.

LLMs pull content that answers a sub-question directly and precisely. Informational density counts. Clarity of the answer counts. Format counts.

## The Signal Most Teams Don't Have

Knowing which fan-outs touch your category is a tangible competitive advantage. Not because it's an esoteric technique, but because it's a signal the vast majority of competitors don't have.

In 2026, most marketing teams are still optimizing for Google keywords. That's useful. It's not enough. Brands that add fan-out logic to their content strategy are the ones starting to appear in AI answers where their competitors are absent.

Today's blind spot is tomorrow's lever.

---

*Vurto identifies the fan-out queries associated with your brand and sector to calibrate GEO content recommendations.*

### E-commerce

### [GEO and E-Commerce: Your Catalog as an AI Visibility Asset](https://vurto.ai/en/resources/geo-and-e-commerce-your-catalog-as-an-ai-visibility-asset) (2026-06-21)

E-tailers have a GEO advantage nobody tells them about.

They already have the content. Thousands of product pages with descriptions, specs, use cases. Informational material that LLMs love to draw from when answering their users' purchase questions.

The problem: this material is in the wrong format. It was built for Google and for human conversion, not for LLMs. And the gap between the two is wider than most people think.

## What LLMs Do with Purchase Queries

"What vacuum cleaner for a 1,000 sq ft home with a dog?" "Which earbuds for working out under $80?" "Best mattress for side sleepers?"

These queries land in ChatGPT, Perplexity, and Gemini in massive volume. Users want recommendations, not lists of results. And LLMs respond with recommendations — often very specific, often with justifications based on product characteristics.

Who gets cited in these recommendations? Rarely the merchants' product pages. Usually comparison sites, buying guides, review articles. Because that content is structured to answer a question, not to sell a product.

The e-tailer loses the direct recommendation and sees their traffic arrive from a third-party comparison site that talks about their product — when they're lucky.

## Why Classic Product Pages Miss LLMs

A typical product page is built like this: image carousel, price, prominent buy button, spec bullet points, short description, dynamically loaded customer reviews, similar product section.

For a human on the site, it works. For an LLM scanning the page for information, it's a nearly empty page. The carousel doesn't load. The reviews are in JavaScript. The description is short and generic. The spec bullet points exist but explain nothing about the use case.

Result: the LLM can't build a rich recommendation from this page. It looks elsewhere.

## The Markdown Conversion: What It Changes

Converting a product catalog to markdown transforms a format designed for visual rendering into one designed for text processing.

In markdown, the product page changes in nature. The title becomes a precisely named entity with its key attributes. The description becomes a prose paragraph that explains who the product suits and why. Technical specs sit in a cleanly structured table. Use cases are described in natural language. Comparisons with alternatives are stated clearly.

This version isn't visible on the site. It's accessible to LLM crawlers via a dedicated URL or through the llms.txt. It doesn't replace the classic product page — it complements it for a different audience.

The impact on AI visibility can be significant. An LLM that finds a well-structured markdown version of a product page has exactly the material it needs to formulate a precise, sourced recommendation.

## The Ten-Criteria Audit

Before rewriting everything, you need to know where the gaps are. An audit against ten LLM readability criteria gives a precise picture of the catalog's state and the right priorities.

These criteria cover: title precision, quality of prose description, presence of explicit use cases, mobilizable numerical data, named entities, review accessibility, technical spec structure, comparative positioning, alignment with llms.txt, and content readability without JavaScript.

Vurto offers this audit across an entire catalog, with per-page scoring and prioritization of which SKUs to tackle first based on commercial importance and LLM visibility potential. The tool then rewrites the pages in real time in markdown according to the identified standards, with validation of the result against each of the ten criteria.

## The Product Categories Where GEO Changes the Most

Not all categories face LLMs equally. Some are massively touched by purchase queries directed at generative engines.

Consumer electronics. Home appliances. Tools. Beauty and skincare. Sports equipment. Books and board games. Any category where users need to compare before buying and where technical criteria matter — that's where LLMs get queried.

In these categories, not being cited in AI recommendations means not existing in a growing share of the purchase journey.

Conversely, in impulse-buy or commodity categories, the impact is less direct. The person buying socks or printer paper isn't asking Perplexity for advice.

## The Right Way to Start

Not by rewriting the entire catalog at once. Start with the 50 to 100 most commercially important SKUs and check their LLM readability score. Often 20% of SKUs drive 80% of traffic. Those are the ones to tackle first.

Then audit the site's technical structure to verify that these pages — even well rewritten in markdown — are actually accessible to LLM crawlers. Perfect content blocked by robots.txt is perfectly useless content.

GEO for e-commerce isn't a revolution. It's a logical extension of existing SEO and content work. But it requires shifting the reference frame: thinking "LLM readability" alongside "human conversion," and producing formats that serve both.

---

*Vurto analyzes product pages against ten criteria and converts them to structured markdown to maximize their presence in LLM recommendations.*

### [Product Pages and AI: Why Your Catalog Is Unreadable](https://vurto.ai/en/resources/product-pages-and-ai-why-your-catalog-is-unreadable) (2026-06-16)

A product page was designed for two readers: the human consumer and Google's bot.

The human looks at the photo, the price, the reviews, the name. Google's bot indexes the title, the meta, the Schema.org structured data. Thirty years of e-commerce refined this alchemy. The product pages of major online retailers are conversion and ranking machines.

They're often unreadable to LLMs.

## What LLMs Read — and What They Ignore

Language models process text. Not JavaScript. Not image carousels. Not tabs hidden behind lazy loading. Not descriptions buried under a "read more" that requires a click.

The typical e-commerce product page structure looks like this: an H1 title, a few bullet points of specs, a condensed description block, visuals, a price, a customer reviews section. For an LLM scanning the page, the useful signal is often buried in HTML noise.

The direct consequence: when a user asks ChatGPT or Perplexity "what's the best [product] for [use case]," the model pulls sources that discuss those products in a dense and structured way — comparisons, expert reviews, buying guides. Rarely the merchant's official product page.

The e-tailer loses the recommendation. Sometimes to a comparison site or media outlet that talks about their product better than they do.

## The Ten Criteria That Make a Product Page Readable for AI

There's nothing mysterious about it. But it requires an honest audit.

**1. Descriptive, precise title.** The title must contain key attributes: brand, model, main feature, target use case. A vague title like "Men's sports shoe" is useless for an LLM trying to answer "what running shoe for a beginner trail runner under $100."

**2. Structured prose description.** Bullet points listing technical specs aren't enough. A written description explaining who the product suits, in what context, and why — that's what the LLM can work with.

**3. Explicit use cases.** "Ideal for..." isn't a use case. "Designed for 2-to-5-day mountain hikes with a load under 33 lbs" is a use case the LLM can actually mobilize.

**4. Honest comparisons.** Product pages carefully avoid comparing themselves to competitors. LLMs love comparisons. A page that clearly positions the product against alternatives — without disparagement — gives models the material to answer comparative questions.

**5. Measurable data.** Weight, dimensions, battery life, yield, capacity, warranty length. LLMs pull numerical data in comparative queries. A purely qualitative description sends the model looking elsewhere — often to a competitor.

**6. Properly named entities.** Brand, technology, material, certification, partner — these entities must appear explicitly in the text, not just in metadata.

**7. Content accessible without JavaScript execution.** If the main description loads in JS, the LLM crawler doesn't see it. Testing a product page with a text-only crawler routinely surfaces surprises.

**8. No generic duplicated content.** Descriptions copied from the supplier or mass-generated without differentiation read as noise. The LLM prefers a source with a distinct angle.

**9. Reviews integrated and accessible.** Customer reviews often contain the most useful phrasing for an LLM ("perfect for wide feet," "holds well on descents," "great for beginners"). If they load dynamically and are invisible to crawlers, that's lost signal.

**10. Links to contextual content.** An isolated product page is a weak signal. A product page linked to a buying guide, a comparison, a blog post — that's a node in a content network that LLMs value more.

## The Specific Challenge of E-Commerce Catalogs

The problem retailers have with LLMs isn't just one badly structured page. It's scale.

A catalog with 5,000 SKUs means 5,000 product pages to reassess through the LLM lens. Manual rewriting is impossible. Generic AI rewriting, without calibration to LLM readability standards, produces text that looks readable to a human but remains poorly structured for generative engines.

This is exactly what Vurto addresses for e-tailers: a real-time catalog rewrite in markdown format, structured according to LLM readability criteria, with each product page scored against ten dimensions before and after rewriting. The output isn't a generic version of the pages — it's a version calibrated so that LLMs can process, understand, and recommend them.

## What This Changes for Sales

Measuring the direct impact of an LLM-ready product page on revenue is tricky, because traffic from AI engines is still partially opaque. But the signals are there.

Direct referrals from ChatGPT and Perplexity are now trackable via UTMs and referrers. Brands watching this traffic see that it converts well — the user arrives with a qualified intent, not mere curiosity. They've already received a recommendation from the AI before landing on the page.

Being the cited reference in that recommendation means capturing high-intent traffic that traditional advertising doesn't reach.

## The Five-Minute Test

Open ChatGPT. Ask: "[your flagship product name]: does it work for [your customers' main use case]?"

Watch the response. If the model cites your product page or your domain — good. If it cites an external comparison, a press article, or worse, a competitor — you know what needs fixing.

---

*Vurto analyzes product pages against ten AI readability criteria and rewrites them in structured markdown to maximize their presence in LLM recommendations.*

### GEO Technical

### [Schema.org in 2026: Which Structured Data Still Matters](https://vurto.ai/en/resources/schema-org-in-2026-which-structured-data-still-matters) (2026-08-24)

Ahrefs tested the theory on 1,885 pages in 2026. Add schema.org markup, measure the effect on citations in AI Overviews. Result: no improvement. Citations even dropped 4.6%. On the other side, 71% of pages cited by ChatGPT contain structured data, and Google itself confirms that schema helps models understand a page's content. Two serious studies, two conclusions that contradict each other. This is exactly the kind of confusion that pushes a marketing team either to dump JSON-LD everywhere as a precaution, or to drop the topic altogether, assuming it no longer matters now that classic SEO has changed nature. Both reactions are mistakes.

## The myth that refuses to die

The misunderstanding comes from confusing correlation with cause. Pages cited by LLMs often have schema, that is true. But they also often have a clear content structure, direct answers, and established domain authority. Schema accompanies those qualities; it does not produce them. A product page stuffed with microdata on a slow site, with vague copy and zero third-party proof, will not get cited any more than before. Google says it in black and white in its documentation: no special files, no specific schema.org markup required to appear in AI features. That is not an invitation to ignore the topic. It is a reminder that markup is not a shortcut to visibility, contrary to what part of the GEO industry claims when it sells schema as a magic formula.

The Ahrefs test confirms this reading without invalidating it entirely. Across 1,885 pages, adding generic, poorly targeted schema on content that already failed to answer a precise question changed nothing. That is consistent. An LLM does not cite a page because it wears a JSON-LD label. It cites a page because it contains the clearest, best-formulated, most verifiable answer to the question asked. Schema helps the model extract that answer faster and with less ambiguity. It does not create the answer.

## What actually works, and what is useless

Not all schema types are equal in front of generative engines. FAQPage comes first, and by a wide margin, because it structures content into self-contained question-answer pairs, exactly the format an LLM can extract and cite without rewriting. Combining FAQPage with Article and HowTo produces nearly twice as many citations as Article alone. The logic is simple: the closer the markup brings the page structure to the format of a conversational answer, the more it eases the model's work.

| Schema type | Real usefulness for GEO in 2026 |
|---|---|
| FAQPage | Strong. Question-answer format directly usable by LLMs |
| HowTo | Strong, especially combined with Article or FAQPage |
| Product | Useful for e-commerce, verifiable price and availability |
| Organization | Useful for entity clarity — who you are, what you do |
| Review / AggregateRating | Useful as third-party proof, provided it is authentic |
| BreadcrumbList | Weak direct impact on citations |
| Generic Article without structured content | Almost none |

The table reads in a precise direction. Markup that describes a content structure already designed to answer a question has a real effect. Markup that merely decorates an existing page without changing its editorial structure has a marginal effect, or none. It is the difference between putting a label on a dish that is already well prepared, and sticking a label on an empty plate hoping it will fill itself.

Product schema deserves a separate mention for e-commerce. Price, availability, verified reviews: this structured data lets an LLM answer a transactional question without hallucinating an outdated price. A product page without Product schema forces the model to guess or to cite a third-party source, often a comparison site that has no obligation to put you forward.

## The invisible infrastructure nobody sees running

The real role of schema in 2026 is not direct citation. It is entity clarity. An LLM that encounters your site must quickly understand who you are, what you sell, where you operate, and what your relationship is with the entities mentioned in your content. Organization schema, with the right fields filled in (legal name, industry, service area, links to social profiles and the knowledge base), reduces ambiguity on that point. It guarantees no citation. It does, however, prevent a model from confusing your brand with a namesake, or from wrongly attributing to you information that belongs to a competitor with a similar name.

This infrastructure function explains why Microsoft and Google confirm that schema helps models understand content, without ever promising that it triggers a citation. Schema builds the context in which your content is interpreted. It never replaces the quality of that content. A company that thinks of GEO only in terms of technical markup makes the classic mistake of confusing the scaffolding with the building.

There is a second, more discreet layer to this. LLMs build internal knowledge graphs from what they crawl. A site with consistent markup across all its pages, entities correctly linked to each other, a clear hierarchy between organization, products and editorial content, becomes easier to integrate into that graph. A site where each page uses different, incomplete, sometimes contradictory markup makes that integration harder. The result is not visible in an immediate citation metric. It shows over time, in the stability and precision of what models know how to say about you.

## Prioritize without drowning in markup

The question to ask is never "how many schema types can I add" but "which pages have content structured enough for the markup to be useful". A poorly written FAQ page, with vague answers, will gain nothing from carrying FAQPage markup. You first have to rewrite the answers so they are self-contained, verifiable, directly citable, then add the markup that makes them easy to extract.

The priority order that makes sense in 2026 comes down to three moves. First, Organization schema on institutional pages, to settle once and for all who you are in the eyes of the models. Then, FAQPage and HowTo on content that already answers precise, frequent questions — the ones your fan-out queries reveal. Finally, Product schema on pages that carry a price and availability, if you are in e-commerce. Everything else — BreadcrumbList, decorative markup on generic content pages — can wait or be ignored without measurable consequence.

Schema.org in 2026 is no longer a magic lever, and it never really was. It is an accelerator of understanding for content that already deserves to be understood. On weak content, it catches nothing up. On strong content, it removes friction between your page and the answer a model will formulate from it.

*Vurto scores your product pages and key pages against ten AI-readability criteria, markup included, to spot where schema actually helps and where it is only decorating a page that needs something else.*

### [LinkedIn and GEO: Why LLMs Love This Platform (and How to Use It)](https://vurto.ai/en/resources/linkedin-and-geo-why-llms-love-this-platform-and-how-to-use-it) (2026-06-23)

## What LLMs Look for in a Source

A large language model, when generating a response, doesn't browse the web the way a human does. It draws on two things: its training memory (the billions of documents ingested before its cutoff date) and, for models with web access, a real-time search layer that retrieves and synthesizes fresh sources.

In both cases, a source's perceived quality depends on specific signals:

- Semantic density: does the content say something precise, or is it filler?
- Factual consistency with other sources: if 12 documents converge on the same information, the model treats it as reliable.
- Clear attribution: who is speaking, in what capacity, on what subject.
- Readable structure: clean HTML, explicit headings, continuous prose.

LinkedIn meets these criteria better than almost any social network. A well-written LinkedIn post cites a figure, names a company, identifies an author with a title and experience. That's exactly what an LLM wants to anchor a factual response.

## Why LinkedIn Survives Training Better

Large language models were trained on Common Crawl, Reddit dumps, Wikipedia, books, and press articles. LinkedIn is part of these corpora, in various forms depending on licensing agreements.

What matters is that LinkedIn is a platform built around professional identity. Every piece of content is attached to a profile: a name, a job title, a company, a sector. This richness of context makes each post a micro-document dense with attribution signals. "Marie Dupont, Marketing Director at X, writes about B2B content strategy" is more useful to an LLM than an anonymous article published on a dateless, authorless blog.

LinkedIn articles — the long-form feature, distinct from short posts — have a stable URL, a structured title, content that tends to be longer and better referenced. They resemble blog posts, but with a LinkedIn identity layered on top. For a model trying to attribute an opinion or expertise, that's ideal.

The profile itself is a source. When an LLM answers "who is the French reference on marketing automation," it can synthesize information from LinkedIn profiles, posts, and articles published on the platform. Your LinkedIn presence isn't just a digital CV. It's a node in the knowledge graph that LLMs draw from.

## What Makes LinkedIn Content Citable

Not all LinkedIn posts are equal from an LLM's perspective. A post with five emojis, a clickbait hook, and a "tell me in the comments" has little chance of being cited by a model answering a serious question. What LLMs retain is high-information-density content.

A few concrete criteria:

**Sourced figures.** A post that says "our conversion rate increased 34% after restructuring our landing pages according to these three principles" is citable. A post that says "here are 5 tips to convert better" isn't really — too generic, too diluted.

**Clear positions.** LLMs synthesize expert opinions. If your content asserts something specific, on a given subject, with your name and title attached, it becomes a potential source for answering "what do experts think about [topic]?"

**Exact industry vocabulary.** Models recognize technical language. A post on GEO that uses the right terms (fan-out queries, grounding, llms.txt, share of voice) will be better understood and better semantically indexed than a post that paraphrases the same concepts in plain language.

**Consistency.** A profile that publishes weekly within a coherent thematic perimeter builds a strong semantic footprint. The model recognizes it as a specialized source on that subject. A profile that publishes on ten different topics generates noise.

**Useful length.** Short posts (under 150 words) perform poorly unless they contain a standout fact. Long-form articles are more easily ingested as documentary sources. The ideal GEO format is probably the mid-length post (300–600 words) with clear structure and actionable information.

## The LinkedIn Strategy for Corporate GEO

For a brand, LinkedIn plays a dual role in GEO strategy.

First role: amplifying the presence of your experts. When your directors, consultants, and product managers publish on LinkedIn using your brand name and product vocabulary, they create distributed occurrences that strengthen your signal in LLMs. One company blog versus ten active profiles: the ten profiles win on volume and signal diversity.

Second role: occupying comparison queries. LLMs often answer questions like "what's the best solution for [problem]" by synthesizing expert opinions and experience reports. LinkedIn is a natural source for this type of content. If your clients publish positive feedback on LinkedIn mentioning your product by name, those posts can end up in the LLM synthesis of a comparison question.

The distinction matters: this isn't about publishing promotional posts. LLMs don't cite disguised ads. They cite expert opinions grounded in concrete experience, data, and precise examples.

## What LinkedIn Can't Do Alone

LinkedIn is a powerful lever, but in isolation it's not enough. LLMs build their responses through triangulation. If your brand appears on LinkedIn but nowhere else — no press coverage, no in-depth articles on your site, no mentions in specialized forums like Reddit or industry communities — the signal is fragile.

The most robust GEO strategy combines multiple surfaces: your site (with content readable by AI crawlers), your LinkedIn presence, mentions in third-party publications, and ideally a presence in community discussion spaces where LLMs look for authentic opinions.

LinkedIn is a central piece of the setup, not the whole setup. What changes compared to classic SEO is that each surface now has direct value for your LLM visibility, independently of backlinks. An excellent LinkedIn post with no backlinks can be cited by ChatGPT. In traditional SEO, it would have carried little weight.

## What You Should Do This Week

Audit your LinkedIn presence through the GEO lens. Ask yourself three simple questions.

Does your profile describe your expertise clearly, using the exact vocabulary of your sector? LLMs read titles, summaries, and experience sections. If your profile uses vague language ("helping companies navigate their transformation"), it won't be cited as a reference on a specific topic.

Do your recent posts contain citable factual information? Figures, examples, clear positions on a subject. Or is it surface content designed to generate likes?

Is your company mentioned in your colleagues' and clients' posts in a substantive way, with precise details about what you do and for whom?

If the answer is no on all three points, you have a straightforward, fast opportunity to improve your GEO signal without touching your site or launching a complex content campaign.

---

*Vurto monitors your presence in ChatGPT, Gemini, Perplexity, and Claude in real time. If you want to know whether your brand or your LinkedIn experts are being cited by LLMs today — and in what context — that's exactly what Vurto measures.*

### [Your Site Is Invisible to AI — and You Don't Know It](https://vurto.ai/en/resources/your-site-is-invisible-to-ai-and-you-don-t-know-it) (2026-06-20)

There's a difference between a site Google can index and a site an LLM can read.

Google sends a crawler that retrieves HTML, follows links, logs tags. It's a well-documented mechanism. Thousands of tools audit it. Millions of developers optimize for it.

LLMs work differently. When a model goes looking for information on your site, it doesn't navigate. It scans. It extracts text. It looks for structured meaning. And if what it finds is opaque, fragmented, buried in JavaScript or poorly tagged HTML, it moves on.

The result: sites that are perfectly optimized for Google — well-ranked, fast-loading, green Core Web Vitals scores — that are nearly nonexistent from an LLM's perspective.

## What "Readable for AI" Actually Means

It's not a question of technical performance. It's not meta-descriptions or alt tags either.

AI readability is the capacity of a page's content to be extracted, understood, and mobilized by a language model when answering a user's question.

It starts with a simple question: if you strip all the JavaScript from your page and keep only the raw HTML, what readable text is left? On many modern e-commerce or SaaS sites, the answer is: not much. The main content is rendered dynamically. Descriptions hide behind accordions. Customer reviews come in via a third-party API. Prices appear after a JavaScript call.

For an LLM, none of that exists.

## The Most Common Blind Spots

The first blind spot is JavaScript-rendered content. An LLM scanning a product page and finding only an empty HTML shell — because all the content is rendered client-side — will simply ignore the page. Modern frontend frameworks are great for user experience. For LLM readability, they're often a disaster.

The second is invisible structure. Dense content without clear hierarchy, without well-positioned headings, without a readable progression — an LLM struggles to use it. It doesn't read like a human. It looks for structural signals to identify what matters.

The third is the absence of markdown. LLMs process markdown natively. Well-structured markdown content is more readily usable than equivalent content in rich HTML. For strategic pages — product pages, comparison pages, buying guides — having a markdown-accessible version is a direct advantage.

The fourth is a poorly calibrated robots.txt. Blocking certain bots without realizing they're precisely the crawlers used by LLMs. GPTBot, ClaudeBot, Googlebot AI Mode, PerplexityBot — agents that can end up blocked by robots.txt rules copy-pasted from templates that predate their existence.

## The AI Readability Audit in Practice

The audit starts with a simple test: access each of your strategic pages via a text-only browser or a JavaScript-free scraper and observe what's left. It's usually revealing.

Then check the coherence between robots.txt and your GEO strategy. If certain crawlers are blocked, is it intentional? If so, which LLMs are affected? Is the impact on visibility accepted?

Check the status of the llms.txt. Does it exist? Is it consistent with the site's actual content? Does it point to the right pages?

Analyze the informational density of key pages. A product page, a comparison page, a pricing page — do they have enough textual substance for an LLM to extract something useful?

Vurto offers a dedicated tool for this diagnosis: a site-wide AI readability check, page by page, that identifies technical blockers, content invisible to crawlers, robots.txt/llms.txt inconsistencies, and pages to prioritize for rewriting or restructuring. It's often the first tool web teams open when they start working on GEO.

## What This Changes for Technical Teams

GEO isn't just a marketing topic. It has direct technical implications.

Server-side rendering or partial server-side hydration for strategic content. Static markdown files for product pages or guides. A robots.txt audit to verify LLM crawler access. llms.txt implementation. Up-to-date and coherent Schema.org structured data.

None of these actions are complicated. But they require coordination between SEO, tech, and content teams that doesn't yet exist in most organizations.

The problem isn't technical. The problem is that nobody has clearly asked the question yet: are our strategic pages readable by LLMs? And in the vast majority of cases, the honest answer would be: not really.

---

*Vurto checks your site's AI readability: content invisible to crawlers, robots.txt, llms.txt, and the structure of strategic pages.*

### [llms.txt: The robots.txt of the AI Era](https://vurto.ai/en/resources/llms-txt-the-robots-txt-of-the-ai-era) (2026-06-17)

In 1994, site owners started placing a `robots.txt` file at the root of their domains. It told web crawlers what to index and what to ignore. Nobody forced them. It became a universal standard because it solved a real problem: giving bots the instructions they needed to do their job properly.

Thirty years later, a new file is emerging with the same logic: `llms.txt`.

Its purpose? Give language models the essential information about your site, structured in a way they can process efficiently — without having to piece it together themselves.

## What llms.txt Does

The `llms.txt` file lives at the root of a domain, just like `robots.txt`. It contains a structured description of the site — in markdown, readable by LLMs — with key information about the entity, its products, its values, its trusted sources, and its strategic pages.

It can include things like:

- What the company does, in one precise sentence
- Main products or services and their characteristics
- Target audience
- The most important pages to know about
- Authors or experts associated with the site
- Links to third-party sources that mention the site (press, studies, comparisons)

The principle: rather than letting an LLM reconstruct an incomplete or distorted image of your brand from scattered fragments, you give it a structured, verifiable, coherent information base to work from.

## Why It Matters for GEO

LLMs have two ways of knowing a brand.

The first is training. Everything ingested before the model's knowledge cutoff. That's where well-known brands live — established facts, information repeated across many sources. For recent, niche, or poorly represented brands, this layer is partial or inaccurate.

The second is real-time search. When the model has access to search tools, it goes looking for current information. And what it finds first are the best-structured, most easily parseable sources.

`llms.txt` plays on both levels. It improves the quality of information available to the model in real time, and it structures that information in a way that maximizes the likelihood the model uses it in its responses.

## What a Poorly Written llms.txt Does (or Doesn't Do)

An empty or generic llms.txt is a missed opportunity. An auto-generated llms.txt without calibration is sometimes worse — it can contain vague descriptions that reinforce an inaccurate image.

The common mistakes:

Description too vague. "Innovative company in the digital sector" tells an LLM nothing. "GEO platform that tracks brand citations in ChatGPT, Gemini, and Perplexity and recommends content actions based on fan-out queries" says a great deal.

No named entities. LLMs operate on entities — names, products, technologies, locations, people. An llms.txt without clear entities is barely usable.

Missing third-party sources. Pointing to press articles, comparisons, and studies mentioning your brand reinforces the credibility of the information in the model's eyes.

Misalignment with the rest of the site's content. If the llms.txt says one thing and the site's pages say another, the model arbitrates based on the strongest signals — not necessarily in your favor.

Contradictory robots.txt. Blocking AI crawlers in robots.txt while maintaining a llms.txt is incoherent. Both files need to be aligned.

## llms.txt and robots.txt: Coherence as a Condition

The llms.txt doesn't work in isolation. It connects with robots.txt, with Schema.org structured data, with the overall quality of the site's content.

The coherence between robots.txt and llms.txt is a point often neglected. Some sites block AI crawlers in robots.txt by instinct — data protection reflex — without realizing it completely neutralizes the llms.txt. Others allow crawling but have an llms.txt that contradicts their product pages.

Vurto includes a dedicated module for auditing and rewriting both files together — robots.txt and llms.txt — to ensure the instructions given to models are coherent, complete, and aligned with the overall GEO strategy. It's one of the starting points of any serious GEO audit.

## How to Write an Effective llms.txt

The recommended structure is simple:

```markdown
# [Company Name]

[Precise description in 2-3 sentences: what you do, for whom, with what difference]

## Products / Services

- [Product 1]: [functional description in one sentence]
- [Product 2]: [functional description in one sentence]

## Target Audience

[Description of the primary audience: sector, size, role, problem solved]

## Key Pages

- [URL page 1]: [what's there]
- [URL page 2]: [what's there]

## Reference Sources

- [Press article link]
- [Comparison link]
- [Study citing the brand]
```

This isn't a fixed format — the standard is evolving fast. But the underlying logic is stable: precise information, named entities, links to third-party sources, coherence with the rest of the site.

## Where the Market Stands in 2026

llms.txt adoption is growing. Among the most GEO-advanced sites, it's become a baseline reflex, just like robots.txt or a sitemap.

For brands that haven't done it yet, it's one of the GEO actions with the best effort-to-impact ratio. A few hours of work, a well-written file, a robots.txt coherence check — and you give LLMs a reliable information base about your brand that they didn't have before.

In a world where generative engines build their responses from whatever signals are available, controlling those signals at the source is basic digital hygiene.

---

*Vurto audits and rewrites llms.txt and robots.txt files to ensure coherence and maximize your site's readability for LLMs.*

### Brand & Reputation

### [Brand Perception in AI: What ChatGPT Really Says About You](https://vurto.ai/en/resources/brand-perception-in-ai-what-chatgpt-really-says-about-you) (2026-06-18)

Ask ChatGPT, Perplexity, and Gemini this question today: "[Your brand name]: what are the pros and cons?"

The answers will surprise you. Sometimes in a good way. Often not.

What LLMs say about your brand is the result of an aggregation of sources: press articles, customer reviews, forums, comparisons, product pages, LinkedIn posts, industry studies. The model synthesizes all of it into a view of your brand — and that view can diverge significantly from your official positioning.

That's AI brand perception. And it's a signal almost nobody is tracking right now.

## Why AI Brand Perception Is Different from Classic Reputation Management

Online reputation management has been around for twenty years. Google Alerts, social mentions, platform reviews. Mature tools for a well-understood problem.

Brand perception in LLMs is different for several reasons.

LLMs aggregate and synthesize. They don't show you a list of mentions — they give you a consolidated view. And in that consolidation, some signals carry more weight than others, following an opaque logic.

LLMs have partial memory. What a model knows about your brand depends on its training cutoff, the sources ingested during training, and what it finds in real-time searches. A negative event from two years ago can still weigh heavily if the sources covering it are well-indexed. A recent positive story may be absent if it hasn't been crawled yet.

LLMs aren't consistent with each other. ChatGPT, Gemini, Perplexity, and Claude don't share the same sources or the same response architectures. Your brand perception can be positive in one and neutral or negative in another.

AI brand perception influences purchase decisions. That's what makes this strategically important. When a prospect asks an AI about your brand before buying — and that's a rapidly growing behavior — what they receive shapes their decision. Not in the same way as a Trustpilot review, but potentially in a more synthetic, more authoritative way.

## The Dimensions of AI Brand Perception

Perception isn't a single dimension. It breaks down into several layers.

Overall sentiment. Do AIs talk about your brand in positive, neutral, or negative terms? This is the most visible layer and the easiest to communicate to leadership.

Associated attributes. When LLMs describe your brand, what adjectives and qualifiers do they use? "Reliable," "premium," "accessible," "innovative," "hard to onboard" — these attributes build an image that influences prospects.

Attributed use cases. Who do AIs recommend your brand to? A tech startup? An enterprise? A beginner? LLMs often have fixed ideas about the use cases of brands, sometimes out of step with your actual commercial reality.

Comparisons. When you're compared to a competitor in an AI response, do you come out ahead, behind, or even? And on which criteria?

Sources mobilized. Which articles, comparisons, and sites do LLMs cite when they talk about your brand? These sources shape perception and may contain outdated or inaccurate information.

## The Common Gaps Between Real Image and AI Image

In practice, the most frequent gaps involve:

Target positioning. A B2B mid-market brand may be described as an SMB solution because the comparisons mentioning it target smaller businesses. The AI copies the bias of its sources.

Features. Recent features launched after the model's cutoff or poorly documented publicly are absent from the description. The brand is presented with its capabilities from eighteen months ago.

Pricing. LLMs often cite outdated price ranges pulled from old comparisons. For brands whose pricing has evolved, this directly affects conversion.

Friction points. Old negative feedback on a specific issue — customer service, a product transition period, a bug fixed long ago — can continue to weigh on AI perception if the sources covering it remain well-indexed.

## How to Correct AI Brand Perception

The fix isn't magic. It takes work on the sources.

Identify which sources carry the most weight. Which articles and comparisons do LLMs most often cite when covering your brand? Those are the sources to prioritize: update them, complement them, or counterbalance them with new ones.

Produce content that challenges incorrect attributes. If AIs associate your brand with a use case you no longer target, create explicit content about your current audience — on your own site and with third-party publishers. Models update their perception as they crawl.

Work on the density of positive named entities. LLMs pull named entities to build their responses. Content that explicitly associates your brand with positive attributes ("fast deployment," "measurable ROI in 30 days," "works for teams without dedicated SEO resources") feeds perception in the right direction.

Monitor continuously, not once. AI brand perception evolves. A minimum monthly check is needed to catch drift before it becomes entrenched.

Vurto integrates this dimension directly into its dashboard: brand sentiment measurement across LLMs, identification of associated attributes, tracking of mobilized sources, and alerts when perception diverges from the defined positioning. It's the monitoring tool that classic brand reputation management doesn't offer yet.

## The Signal No Marketing Team Should Ignore

What AIs say about your brand is what millions of users receive as their first answer about you.

Not an ad. Not a sales pitch. A synthesis that the model presents as factual information.

Ignoring it means letting others shape your image — your past sources, your old reviews, your better-AI-referenced competitors — without ever taking back control.

The good news: this perception is correctable. It takes time, content, and method. But it's not set in stone.

---

*Vurto measures brand perception across LLMs: sentiment, associated attributes, mobilized sources, competitor comparisons.*

### Monitoring & Analytics

### [Measuring GEO ROI: The Question That Blocks Every Budget](https://vurto.ai/en/resources/measuring-geo-roi-the-question-that-blocks-every-budget) (2026-09-21)

70.6%. That is the share of traffic generated by AI chatbots that arrives on a site with no origin trace at all. No "referred from ChatGPT", no "referred from Perplexity". Nothing. Google Analytics dumps it all into the "Direct" bucket, the same one used for a visitor who typed the site address from memory. A GEO budget (optimizing visibility in chatbot answers from ChatGPT, Claude, or Gemini, as opposed to classic Google search ranking) can therefore produce real results and still remain invisible in the dashboard used to justify it.

That is the heart of the problem. A marketing team invests in GEO, gets citations in AI answers, sometimes sees traffic rise, but cannot answer the simplest question a CFO will ask: how much does it return. Not because GEO returns nothing. Because measurement tools were built for a world where people clicked blue links, not for a world where they get a direct answer inside a conversation window.

## Why the click is no longer enough

Return on investment (the calculation that compares what an action cost to what it returned) has always relied on a chain of trackable clicks. A user clicks a Google result, lands on a page, fills a form or buys a product. Each step leaves a trace, each trace feeds a dashboard.

That chain breaks with AI chatbots, for three precise technical reasons. The mobile apps for ChatGPT or Claude do not pass origin information when they open a link. Many users copy-paste an address straight from the chatbot answer instead of clicking, which erases any provenance trail. And ChatGPT Plus, like Google's AI mode, deliberately adds a technical instruction that prevents destination sites from knowing where the visitor came from.

As a result, a site can see its "Direct" traffic double without anyone on the team understanding why. The answer is almost always the same: part of that supposedly originless traffic actually comes from AI chatbots. And this ghost traffic is not anecdotal. It converts at 10.2%, versus 2.46% for non-AI traffic classified elsewhere. A visitor from a chatbot buys or fills a form four times more often than an average visitor, but nobody knows because nobody sees it.

Before trying to calculate ROI, the first task is therefore to repair measurement itself. A channel grouping that manually recognizes known AI domains (chatgpt.com, perplexity.ai, claude.ai and the like) recovers part of the identifiable traffic. For the rest, the traffic that arrives with no trace at all, only server-log analysis, more tedious but more reliable than classic analytics tools, can rebuild a faithful picture.

## The metrics that replace the click

Once measurement is fixed, a deeper problem remains. Even when well measured, the click tells only a small part of the GEO story. A majority of Google searches now end with no click at all, and that share rises sharply whenever an AI-generated summary appears at the top of the results page. A user who asks ChatGPT a question and gets a satisfying answer, citing the brand along the way, often has no reason to click anywhere. The brand was still seen, named, recommended. That value does not land in any clicks column.

Substitution metrics are then needed, ones that measure presence rather than passage. The first is citation frequency: on a sample of queries representative of the sector, in how many answers does the brand appear, and at what rank in the source list. The second is share of voice, which compares that frequency to competitors on the same queries. The third is the sentiment attached to the citation: an answer that mentions the brand in positive terms does not weigh the same as a neutral mention in the middle of a list, or worse, a mention paired with a caveat about quality or price. The fourth is the nature of the sources the chatbot goes looking for to build its answer, because a brand absent from the pages the AI consults to form an opinion has no chance of being cited, whatever its budget.

These four metrics do not replace the financial calculation. They build the base without which that calculation means nothing. A GEO budget that moves citation frequency and share of voice forward, month after month, produces an effect that will eventually translate into traffic and sales, even if the exact path between the two stays blurry. Conversely, a budget that moves none of these four numbers will never produce a return, no matter how good the published content is.

## What a chatbot visitor is really worth

The most counterintuitive point in this whole topic is that AI traffic, when measured correctly, converts clearly better than classic traffic. Several studies published between late 2025 and 2026 agree on an order of magnitude: between four and five times the average conversion rate of classic organic search, with gaps ranging from 1.3x on impulsive online purchases up to more than twenty times in some B2B sectors where the buying decision is long and deliberate.

| Traffic origin | Average conversion rate |
|---|---|
| Classic Google search | 1.76% |
| Claude | 5.0% |
| Perplexity | 10.5% |
| ChatGPT | 15.9% |

The logic behind these numbers is simple once you see it. A user who clicks a Google result is still comparing, opens several tabs, hesitates. A user who already queried a chatbot has gotten a synthesis, asked follow-up questions, eliminated some options. When they finally click through to a site, it is often because they are ready to act. The chatbot did part of the qualification work upstream.

This data changes how a GEO budget must be read. A lower volume of AI traffic than classic Google volume is not necessarily a failure, if that rarer traffic converts four times better. The right question is not "how many visitors did GEO bring", it is "how many sales or leads did this traffic, even reduced, produce". It is a complete reversal of how classic search ranking learned to think for twenty years, when volume was king.

## Building a formula that holds up

Without waiting for a perfect tool that does not exist yet, a company can build a reasonable GEO ROI measure on four pillars.

The first is a starting snapshot, taken before any investment: current citation level, share of voice versus competitors, dominant sentiment. Without that baseline, there is no way to know whether the efforts produce an effect.

The second is regular tracking of those same visibility metrics, month after month, to objectify progress rather than relying on a feeling.

The third is reconstituting traffic that actually came from chatbots, via a correctly configured channel grouping and, when possible, server-log analysis to recover what escapes standard tools.

The fourth is linking that reconstituted traffic to the conversions it produces, keeping in mind that its conversion rate will likely be higher than other channels, not lower.

Once these four building blocks are in place, the calculation becomes: the value of sales or leads attributable to that traffic, minus what GEO cost, all relative to the starting investment. It is not exact science. It is a considerable improvement over the total absence of measurement, which remains the norm today in most companies that are already investing in this topic.

The temptation, faced with this complexity, is to wait for analytics tools to catch up before acting. That is the inverse error of spending without ever measuring. Companies that start today tracking their citation frequency and share of voice are building a history. Those that wait will start in a year with no baseline at all, at the exact moment their competitors already have twelve months of it.

*This is exactly what Vurto does: track citation frequency, share of voice, and brand sentiment across ChatGPT, Gemini, Perplexity, and Claude, month after month, to give marketing teams the measurement base that most GEO budgets still lack.*

### [AI Chatbot Traffic Is Tiny. It's Worth Gold](https://vurto.ai/en/resources/ai-chatbot-traffic-is-tiny-its-worth-gold) (2026-09-10)

A US agency, Seer Interactive, compared two sources of visitors on the same site: those arriving from ChatGPT, and those arriving from a classic Google search. Result: 15.9% of the former end up buying or filling out a form, versus 1.76% of the latter. Almost nine times more. Microsoft, using its Clarity analytics tool, observed the same thing across more than 1,200 media and publisher sites: visitors coming from a chatbot sign up eleven times more often than those coming from a classic Google search. A study of 94 e-commerce sites measured a more modest but real gap: 31% more conversions for ChatGPT traffic compared to a Google search that doesn't include the brand name.

These numbers vary from study to study, sometimes a lot, because the calculation methods differ. But they all point in the same direction. Traffic sent by an AI chatbot converts noticeably better than classic traffic. The problem is that it accounts for barely more than 1% of a site's total traffic, according to measurements from the analytics tool Conductor. A marketing team that only looks at volume misses the most useful piece of information on its dashboard.

## Why a visitor from a chatbot buys more often

The answer lies in how the question was asked before the person even arrived on the site. When someone types "best invoicing software for freelancers" into Google, they get ten links and still have to compare, sort, and hesitate. When they ask the same question to ChatGPT, the model has already done that work. It has read several sources, compared the options, and returns only the two or three names that best answer that person's specific request. The visitor who clicks the final link is no longer searching, they're verifying a decision that's already been nearly made.

It's the difference between a customer who walks into a store to browse, and a customer who walks in already knowing what they want to buy, on the recommendation of a friend they trust. The chatbot plays that role of trusted intermediary. It has filtered, explained, compared. The person landing on your site no longer needs to be convinced of the product's general value, only reassured on the details.

A report published in March 2026 confirms the scale of the phenomenon: visitors coming from AI assistants like ChatGPT or Perplexity convert 42% better than non-AI traffic, across all channels combined. The quality of the intent that precedes the visit changes the nature of the visitor itself.

Take a concrete example. Someone looking for payroll software for ten employees types their question into Perplexity. The engine answers by citing three tools, each with an indicative price, a strength, and a limitation. They click on the one that best fits their situation, land on a pricing page, and request a demo that same day. The same buyer, starting from a classic Google search, would probably have opened five or six tabs, compared reviews for an hour, then put off the decision until the next day. The chatbot didn't just shorten the path. It did the filtering on their behalf, and that filtering translates directly into conversion rate.

## Advertising is still useful, but for a different goal

This isn't about pitting advertising against visibility in AI chatbots as two competing strategies. They serve different needs. Advertising pays for immediate, controllable volume: you know how much you're spending, how many visitors you're getting, and you can adjust the budget from one day to the next. Getting cited by an AI chatbot doesn't work the same way. It's built over time, through clear content, visible customer reviews, technical documentation that the models reading it can access.

The conversion gap between the two channels doesn't mean you should cut your advertising budget. It means that a euro invested in the quality of what AI chatbots find about a brand produces a return that shows up in no advertising campaign dashboard, and that return will grow mechanically as buyers increasingly use these tools.

## What this changes in how you measure performance

Most marketing dashboards are built around volume: number of visits, number of ad clicks, number of impressions. That logic makes sense for paid advertising, where each visit has a direct, predictable cost. It becomes misleading for traffic generated by AI chatbots, where volume will stay modest for a long time but where each visit carries more weight.

The table below summarizes the gap observed across several recent studies, with different methodologies, which explains the variation in the numbers:

| Traffic source | Observed conversion rate | Relative volume |
|---|---|---|
| Classic Google search | Baseline (1x) | High |
| Visitors from an AI chatbot | 4x to 11x depending on the study | Low, often under 1% of the total |
| Paid advertising | Variable, heavily dependent on cost per click | High but costly to sustain |

Comparing a channel that accounts for 1% of traffic to a channel that accounts for 60% in raw volume doesn't make sense. The right question isn't "how many visitors does the chatbot send us," it's "what is the value of a visitor sent by a chatbot, compared to that of a visitor from an ad." On that basis, a budget devoted to being well cited in AI chatbot answers can pay off more, visitor for visitor, than an equivalent ad budget, even if the total number of visitors remains incomparable for now.

## A signal that will carry increasing weight

Traffic from AI chatbots remains marginal today, but its trajectory is clear: it grows every quarter, while classic organic traffic flattens out. A team that invests now in its visibility within the answers given by ChatGPT, Perplexity, or Gemini isn't just building one more channel. It's capturing a stream of pre-qualified visitors, fewer in number but far more likely to become customers, before this channel gets big enough for everyone to chase it at once.

The logic of classic advertising is still useful for generating volume quickly. But a budget aimed entirely at volume, with nothing invested in the quality of what arrives via generative engines, ignores the underlying trend in buying behavior. The next quarter won't be won just by attracting more people, but by attracting the right people, at the right moment in their decision.

*Vurto measures the share of your traffic that genuinely comes from citations earned in ChatGPT, Perplexity, Gemini, and Claude, to tell apart what's small in volume but big in value.*

### [The GEO Metrics That Actually Matter](https://vurto.ai/en/resources/the-geo-metrics-that-actually-matter) (2026-06-22)

SEO has its metrics. Average position, click-through rate, impressions, organic traffic. Years of practice have built a shared reference frame. Everyone knows what a good DA means. Everyone reads the same dashboards.

GEO doesn't have that yet. The market is young. Tools are being built. And in that fog, many GEO dashboards measure what's easy to measure — not what counts.

Here are the signals worth tracking.

## Citation Rate, Not Raw Presence

The first obvious metric is presence: is my brand mentioned in LLM responses? That's level zero. Useful for detecting total absence. Not enough to steer by.

What counts is the weighted citation rate: across prompts related to my category, in what proportion of responses do I appear? And above all, is that proportion moving over time?

A citation rate that grows is a sign that content, activation, and monitoring work is paying off. A stable rate despite content investment signals either more active competition or a technical readability problem neutralizing the effort.

## Sentiment by LLM

Being cited is good. Being cited positively is the goal.

Brand sentiment in LLMs is a metric that didn't exist two years ago. The most advanced GEO tools now measure it: for each LLM, what's the overall tone of brand mentions? Positive, neutral, negative? And which specific attributes drive that sentiment?

The multi-LLM dimension matters here. Positive sentiment on Perplexity and negative on ChatGPT is a common situation. It usually signals different sources being mobilized by each model — an old comparison visible to ChatGPT, recent positive press accessible to Perplexity. Identifying that is knowing where to act.

## Share of Voice by Prompt Category

A competitor can have a lower overall citation rate than you and still beat you on the prompts that convert the most.

Segmentation by prompt category is essential: comparison prompts ("X vs Y"), purchase intent prompts ("best X for Y"), discovery prompts ("what is X"), validation prompts ("is X reliable"). Across each of these angles, share of voice can vary dramatically.

A B2B brand might be well-positioned on discovery prompts and absent from comparison prompts — where purchase decisions happen. That's a precise diagnostic that guides content priorities.

## Quality of Sources Mobilized

LLMs don't cite all the content that shaped their knowledge. They cite a subset of sources — typically the most recent, best-structured, or highest-trust-signal ones.

Tracking which sources LLMs mobilize when they talk about your brand is a powerful diagnostic metric. If it's consistently third-party comparisons you don't control, dating back two years, containing outdated information — you know where to work.

If your own pages start appearing as sources, that's a signal of technical progress (AI readability, llms.txt) and editorial progress.

## Trend Over Time, Not Snapshots

All these metrics are nearly useless as point-in-time readings. GEO is a medium-term practice. A citation rate moving from 12% to 18% over six months on purchase prompts in your category is a strong signal. A snapshot at 18% without a baseline is noise.

Trend measurement requires consistency. Weekly for sentiment and citation rate on key prompts. Monthly for competitive share of voice. Quarterly for the audit of mobilized sources.

Vurto structures this monitoring continuously, with granularity by LLM, by prompt category, and by competitor. The dashboard gives access to the trend, not just the current number — which is rare among market tools, most of which offer static views.

## What Shouldn't Be in a GEO Dashboard

A few metrics that show up frequently and shouldn't be steering a GEO strategy.

Raw total mention count. Without weighting for context quality or prompt relevance, this number means little. A brand can be mentioned 500 times a week in negative or off-target contexts.

Rankings in "top GEO tools" lists. Some platforms build their own prompt bases biased toward benchmarks that favor them. Watch out for self-referential rankings.

Presence on obscure LLMs with near-zero market share. Monitoring ten LLMs of which eight have 0.1% usage share is noise. Focusing on ChatGPT, Gemini, Perplexity, and Claude covers the vast majority of real generative traffic.

## The Question to Ask Every Week

Is my LLM visibility growing on the prompts that match my actual buyers?

Not "am I mentioned," but "am I mentioned in the right conversations, with the right framing, from the right sources?"

That's the question that turns GEO monitoring into strategic management.

---

*Vurto monitors citation rate, brand sentiment, competitive share of voice, and mobilized sources by LLM — continuously.*

## Academy (learning guides — full content)

Full Vurto Academy curriculum with the full content of each lesson, module by module (74 lessons).

### Technical setup

### [Why everything starts with the technical layer](https://vurto.ai/en/academy/setup-technique/pourquoi-tout-commence-par-la-technique) (2026-05-17)

Without a solid technical foundation, content cannot be crawled, indexed, or cited by a generative AI — the technical layer is the entry condition for any visibility strategy, not a step you handle "later." It governs a three-link chain — crawl, indexing or ingestion, then citation — and if the first link breaks, the other two never get a chance to happen.

## The chain: crawl → indexing/ingestion → citation

Before a page can show up in Google results or in a ChatGPT answer, it first has to be found. That's the job of **crawling**: a bot (Googlebot, Bingbot, or an AI crawler like GPTBot) follows links across the web, reads a page's HTML, and decides whether to keep it for later use.

Next comes **indexing** on the classic search engine side: the page gets added to a database that lets it surface for relevant queries. On the generative AI side, the equivalent is usually called **ingestion**: content gets folded into document stores or indexes the model can consult when generating an answer — through real-time web search for Perplexity or browsing-enabled ChatGPT, or through datasets used at different stages depending on the provider.

The last link, **citation**, depends entirely on the first two. A generative AI cannot cite a page it was never able to read, and it won't read it if a bot could never reach it in the first place. It's a simple point that gets overlooked constantly: teams optimize content with citation in mind while the actual blocker sits one or two steps earlier, at the level of technical access.

## What breaks this chain in practice

In most technical audits, the same blockers show up again and again:

- An overly restrictive **robots.txt** that blocks entire sections of a site, sometimes by mistake or out of excessive caution.
- Client-side JavaScript rendering with no server-side rendering or pre-rendering, which some crawlers fail to execute properly.
- **Orphan pages** with no internal links pointing to them, invisible to a bot that navigates link by link.
- Server errors (5xx), redirect loops, or slow response times that push bots to reduce how often they come back.
- No up-to-date **XML sitemap**, leaving search engines without a clear list of pages to explore.
- Duplicate content or a confusing URL structure, which spreads crawl budget thin across pages with no real value.

Each of these looks minor in isolation. Combined, they explain why some sites with genuinely solid content stay invisible.

## SEO and GEO share the same foundation

GEO doesn't replace technical SEO — it builds on top of it. AI crawlers operate through mechanisms fairly close to those of classic search engines: they read HTML, generally respect the robots.txt file, and are sensitive to whether a page is actually reachable. It's generally observed that a page invisible to Googlebot also has little chance of being properly read by an AI crawler, since the same technical obstacles — unrendered JavaScript, server errors, robots.txt blocks — tend to affect both kinds of bots in similar ways.

That's why a GEO strategy that ignores the technical layer starts at a structural disadvantage. You can produce the best-structured content in the world — if it isn't reachable, it stays invisible both to someone typing a Google query and to someone asking Claude or Perplexity a question.

## A quick test to see where you stand

Before diving into tools, a few immediate checks give you a first diagnosis without installing anything:

- Type `site:yourdomain.com` into Google: if the number of results is far lower than your site's actual page count, a chunk of your content is likely not indexed.
- Open `yourdomain.com/robots.txt` in a browser: if a `Disallow: /` line appears under `User-agent: *`, your entire site is closed off to bots, including AI crawlers.
- Turn off JavaScript in your browser (or use Search Console's "Inspect URL" tool once it's set up) and reload a key page: if the main content disappears, a bot that doesn't execute JavaScript properly will see the same thing.
- Check whether a sitemap exists at `yourdomain.com/sitemap.xml`: its absence isn't a dealbreaker on its own, but it deprives search engines of a useful reference point.

These four checks take less than five minutes and are often enough to tell whether a visibility problem comes from an obvious technical block before digging any further.

## Where to actually start

Before working on substance — content structure, direct answers, structured data — you need to secure the basics. In order:

1. Check what's already indexed today (a `site:yourdomain.com` search on Google, or a direct look at Search Console if it's already set up).
2. Install and configure Google Search Console for a reliable view of indexing status.
3. Audit the robots.txt file and XML sitemap to make sure nothing important is being excluded.
4. Check loading speed and Core Web Vitals, which influence crawl budget.
5. Only after that, work on structuring content to make it "GEO-ready."

This lesson sets the general frame. The next ones get concrete: installing [Google Search Console](/en/academy/setup-technique/installer-google-search-console), understanding the [robots.txt file and controlling AI crawlers](/en/academy/setup-technique/robots-txt-et-crawlers-ia), and digging into [why AI sometimes ignores your site](/en/academy/comprendre-le-geo/pourquoi-les-ia-ignorent-votre-site) once the technical basics are in place.

### [Installing and configuring Google Search Console](https://vurto.ai/en/academy/setup-technique/installer-google-search-console) (2026-05-18)

Setting up Google Search Console takes three steps: create a property (domain or URL-prefix), prove you own it, then submit your XML sitemap. The whole process takes about fifteen minutes and is the first building block of any technical tracking, for SEO as much as for GEO.

## Create an account and choose the right property type

Go to `search.google.com/search-console` with a Google account and click "Add property." You'll be offered two options:

- **Domain property**: covers the entire domain, across all subdomains and protocols (http, https, with or without www). This is the recommended option in most cases, since it gives you a complete view and avoids duplicating properties.
- **URL-prefix property**: covers only one exact URL (for example, only `https://www.yourdomain.com`). Useful if you can't edit your domain's DNS records, or if you want to track a specific subfolder in isolation.

If you have control over your domain's DNS configuration, go with the domain property — it simplifies everything else downstream.

## Verifying the property: the available methods

Google needs to confirm you actually control the site before granting access to its data. Several methods exist depending on the property type you chose:

1. **DNS record (TXT)** — the only method available for a domain property. Google gives you a value to add as a TXT record with your host or DNS registrar. Propagation can take anywhere from a few minutes to 48 hours.
2. **HTML meta tag** — for a URL-prefix property, Google provides a `<meta name="google-site-verification" content="...">` tag to insert in the `<head>` of the homepage.
3. **HTML file upload** — a uniquely named file (e.g. `google1234567890.html`) provided by Google, placed at the site root and reachable directly at `yourdomain.com/google1234567890.html`.
4. **Google Analytics or Google Tag Manager** — if either is already installed and you have admin rights on it, verification happens automatically with no extra step.

Pick whichever method matches your technical access: DNS if you manage the domain name, meta tag or HTML file if you only have access to the site's code.

## Submitting the sitemap

Once the property is verified, go to the "Sitemaps" entry in the side menu. Enter your XML sitemap URL (usually `sitemap.xml` or `sitemap_index.xml`, appended after your domain) and click "Submit." This step isn't required for Google to crawl your site, but it speeds up discovery of new pages and gives Google an explicit list of the URLs you consider important.

If you're not sure whether a sitemap already exists, just test `yourdomain.com/sitemap.xml` directly in your browser. Most CMSs (WordPress with Yoast or Rank Math, Shopify, Webflow) generate one automatically.

## Extra settings worth doing right away

A handful of settings, done once the property is verified, save time down the line:

- **Link Google Analytics 4** if you already have it installed, so you can later cross-reference crawl and traffic data.
- **Add other team members as users** (Settings > Users and permissions menu), with an access level that matches their role (read-only or full owner).
- **Check the "Configuration" > "Changes"** section, which lists structural changes Google has detected (title tag changes, canonical changes, etc.) — useful for spotting unintended modifications.
- If the site is multilingual, don't expect Search Console to handle international targeting automatically: that depends on your hreflang implementation, which needs to be checked separately.

Once this setup is done, the next step is learning to [read your data in Google Search Console](/en/academy/setup-technique/lire-donnees-google-search-console) to turn this access into an actionable diagnosis. It's also worth [installing Bing Webmaster Tools](/en/academy/setup-technique/installer-bing-webmaster-tools), which follows the same logic and matters for visibility in Copilot's answers.

### [Reading your Google Search Console data](https://vurto.ai/en/academy/setup-technique/lire-donnees-google-search-console) (2026-05-19)

Four Google Search Console reports deserve regular attention: Performance (what actually drives clicks), Indexing (what's really indexed), Sitemaps (what Google managed to crawl), and Core Web Vitals (loading experience). Together, they give you a complete diagnosis without needing any third-party tools.

## The Performance report: queries, pages, clicks, impressions

This is the most-viewed report, and often the most poorly read one. It crosses four metrics: **impressions** (how many times a page showed up in results), **clicks**, **CTR** (click-through rate, clicks divided by impressions), and **average position**.

A few useful readings:

- A page with a lot of impressions but a low CTR usually points to a weak title tag or meta description, not a content problem.
- An average position stuck around page 2 (positions 11-20) signals a page Google considers relevant enough to exist, but that's missing the signal needed to break into page 1 — often internal linking or content depth.
- The query filter lets you spot keywords you're already ranking for without realizing it, sometimes unexpectedly — a solid starting point for complementary content.

Always segment by device type (mobile vs. desktop) and by country if your audience is international: global averages often hide significant gaps.

## Coverage / Indexing: what's indexed and what isn't

The "Pages" report (formerly "Coverage") lists your URLs by status: indexed, or excluded with a specific reason. The most common reasons worth watching:

- **"Crawled – currently not indexed"**: Google visited the page but chose not to index it, usually due to low perceived value or duplicate content.
- **"Discovered – currently not indexed"**: Google knows the URL exists but hasn't crawled it yet, generally a sign of limited crawl budget or low perceived priority.
- **"Page with redirect"** or **"404 error"**: fix it or leave it as-is, depending on whether it's intentional or an oversight.
- **"Duplicate, Google chose different canonical than user"**: Google picked a different canonical URL than the one you declared, often a sign of poorly managed duplicate content.

The goal isn't 100% of pages indexed at any cost: some pages (filters, technical pagination, internal search results) have no reason to be. The goal is for the pages that matter to be indexed, and to understand why the ones that aren't, aren't.

## Sitemaps: checking everything is accounted for

In the Sitemaps report, Google shows the number of URLs discovered in the submitted sitemap. A large gap between the number of URLs in the sitemap and the number of pages actually indexed (visible in the Pages report) is a warning sign — it can point to a poorly generated sitemap, outdated URLs, or low-quality pages Google refuses to index despite their presence in the sitemap.

## Core Web Vitals in Search Console

The Core Web Vitals report aggregates real user data (from the Chrome User Experience Report) and buckets your pages into three categories — good, needs improvement, poor — across three metrics: LCP (loading time of the largest visible element), INP (responsiveness to interactions), and CLS (visual stability). Unlike a one-off testing tool, this report reflects actual visitor experience, grouped by URL pattern, which helps with prioritization: it's more valuable to fix a page template used by 500 URLs than to fix one isolated page.

## Secondary reports worth not ignoring

Beyond the four main reports, two others deserve a regular look even though they get checked less often:

- **The Links report** lists your most frequent external backlinks, your most internally linked pages, and the anchor text used. A central page on your site that gets very few internal links is a sign it risks being under-prioritized by bots, regardless of how good its content is.
- **The "Security Issues" report** flags any detected hacking or misleading content identified by Google. A compromised site can see its crawl frequency drop sharply, an easy thing to miss if this report never gets checked for lack of a visible alert elsewhere.

These two reports are rarely the priority on a first pass, but ignoring a security issue for several months can undo all the effort put into every other lever.

## Cross-referencing the data to prioritize

The most effective method is cross-referencing reports rather than reading them in isolation:

1. Spot high-potential pages in Performance (lots of impressions, average position 8-20).
2. Check in Pages whether they're fully indexed without restriction.
3. Check in Core Web Vitals whether their template is being held back by a speed issue.
4. Prioritize fixes on the templates that affect the most high-potential pages at once.

This cross-reading is exactly the foundation of a full technical audit: the next lesson details how to [audit your site with generative AI and GSC data](/en/academy/setup-technique/auditer-son-site-avec-ia-et-gsc) to speed up this prioritization, before moving on to the [complete technical checklist](/en/academy/bases-seo-techniques/checklist-technique-complete).

### [Auditing your site with generative AI + GSC data](https://vurto.ai/en/academy/setup-technique/auditer-son-site-avec-ia-et-gsc) (2026-05-20)

Exporting your Google Search Console data as CSV and feeding it to ChatGPT or Claude with a structured prompt gets you a prioritized list of technical fixes within minutes — as long as you cross-reference several exports and verify every recommendation before applying it.

## Why run GSC data through a generative AI

Google Search Console shows raw data: tables of queries, pages, indexing statuses. Nothing in it automatically tells you what to fix first. A generative AI is useful here not because it "knows" your site, but because it can quickly cross-reference several columns of data, spot recurring patterns (say, a page template that keeps showing up among non-indexed URLs), and turn a technical diagnosis into plain language for a team that isn't made up of SEO specialists.

This isn't a replacement for human analysis, it's an accelerator: what would take an hour of manual sorting in a spreadsheet can be done in a few minutes, provided you feed it clean data and a precise prompt.

## Exporting the right data from Search Console

Before involving an AI, prepare your exports. Three reports are enough for a first-pass audit:

1. **Pages report** (Indexing): export the full list of URLs with their status and exclusion reason. This is where most technical blocking signals live.
2. **Performance report**: export data by page (not just by query), over a period of at least 3 months, with impressions, clicks, CTR, and average position.
3. **Core Web Vitals report**: export the breakdown by URL group if volume allows.

Each export is done via the "Export" button in the top-right corner of each report, as CSV or Google Sheets. Clean up the files if needed (empty columns, duplicates) before handing them over.

## Building an effective audit prompt

A vague prompt gets a vague result. Structure your request with context, objective, and expected format. For example:

> "Here's an export of my Pages report from Google Search Console for [domain name], a [type of business] site. Analyze the exclusion statuses and identify which reasons come up most often. For each recurring reason, explain the likely cause and propose a concrete fix. Rank the results by number of affected URLs, from most to least impactful."

For the Performance report, a useful prompt is to ask it to spot high-potential pages: "Identify pages with more than 500 impressions and an average position between 8 and 20 — these are candidates for targeted optimization rather than a full rewrite."

Avoid generic prompts like "how do I improve my SEO": without attached data and format constraints, the answer stays theoretical and hard to act on.

## What an AI is good at spotting — and what it isn't

A generative AI excels at: spotting statistical patterns across a large number of rows, grouping URLs by template or site section, and translating a technical report into plain language for a marketing team.

It's more limited when it comes to: verifying in real time whether a fix has actually been applied (it has no access to your live site unless you explicitly provide it), judging the real severity of an issue without business context (a poorly indexed page can be entirely normal if it holds little value), and replacing an actual technical test (it can't run PageSpeed Insights on your behalf, for instance). Every recommendation it generates should therefore be verified before implementation, especially anything touching robots.txt or redirects, where a mistake can deindex important pages.

## Turning the analysis into a prioritized action plan

Once you have the analysis, always ask for a usable format: a table with columns "issue," "number of affected URLs," "proposed fix," "estimated effort." That table, not the raw text of the answer, is what becomes your technical roadmap.

This method builds directly on the report-reading approach from the previous lesson on [reading your Google Search Console data](/en/academy/setup-technique/lire-donnees-google-search-console). Once technical fixes are underway, it's worth checking that your [site speed and Core Web Vitals](/en/academy/setup-technique/vitesse-du-site-core-web-vitals) are keeping up, then digging into [how a search engine reads a page](/en/academy/bases-seo-techniques/comment-un-moteur-lit-une-page) to understand the fundamentals behind these diagnostics.

### [Installing Bing Webmaster Tools](https://vurto.ai/en/academy/setup-technique/installer-bing-webmaster-tools) (2026-05-21)

Bing Webmaster Tools takes a few minutes to set up, with a direct import available straight from Google Search Console — it's the quickest step in this module, and one of the most overlooked, even though it directly affects your visibility inside Copilot.

## Why Bing matters for GEO

It's natural to associate Bing with a modest market share compared to Google in classic search. But Bing plays a different role in the generative AI ecosystem: Copilot, Microsoft's assistant, relies in part on Bing's index for its web-search-backed answers. A site that's absent or poorly indexed on Bing has fewer chances of being picked up by that assistant, regardless of its standing on Google. That alone is reason enough not to skip this tool, even if your current Bing traffic looks marginal.

## Create an account and add your site

Go to `bing.com/webmasters` and sign in with a Microsoft account (or create one). Click "Add a site" and enter your domain's URL.

## Import directly from Google Search Console

If your site is already verified in Google Search Console, Bing Webmaster Tools offers an import option: you authorize access to your Google account, and Bing automatically pulls in the list of your verified properties, your already-submitted sitemap, and a sample of your performance data. This is by far the fastest method — it skips a new property verification and a new sitemap export.

## Verifying ownership if you don't import

If you'd rather verify manually, or your site isn't on Google Search Console yet, three methods are offered:

1. **HTML meta tag** to insert in the `<head>` of the homepage.
2. **XML file upload** at the site root (similar to Google's HTML file, but in XML format).
3. **DNS record (CNAME)** to add with your registrar, for domain-level verification.

Pick whichever matches your level of access, exactly as with Google Search Console.

## Exploring Bing-specific reports

Once the property is active, Bing Webmaster Tools offers reports similar to Google Search Console's (Performance, Crawl, Site Index) but also a few tools of its own:

- **SEO Reports**, which automatically lists technical issues detected on your pages (missing titles, duplicate meta tags, images without an alt attribute), with no extra setup required.
- **URL Inspection**, the equivalent of Google's URL inspection, which lets you check how Bingbot sees a specific page and request an on-demand crawl.
- **Backlinks**, a report listing external links pointing to your site, useful alongside Google Search Console's Links report for cross-referencing both sources.

These reports need less upkeep than Google's, but are worth a monthly check once the tool is set up, particularly the SEO Reports section, which often surfaces simple technical fixes.

## Submitting the sitemap and setting the basics

Once the property is verified, go to the "Sitemaps" menu and submit your XML sitemap URL, just like with Google. Then check two useful settings:

- **Targeted country/region**, in the site settings, if your business is local or national.
- **Crawl control**, a Bing-specific feature that lets you throttle how often the bot visits if your server has limited resources — only touch this if you're seeing a real impact on server performance.

Setting up Bing Webmaster Tools rounds out the technical tracking foundation started with [Google Search Console](/en/academy/setup-technique/installer-google-search-console). With both tools in place, the next step is adding [Google Analytics 4](/en/academy/setup-technique/installer-google-analytics-4) to measure, this time, the traffic actually generated by visitors coming from generative AI.

### [Installing Google Analytics 4](https://vurto.ai/en/academy/setup-technique/installer-google-analytics-4) (2026-05-22)

Installing Google Analytics 4 properly takes four steps: create the property and data stream, add the tracking code, wire up consent management, then verify data is actually coming in before you rely on it to analyze traffic from generative AI. This installation builds on the foundation already laid with [Google Search Console](/en/academy/setup-technique/installer-google-search-console) and [Bing Webmaster Tools](/en/academy/setup-technique/installer-bing-webmaster-tools): it adds traffic measurement on top of crawl measurement.

## Create a GA4 property and data stream

On `analytics.google.com`, create a property (if you're starting from scratch, GA4 is now the only version Google offers). Fill in the timezone and currency, which can't easily be changed later. Then add a **data stream** of type "Web" with your site's URL: this stream generates a measurement ID (formatted `G-XXXXXXX`) used to connect your site to this property.

## Install the tracking code

Two main methods:

1. **Direct gtag.js tag**: paste the code snippet Google provides into the `<head>` of every page on the site. Simple for a small static site, but rigid if you plan to add other tracking tools later.
2. **Google Tag Manager (GTM)**: recommended as soon as the site has more than a handful of pages, or if you expect to add other tools (ad pixels, other analytics). GTM installs once via two code snippets (one in the `<head>`, one right after the `<body>` opens), and GA4 is then configured as a tag inside GTM — no need to touch the site's code again for every change.

Most CMSs (WordPress, Shopify, Webflow) also offer native GA4 or GTM integration through a dedicated settings field, which avoids editing source code directly.

## Handle consent before any data collection

If your audience includes European users, data collection through GA4 falls under GDPR and generally requires prior consent for non-essential cookies. In practice, that means connecting GA4 to a consent banner (CMP) that blocks analytics tags from firing until the user has given their approval. Google offers a native option, **Consent Mode**, which lets GA4 keep modeling some data even without consent, without storing identifying cookies. This technical point goes beyond the scope of this lesson, but it directly affects how reliable your data is: an installation that ignores consent risks skewed data or a compliance problem.

## Verify the data is actually coming in

Once installed, don't assume "it works" without checking. Two tools let you confirm this immediately:

- The **"Realtime"** report in GA4: open your site in a tab, browse a few pages, and check that your visit shows up within a few minutes.
- The **Google Tag Assistant** browser extension or GTM's debug mode, which show exactly which tags fire, in what order, and with what data.

If nothing shows up after a few minutes, the most common causes are a mistyped measurement ID, an ad blocker preventing the script from loading, or a misconfigured consent banner blocking everything by default.

## The key events to enable from the start

GA4 automatically collects several events (page views, scroll, outbound clicks, file downloads) with no extra setup, as long as "Enhanced measurement" is enabled in the data stream settings — on by default, but worth double-checking. Beyond these automatic events, define at least one conversion event specific to your business (form submission, sign-up, purchase) so you can later measure whether traffic from generative AI converts differently than classic organic traffic.

Once GA4 is installed and reliable, the next step is learning to [configure the GA4 reports that actually matter](/en/academy/setup-technique/configurer-ga4-rapports-utiles), particularly to isolate referral traffic from generative AI and track how it evolves over time.

### [Configuring GA4: the reports that actually matter](https://vurto.ai/en/academy/setup-technique/configurer-ga4-rapports-utiles) (2026-05-23)

To track traffic from generative AI in GA4, isolate the referring domains (chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com...) in the acquisition report, ideally through a custom channel, then follow its trend over time rather than looking at a single snapshot.

## Isolating referral traffic from generative AI

By default, GA4 classifies incoming traffic using predefined channel rules (Organic Search, Direct, Referral, Paid Search...). Traffic from a click on a link cited by ChatGPT or Perplexity usually lands in the **"Referral"** channel, lumped together with any other site linking to you — which makes it invisible unless you dig further.

To isolate it, open the **Acquisition > Traffic acquisition** report, then add a secondary dimension for **"Session source"**. Filter on the relevant domains: `chatgpt.com`, `chat.openai.com`, `perplexity.ai`, `gemini.google.com`, `copilot.microsoft.com`, `claude.ai`. This list evolves with usage and should be reviewed regularly.

## Creating a custom channel or a dedicated segment

Manually filtering every time you check gets tedious fast. Two more durable options:

- **A custom channel group** (Admin > Data channel groups), where you create a rule like "if session source contains chatgpt.com OR perplexity.ai OR gemini.google.com OR copilot.microsoft.com OR claude.ai, classify as 'Generative AI' channel." This channel then shows up like any other in standard reports, allowing ongoing tracking without redoing the filter every time.
- **A custom exploration** (Explore menu), letting you build a table crossing session source, number of sessions, conversion rate, and landing pages, with export available for external tracking.

The custom channel is the more durable solution over time: once created, it applies retroactively to some extent and stays available across all standard reports without repeated manual work.

## Reports worth checking beyond acquisition

Once the channel is isolated, three angles add more value than a simple session count:

- **Landing pages**: which pages on the site actually receive this traffic? This reveals, indirectly, which content is being picked up by AI.
- **Engagement rate and session duration**: does AI-driven traffic behave differently from classic organic traffic (shorter sessions, higher bounce rate, or on the contrary stronger engagement)? There's no universal rule here — it depends heavily on industry and content type, and is worth observing in your own data rather than assuming.
- **Conversions by channel**: if a conversion event is set up, compare its rate on the generative AI channel against other channels, to assess the real value of this traffic rather than just its volume.

## Tracking the trend over time

A one-off reading of AI traffic has little value on its own — the trend is what matters. Set up simple monthly tracking (number of sessions on the generative AI channel, compared to the previous month and to the previous year if volume allows). A gradual rise in this channel, even from a small base, is a signal that your content is starting to get picked up by answer engines — a signal that complements AI visibility scores measured elsewhere, but this time measured on real traffic rather than prompt tests.

## Setting up an automatic alert

Rather than relying on remembering to check manually, GA4 lets you create **custom insights** (Admin > Custom insights) that trigger an email notification when a defined condition is met — for example, a change in session count on the "Generative AI" channel beyond a certain threshold over a week. This kind of alert removes the need to remember to check the report every month, and lets you react faster if an unusual spike or drop shows up, particularly after a robots.txt change or a content overhaul.

## Limits worth knowing

This tracking has blind spots worth keeping in mind:

- Some traffic from AI answers isn't necessarily attributed correctly: links shared from a mobile app or copy-pasted into a new tab can land as "Direct" traffic, invisible in the dedicated channel.
- Consent banners (see the lesson on installing GA4) can reduce collected data volume if some visitors decline tracking.
- Referring domain naming changes over time: a new AI engine or a domain change on the provider's side can make traffic disappear from your custom channel without warning, which is why the channel rule is worth revisiting regularly.

With this tracking in place, you have a direct measure of AI traffic to weigh against the technical fixes already made through [Google Search Console](/en/academy/setup-technique/lire-donnees-google-search-console) and your [site speed](/en/academy/setup-technique/vitesse-du-site-core-web-vitals) checks. It's also a solid foundation before tackling the next module's core question: [why AI sometimes ignores your site](/en/academy/comprendre-le-geo/pourquoi-les-ia-ignorent-votre-site).

### [Site speed and Core Web Vitals](https://vurto.ai/en/academy/setup-technique/vitesse-du-site-core-web-vitals) (2026-05-24)

Site speed influences both the crawl budget bots devote to your site and the actual experience of your visitors: a slow site gets crawled less often and less deeply, and loses visitors before they even reach the content. Core Web Vitals are the standard way to measure that speed in a way that's comparable across sites.

## Why speed affects crawl budget and experience

**Crawl budget** refers to the number of pages a bot is willing to explore on your site within a given time. That budget isn't unlimited: it depends partly on how quickly your server responds. It's generally observed that a slow or unstable site pushes bots to reduce how often they visit and how many pages they crawl per visit, which mechanically slows down the discovery of new content and its update in the index.

On the user side, the effect is more direct: a long loading time increases the abandonment rate before content even renders. For a GEO objective, this matters twice over: a slow page has less chance of being crawled by an AI crawler, and less chance of being read all the way through by a visitor who clicked through from a generated answer.

## The three Core Web Vitals metrics explained

Google defined three metrics that together make up Core Web Vitals:

- **LCP (Largest Contentful Paint)**: the time it takes to render the largest visible element on screen (often a hero image or a headline). Generally recommended target: under 2.5 seconds.
- **INP (Interaction to Next Paint)**: the delay between a user interaction (click, keypress) and the corresponding visual update. This metric replaced FID (First Input Delay) in 2024. Generally recommended target: under 200 milliseconds.
- **CLS (Cumulative Layout Shift)**: a measure of unexpected visual shifts during loading (a button that moves right before you click it, for instance). Generally recommended target: under 0.1.

These thresholds are indicative benchmarks published by Google, not absolute rules that apply everywhere: a site with a lot of complex interactive content may tolerate slightly different margins depending on its use case.

## Measuring with PageSpeed Insights (and why not only that)

`pagespeed.web.dev` is the reference tool for testing a specific page. It provides two types of data, worth telling apart:

- **Field Data**, sourced from the Chrome User Experience Report, reflecting the real experience of your visitors over the last 28 days — only available if your site gets enough Chrome traffic.
- **Lab Data**, generated from a simulated test at the moment you run the analysis — useful for diagnosis, but can vary from one run to the next depending on the simulated network conditions.

The Core Web Vitals report in Google Search Console, covered in an earlier lesson, complements this tool by giving an aggregated site-wide view rather than a page-by-page one, which helps prioritize by template rather than by isolated URL.

## Mobile and desktop: two separate realities

Core Web Vitals are evaluated separately for mobile and desktop, and the results often diverge quite a bit between the two. Since most web traffic today happens on mobile, and Google uses "mobile-first" indexing (your page's mobile version is the reference used for crawling and evaluation), it's advisable to treat mobile results as the priority whenever there's a trade-off to make. A site that scores well on desktop but poorly on mobile is still, from Google's standpoint, a site with a speed problem to fix.

## The most common fixes, and their order of priority

In most audits, the fastest wins come from:

1. **Image compression and resizing**: serve images at the size they're actually displayed, in a modern format (WebP or AVIF), with lazy loading for off-screen images.
2. **Reducing blocking JavaScript**: defer loading of scripts that aren't essential to the initial render (live chat, third-party widgets, secondary ad pixels).
3. **Reserving space for dynamic elements**: define explicit dimensions for images, videos, and ads before they load, to avoid layout shifts (CLS).
4. **Caching and CDN**: use a content delivery network to bring static files closer to visitors, and configure appropriate browser caching.
5. **Hosting choice**: cheap shared hosting is often the hardest limiting factor to fix without switching technical solutions.

## Tracking the impact over time

After each significant fix, re-test with PageSpeed Insights and watch Search Console's Core Web Vitals report over the following weeks rather than the following hour: field data is based on a rolling 28-day window, so it doesn't react instantly to a change.

Once speed is under control, the site's technical foundation is largely in place. It's time to move on to controlling bot access with the lesson on [the robots.txt file and controlling AI crawlers](/en/academy/setup-technique/robots-txt-et-crawlers-ia), then to understand in detail [how a search engine reads a page](/en/academy/bases-seo-techniques/comment-un-moteur-lit-une-page).

### [The robots.txt file and controlling AI crawlers](https://vurto.ai/en/academy/setup-technique/robots-txt-et-crawlers-ia) (2026-05-25)

The robots.txt file, placed at the root of your site, tells each bot which sections it can or can't crawl — including generative AI crawlers, each identified by its own user-agent (GPTBot, ClaudeBot, PerplexityBot, Google-Extended...) that you can allow or block independently of one another.

## What robots.txt does (and doesn't do)

Robots.txt gives **crawling** instructions, not privacy guarantees or guaranteed indexing control. Three important clarifications:

- Blocking a URL in robots.txt stops a compliant bot from crawling it, but doesn't necessarily stop it from appearing in an index if it's already known through other means (external links, for instance). To actually remove a page from an index, the `noindex` tag or a removal request through Search Console are the right tools, not robots.txt.
- The file works on good faith: a legitimate, declared bot (Googlebot, GPTBot, ClaudeBot) generally respects the rules it finds there, but nothing technically stops a malicious or poorly coded bot from ignoring them.
- Rules apply per **user-agent**: you can write different rules for each bot, which is exactly what lets you arbitrate between classic SEO and GEO.

## Basic syntax

A robots.txt file is made of blocks, each targeting one or more user-agents:

```
User-agent: [bot name or * for all]
Disallow: [blocked path]
Allow: [allowed path, takes priority over a broader Disallow]
```

Simple example blocking an admin folder for all bots, but allowing everything else:

```
User-agent: *
Disallow: /admin/
Allow: /

Sitemap: https://yourdomain.com/sitemap.xml
```

The `Sitemap:` line at the end, while optional, is good practice: it tells bots directly where to find the full list of your pages, independent of any submission made through Search Console or Bing Webmaster Tools.

## The main AI user-agents to know

Here are the most common crawlers as of today, keeping in mind this list evolves with the market:

- **GPTBot** (OpenAI): used to collect content for model training.
- **OAI-SearchBot** (OpenAI): dedicated to ChatGPT's web search feature, distinct from GPTBot.
- **ChatGPT-User** (OpenAI): triggers when a user explicitly asks ChatGPT to look at a page in real time.
- **ClaudeBot** and **anthropic-ai** (Anthropic): content collection by Anthropic.
- **PerplexityBot**: crawling for indexing by Perplexity; **Perplexity-User** handles real-time browsing requests initiated by a user.
- **Google-Extended**: specifically controls how your content is used by Google's AI models (Gemini), separate from Googlebot, which handles classic indexing.
- **Bingbot** covers both classic Bing search and, to some extent, the data used by Copilot.

Each of these user-agents can be targeted individually in robots.txt, which lets you, for example, allow Googlebot without allowing Google-Extended, or allow PerplexityBot without allowing GPTBot.

## Deciding: allow, block, or differentiate by crawler

The right call depends on your objective, and there's no universal answer:

- **Blocking a training crawler** (GPTBot, ClaudeBot for collection, Google-Extended) reduces use of your content for training future models, but has no effect on your presence in real-time answers generated by those same assistants, which often go through a separate web search mechanism.
- **Blocking a search/citation crawler** (OAI-SearchBot, PerplexityBot, Perplexity-User) directly reduces your chances of being cited in those tools' answers, since that's the bot that lets the AI discover and read your page at the moment it answers a question.
- Most sites looking to grow their GEO visibility have an interest in **allowing search/citation crawlers**, and deciding case by case on training crawlers based on their own IP and content policy.

Example robots.txt that blocks training but allows citation:

```
User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: ClaudeBot
Allow: /

User-agent: *
Allow: /

Sitemap: https://yourdomain.com/sitemap.xml
```

## The case of firewalls and bot management tools

Beyond the robots.txt file itself, some sites use an application firewall or a service like Cloudflare to block or throttle bots at the network level, independent of what's declared in robots.txt. These settings are sometimes turned on by default with broad blocking rules against "bots" in general, which can block a legitimate AI crawler without any trace of it in your robots.txt. If you notice an AI crawler you've explicitly allowed never shows up in your logs, check these network-level settings in addition to the robots.txt file itself: both layers need to line up for your rules to actually produce the intended effect.

## Checking and testing your robots.txt

After any change, check the file directly at `yourdomain.com/robots.txt` in a browser, and use the robots.txt tester built into Google Search Console ("Settings" > "robots.txt" report) to confirm a specific URL is allowed or blocked as intended. A syntax mistake in this file (a `Disallow: /` left under the wrong user-agent, for instance) can deindex an entire site by accident — treat it with the same care as a production change.

Robots.txt works together with two other files: the [XML sitemap and the llms.txt file](/en/academy/setup-technique/llms-txt-et-sitemap-xml), which round out the setup by explicitly telling AI systems your pages and your preferences. To go further on the bots themselves, the lesson [AI crawlers: who reads you, and why](/en/academy/setup-technique/crawlers-ia-qui-vous-lit) details how to spot them directly in your server logs, beyond what they declare in their user-agent.

### [The llms.txt file and the XML sitemap](https://vurto.ai/en/academy/setup-technique/llms-txt-et-sitemap-xml) (2026-05-26)

The llms.txt file is an emerging, not universally adopted standard that offers a structured Markdown summary of your site for generative AI systems; the XML sitemap, on the other hand, remains the established and widely supported standard for telling search engines the full set of pages to crawl. The two play different roles and complement each other rather than compete.

## What llms.txt is

Proposed in 2024, llms.txt is a Markdown text file placed at the site root (`yourdomain.com/llms.txt`), meant to give a generative AI a condensed, readable summary of the site's content: an overview, main sections, links to the most important pages or documents. The underlying idea is to offer a compact entry point that's easier for a language model to process than a full site crawl.

It isn't an official standard adopted or confirmed by the major AI providers — OpenAI, Anthropic, Google, and Perplexity haven't made any formal commitment to systematically use it. It's a community proposal, championed in particular by parts of the developer ecosystem, gaining traction without being any guarantee of adoption.

## What it delivers today (and its limits)

In practice, publishing an llms.txt doesn't guarantee better citation in AI answers: no major provider has confirmed that its models or crawlers systematically consult this file when responding to a user query. What can reasonably be expected from it today:

- A signal of good practice and clear structuring, useful should tools come to rely on it more broadly in the future.
- A useful resource for cases where a developer or a specifically configured AI agent explicitly looks for this file (for example, certain coding assistants that use it to understand a project's documentation).
- A clarifying exercise in its own right: writing an llms.txt forces you to synthesize what actually matters on the site, which has value even independent of whether an AI actually reads it.

It in no way replaces the XML sitemap, robots.txt, or properly implemented Schema.org markup, which remain the best-established levers today.

## How to generate and structure an llms.txt

The proposed format stays simple, in Markdown:

```
# Site name

> Short one- or two-sentence description of what the site does and its value.

## Documentation
- [Page name](https://yourdomain.com/page): one-line description

## Resources
- [Resource name](https://yourdomain.com/resource): one-line description
```

For a content site (like a blog or a resource center), favor a list of your most foundational guides or articles rather than an exhaustive list of every page — the goal is to synthesize, not duplicate the sitemap. No universal automatic generator exists today for every CMS: writing it manually, or semi-automatically by pulling your main titles and meta descriptions, remains the most reliable method.

## A refresher on the XML sitemap and best practices

The XML sitemap, meanwhile, is a standard that's been supported by every major search engine for a long time. A few best practices worth checking:

- **Cap at 50,000 URLs and 50MB uncompressed per file**: beyond that, use a sitemap index referencing multiple sitemaps.
- **Only include canonical, indexable, up-to-date URLs**: a sitemap containing `noindex` pages, redirects, or 404s sends a negative quality signal.
- **Update it automatically** through the CMS rather than by hand, to avoid drift from what's actually published.
- **Include a reliable last-modified date** (`<lastmod>`), since some engines use it to prioritize re-crawling.

## Where to place these files and how to keep them current

Both files follow the same location logic: at the domain root, directly reachable (`yourdomain.com/llms.txt`, `yourdomain.com/sitemap.xml`), with no authentication or redirect in the way. For the sitemap, automatic updates through the CMS are almost always possible and preferable. For llms.txt, which remains a more editorial document, a quarterly manual review is usually enough to reflect newly published foundational content — no need to regenerate it every time a new article goes live if the site's overall structure hasn't changed.

## Two files, complementary rather than competing

The XML sitemap says "here are all my pages, go crawl them"; llms.txt, in its current intended use, says more "here's the essential to understand if you need to summarize my site quickly." Publishing both costs virtually nothing in implementation time and covers both approaches without having to bet on the future adoption of either.

With robots.txt, the sitemap, and llms.txt in place, the technical layer for controlling crawlers is complete. What's left is understanding more precisely [who reads you and why](/en/academy/setup-technique/crawlers-ia-qui-vous-lit), then moving on to the [technical SEO basics](/en/academy/bases-seo-techniques/comment-un-moteur-lit-une-page) to look at how content itself gets interpreted once it's reachable.

### [AI crawlers: who reads you, and why](https://vurto.ai/en/academy/setup-technique/crawlers-ia-qui-vous-lit) (2026-05-27)

AI crawlers fall into two families with distinct purposes — collecting data to train models, and retrieving content in real time to answer a specific question — and you spot them concretely in your site's server logs, not just by trusting their user-agent declarations.

## Two families of AI crawlers: training vs. real-time retrieval

Not all AI crawlers pursue the same goal, which directly changes how you should treat each one:

- **Training crawlers** scan the web to build or enrich the datasets used when training future models. Their visit has no immediate effect on your visibility in answers generated today; it concerns future versions of the model, on a timeline providers don't make public.
- **Real-time retrieval crawlers** (often called *retrieval* or *browsing*) trigger the moment a user asks a question that requires a current web search. This is the type of crawler that lets a ChatGPT, Perplexity, or Copilot answer cite your page with a clickable link, right now.

Blocking one doesn't necessarily affect the other: that's exactly why the lesson on [robots.txt and controlling AI crawlers](/en/academy/setup-technique/robots-txt-et-crawlers-ia) recommends treating them separately rather than as a single "AI" block to allow or deny wholesale.

## An overview of the major crawlers by provider

As a reference point, keeping in mind this list evolves with provider announcements:

- **OpenAI**: `GPTBot` (training), `OAI-SearchBot` (indexing for ChatGPT's web search), `ChatGPT-User` (browsing triggered by a live user request).
- **Anthropic**: `ClaudeBot` (collection, including for search-related uses), plus `Claude-User` and `Claude-SearchBot` depending on recent browsing use cases.
- **Perplexity**: `PerplexityBot` (indexing), `Perplexity-User` (real-time browsing triggered by a user question).
- **Google**: `Google-Extended` (controls usage by Gemini models), distinct from `Googlebot`, which handles classic search indexing.
- **Microsoft**: `Bingbot`, whose data partly feeds Copilot's answers.

Each provider documents its user-agents on its own technical pages; these declarations change over time, so it's worth re-checking periodically rather than relying on a fixed list.

## Spotting AI crawlers in server logs

User-agent declarations remain the most reliable source, but they aren't enough on their own: a bot may present itself under a name that's no longer current, or an unscrupulous operator could spoof a well-known user-agent. For a deeper check:

1. **Access raw server logs** (through your host, FTP/SSH access, or a tool like Cloudflare, which centralizes this data for sites using it as a proxy).
2. **Filter requests by user-agent string**, searching for known names (`GPTBot`, `ClaudeBot`, `PerplexityBot`, etc.) with a simple command-line tool (`grep` on a log file) or a log analysis dashboard.
3. **Cross-reference the user-agent with the originating IP address**: legitimate providers publish IP ranges or verification methods (reverse DNS lookup) that let you confirm a request genuinely comes from them and not an impersonator.
4. **Note the frequency and the pages visited**: a crawler that keeps coming back to the same core content pages is a sign of ongoing interest in that content, unlike a single one-off visit.

This log analysis is more technical than reading Google Search Console, but it offers a view neither GSC nor Bing Webmaster Tools provide: concrete proof that an AI crawler actually visited, independent of any indexing or citation observed afterward.

## What the presence of these crawlers reveals (and doesn't)

Seeing an AI crawler in your logs confirms your site is reachable and being read by that bot — a positive signal on the first step of the chain described in the lesson [why everything starts with the technical layer](/en/academy/setup-technique/pourquoi-tout-commence-par-la-technique). It doesn't, however, guarantee a future citation in a generated answer, nor a positive judgment by the model on content quality: being read by a crawler is a necessary condition, not a sufficient one.

Conversely, the total absence of a particular crawler from your logs over an extended period is a warning sign worth checking against your robots.txt and overall site accessibility, rather than a hasty conclusion that the provider simply isn't interested in your content.

## The special case of sites behind a CDN

If your site sits behind a CDN or a protection service like Cloudflare, your origin host's logs may only show a fraction of real traffic, with most requests absorbed and filtered upstream. In that case, the CDN's own dashboard holds the more complete crawl data, often with a ready-made categorization of known bots ("Verified bots"), which simplifies identification compared to manually parsing raw logs.

## Turning this data into action

Manually monitoring logs takes time and technical skill that not every team has in-house. For a tooled approach that automates this tracking and ties it directly to a robots.txt audit, the lesson [tracking AI crawlers and auditing your robots.txt with Vurto](/en/academy/avec-vurto/suivre-les-crawlers-ia-avec-vurto) covers how to industrialize this monitoring.

This technical module wraps up here. The logical next step is diving into the core of GEO with [why AI sometimes ignores your site](/en/academy/comprendre-le-geo/pourquoi-les-ia-ignorent-votre-site).

### Technical SEO basics

### [How a search engine (and an AI) reads your page](https://vurto.ai/en/academy/bases-seo-techniques/comment-un-moteur-lit-une-page) (2026-05-28)

A search engine and a generative AI don't "look" at your page the way a human visitor does. They fetch a stream of HTML, turn it (or not) into a data structure called the DOM, then extract raw text from it and split that text into chunks for indexing or for use as an answer. Understanding this path changes, concretely, how you should write and structure your pages.

## The path from raw HTML to rendering

When a crawler visits a URL, it first receives an HTTP response containing the raw HTML sent by your server, before any JavaScript execution. Most bots analyze this initial HTML first, because fully executing it (the "render") costs computing time and memory, multiplied by the billions of pages crawled every day.

Google can render JavaScript in a second pass, through a separate queue that can take anywhere from a few seconds to several days depending on current load and the priority assigned to your site. In other words, there's a gap — sometimes significant — between when your raw HTML is read and when its fully rendered version (after JavaScript execution) is taken into account. Generative AI crawlers (ChatGPT, Claude, Perplexity in particular) are generally more limited on this point: they tend to rely more on static HTML or on versions already indexed by third-party search engines, rather than on a full, systematic JavaScript render on every visit.

This difference in handling explains why two tools can end up with a different picture of the same page: one may have waited for the full render, the other may not have.

## What a crawler sees vs. what a browser sees

A browser executes CSS and JavaScript, loads fonts, applies animations, and displays a complete visual render, pixel by pixel. A crawler sees nothing visual: it processes a string of characters (the HTML) and, possibly, a tree structure built from that string.

Concretely, if content exists only visually — information encoded solely in an image with no alt text, text generated by a CSS animation, content injected after a user click, a tooltip that only appears on hover — it's likely that no bot will ever perceive it as usable text. It's not that it deliberately ignores it: it simply has no way to reproduce the human interaction that triggers its appearance.

## The DOM and text extraction

The DOM (Document Object Model) is the tree representation of the HTML: each tag becomes a "node" with parents, children, and attributes. Engines and AIs extract usable text from this tree structure, relying in particular on:

- the tag hierarchy (`<header>`, `<nav>`, `<main>`, `<article>`, `<footer>`) to distinguish main content from peripheral elements;
- heading tags (`<h1>` through `<h6>`) to identify the logical structure of the content;
- the order of the HTML in the document (not the visual order imposed by CSS).

An often-overlooked point: if you use `position: absolute` or `flex-direction: row-reverse` to visually reorder blocks, the order read by the machine remains the one in the source code, not the one displayed on screen. A paragraph placed last in the HTML but visually moved to the top of the page via CSS will still be treated, by most extraction systems, as the last content block — which can distort the logical structure perceived by an automated reader.

## JavaScript: the invisible-content trap

Sites built with client-side JavaScript (some React, Vue, or Angular apps without server-side rendering) carry a real risk: if the text content only appears after the script runs, and the bot visiting the page doesn't execute it or only partially does, that content is simply absent from what gets indexed. The risk is especially high for blocks loaded via aggressive lazy loading, content hidden behind a tab closed by default, or data injected through an asynchronous API call after the page's initial load.

Several approaches limit this risk:

```html
<!-- Server-side rendering (SSR): the initial HTML already contains the text -->
<article>
  <h1>Title visible immediately in the source HTML</h1>
  <p>This paragraph is present before any JS runs.</p>
</article>
```

SSR (server-side rendering), static rendering (SSG), or progressive hydration guarantee that essential text is present from the very first HTTP response, without depending on the visiting bot executing JavaScript. These approaches aren't just technical options reserved for developers: they have a direct impact on how visible your content is to any automated system, whether a classic search engine or a generative AI browsing the web to build an answer.

## What this concretely means for your content

Three direct implications for writing and building your pages:

1. Important text content (arguments, answers to questions, figures) must exist in the raw HTML, not only after interaction or animation.
2. The structure of the source code should reflect the logical reading order, independent of the visual formatting imposed by CSS.
3. Clear semantic markup helps separate main content from noise (menus, ads, repetitive footers) — a point covered in the lesson on [semantic HTML5](/en/academy/bases-seo-techniques/html5-semantique).

To check concretely what a bot can read, compare the raw HTML returned by your server (right-click → "view page source", not the inspector that shows the post-render DOM) with what's displayed on screen. Any notable gap — a block of text present visually but absent from the source code — is a signal to fix as a priority.

This question of the technical path content takes ties directly into two topics from the next module: how an LLM [chunks your content](/en/academy/comprendre-le-geo/le-chunking-comment-le-llm-decoupe-le-contenu) once the text is extracted, and how to manage crawler access via the [robots.txt file](/en/academy/setup-technique/robots-txt-et-crawlers-ia).

### [The title tag](https://vurto.ai/en/academy/bases-seo-techniques/la-balise-title) (2026-05-29)

The `<title>` tag remains, even today, one of the most-read signals by search engines and one of the first elements a generative AI uses to understand what a page is about before even reading its content.

## The exact role of the title tag

The `<title>` serves three distinct purposes: it's the text shown in the browser tab, it's (most often) the clickable link shown in search results, and it's a strong signal used by ranking algorithms to assess a page's relevance to a query. For a generative AI browsing the web to answer a question, a clear, specific title helps it quickly decide whether the page is worth opening and citing, before even analyzing the body text.

```html
<head>
  <title>Title Tag SEO: Role, Length, and Examples | Vurto Academy</title>
</head>
```

This triple role explains why the title should never be treated as a minor technical detail left to the CMS: it's often the very first contact, human or automated, with your page.

## Recommended length

There's no strict character limit: Google and other engines truncate the display based on the pixel width of the text, which varies by character (a lowercase "i" takes up less room than an uppercase "W"). In practice, aiming for roughly 50 to 60 characters (including spaces) generally avoids truncation on both desktop and mobile. Beyond that, the title still exists in the HTML and can still be read in full by an AI, but it risks being cut off in the visual display of search results, which can hurt click-through rate if the essential information sits at the end of the sentence.

A title that's too short (under 30 characters) isn't ideal either: it often leaves unused space in the display and misses the chance to clarify the intent or entity involved.

## Writing a title around intent and entity

A good title answers two questions at once: which search intent does this page satisfy, and which entity (brand, product, concept) is clearly associated with this content? Avoid vague phrasing that could apply to any page on the site.

**Before**: `Home | Solutions`
**After**: `Invoicing Software for Freelancers | ProductName`

**Before**: `Our offer`
**After**: `2026 Pricing and Subscription Plans | ProductName`

**Before**: `Blog`
**After**: `How to Reduce Payment Delays: 4 Concrete Methods`

The second example contains an entity (the product), an intent (comparing pricing), and a time marker that helps signal how fresh the information is — something generative AIs factor in when choosing between several competing sources on the same topic.

## One title per page, never duplicated

Every page should have a unique title across the entire site. Duplicated or near-identical titles (often auto-generated by a poorly configured CMS, or by a non-personalized product page template) dilute the engine's ability to tell your pages apart, and make it harder for an AI trying to pick the most relevant page among several similar URLs. A quick audit involves extracting the list of every title on the site and spotting duplicates or near-duplicates — particularly useful after a CMS migration or a bulk addition of product pages.

## Common mistakes to avoid

- Repeating the brand name at the start of every page without adding useful information, wasting the first characters, the most visible ones in search results.
- Stacking keywords without forming a readable sentence (`Invoicing, software, invoice, SMB, free`), a practice that brings no benefit anymore and can even hurt perceived readability.
- Leaving the default title generated by a CMS (often identical to the H1, or empty), which robs the page of a chance to state its own specific intent.
- Frequently changing the title of a page that already ranks well, without a substantive reason — an unstable signal that can temporarily disrupt how the engine re-evaluates the page.

## The title in AI-generated results, and how it evolves over time

In search interfaces that embed AI-generated answers, the title keeps a particular role: when the AI cites several sources to build its answer, this text often serves as the visual anchor identifying each source in the reference list. A vague title ("Home", "Product page") makes that identification harder and can reduce the odds that a user then clicks through to your page to verify the information — even if the content was actually used in the answer.

This is also why a title built around a clear entity (the product name, the brand, the concept covered) stays valuable even in a world where the user no longer scans a list of ten blue links but reads a synthesis directly: the title still carries the identity of the cited source.

The title isn't set in stone either. It's legitimate to rework it when:

- the page's content has significantly evolved and the title no longer reflects it;
- the click-through rate observed in tracking tools (Search Console or equivalent) is unusually low relative to the page's position;
- the same intent is now covered by a competitor with more precise phrasing than yours.

Conversely, changing the title of an already stable, well-ranked page without a reason is an unnecessary risk: the engine has to re-evaluate the relevance of the new text, which can temporarily cause the ranking to fluctuate, with no guarantee of improvement.

The title rarely works alone: paired with a well-written [meta description](/en/academy/bases-seo-techniques/la-meta-description) and a coherent [heading structure](/en/academy/bases-seo-techniques/les-balises-hn), it forms the first impression an engine or an AI gets of your page before even analyzing its full content.

### [The meta description](https://vurto.ai/en/academy/bases-seo-techniques/la-meta-description) (2026-05-30)

The meta description has no proven direct effect on a page's ranking, but it influences click-through rate in search results and serves as a ready-made summary for an AI that needs to quickly synthesize what a page is about.

## What the meta description actually does

Contrary to a common belief, the meta description is not a direct ranking factor for Google: stuffing it with extra keywords won't mechanically improve your position. Its real role lies elsewhere, playing out on three levels:

- it's often displayed as the snippet under the link in search results, and therefore influences the decision to click, all else being equal in the ranking;
- when an engine judges that a passage from the page's content better matches the user's query, it replaces your description with an automatically generated excerpt — yours is only a suggestion, not a guaranteed display;
- for a generative AI quickly listing or summarizing several sources, a well-written meta description can serve as a reference summary if the page's content confirms what it announces, speeding up its synthesis task.

```html
<head>
  <meta name="description" content="Learn how to write an effective meta description: ideal length, click-focused phrasing, concrete before/after examples.">
</head>
```

## Writing to entice AND inform

A good meta description does two things at once: make people want to click (a concrete benefit, an engaging tone) and accurately inform about the page's content (so an AI reading it without opening the page understands what it's about). These two goals aren't contradictory as long as you avoid empty marketing filler.

**Before**: `Welcome to our site, discover our best offers and innovative solutions for your business.`

**After**: `Compare 4 methods to reduce customer payment delays, with the pros and limits of each.`

**Before**: `Our team of experts supports you every day with all your needs.`

**After**: `A 6-step guide to setting up e-invoicing before the 2027 regulatory deadline.`

The two second examples promise specific, verifiable content (a number of methods or steps, a clear topic, sometimes a deadline), which helps both a human scanning results and an AI that needs to judge a source's relevance before citing or summarizing it.

## Recommended length

As with the title, there's no strict character limit, but display tends to get truncated around 150 to 160 characters on desktop (slightly less on mobile, where available space is more limited). Write the essential sentence first, even if the end gets cut off visually — the full text remains readable in the source code by bots, even beyond the visual display limit.

## When the meta description isn't enough

If your meta description promises content the page doesn't actually deliver (a checklist announced but missing, a figure cited but not found in the text, a number of steps that doesn't match what follows), an AI checking consistency between the summary and the actual content may discard the page as an unreliable source, or worse, flag that inconsistency in its answer. Consistency between promise and content therefore matters more than the hook itself: a plain but accurate description beats a catchy but misleading one.

## Common mistakes

- Leaving the field empty and relying solely on automatic extraction, which produces an unpredictable and sometimes unengaging result.
- Duplicating the same description across several pages of the same site, making them impossible to tell apart in results.
- Stuffing the description with keywords instead of forming a natural sentence, an outdated practice that hurts readability.
- Copying the first sentence of the content verbatim, without any effort to synthesize, which rarely produces the most compelling or informative summary.

## Adapting the description by page type

Not every page calls for the same phrasing. A product page benefits from mentioning a concrete benefit and, where relevant, a differentiating detail (delivery time, warranty, compatibility). A blog post or guide benefits from announcing the content's structure (number of steps, methods, criteria) rather than a generic hook. A category page benefits from a description that clearly states the scope covered ("Comparison of N available models, filterable by price and use case"), to avoid any ambiguity with neighboring pages on the same site.

In all three cases, the principle stays the same: give a pressed-for-time reader, human or automated, enough to judge in one sentence whether the page matches what they're looking for.

A well-written meta description doesn't guarantee it will be displayed as-is, either: the only way to know what Google or another platform actually shows for a given page is to search for that page directly and look at the result. If the displayed snippet consistently differs from your description on a targeted query, it's often a sign the engine judges another passage of the page more relevant to that specific query — a useful signal for adjusting either the description or the content itself.

The meta description works in tandem with the [title tag](/en/academy/bases-seo-techniques/la-balise-title) and with [Open Graph tags](/en/academy/bases-seo-techniques/les-balises-open-graph) for social sharing — together, the three form your page's identity card as seen from the outside, before anyone even reads its full content.

### [Heading tags (H1, H2, H3)](https://vurto.ai/en/academy/bases-seo-techniques/les-balises-hn) (2026-05-31)

Heading tags (H1, H2, H3...) give the engine and the AI a hierarchical outline of your content, independent of its visual styling. A clean hierarchy makes the page easier to understand as a whole and easier to split into usable chunks.

## Correct semantic hierarchy

Heading tags form a logical tree: the H1 is the page's main title, the H2s are its main sections, the H3s are subsections of those H2s, and so on. This isn't a styling tool (making text bigger) — it's a tool for structuring meaning.

```html
<h1>Complete Guide to E-Invoicing</h1>
  <h2>Why e-invoicing is becoming mandatory</h2>
  <h2>Accepted file formats</h2>
    <h3>Factur-X format</h3>
    <h3>UBL format</h3>
  <h2>How to choose a dematerialization platform</h2>
```

Skipping a level (jumping from an H1 straight to an H3, for instance) breaks this logic and makes it harder for a machine to understand at what level of detail a given passage sits.

## One H1 per page

The simplest rule to remember: a page should have a single H1, summarizing the page's central topic — generally close to, but not necessarily identical to, the content of the [title tag](/en/academy/bases-seo-techniques/la-balise-title). Multiple H1s on the same page (common with certain themes or visual builders, or with layout components reused without much thought) dilute the signal about the main topic.

The H1 and the title can differ slightly in wording: the title is often optimized for display in search results (including the brand name, for instance), while the H1 can focus purely on the topic, without a brand element. This difference is normal and not a problem, as long as both texts clearly describe the same central topic without contradicting each other.

## Structuring headings as questions to ease chunking

An LLM splits a page's content into blocks (chunks) to index it or extract an answer from it — a mechanism covered in the lesson on [chunking](/en/academy/comprendre-le-geo/le-chunking-comment-le-llm-decoupe-le-contenu). Phrasing your H2s and H3s as concrete questions increases the odds that a whole block can autonomously answer a user query, without depending on the context of neighboring paragraphs.

**Before**: `Benefits`
**After**: `What are the benefits of e-invoicing for a small business?`

**Before**: `Implementation`
**After**: `How to set up e-invoicing in 3 steps?`

An H2 phrased as a question, followed by a direct answer in the first sentences of the paragraph, matches exactly the format an AI looks to extract when building its own answer. This logic doesn't apply uniformly to every H2 on a page: an article's outline can legitimately mix declarative headings ("Definition and stakes") with interrogative ones ("How much does setup cost?"), as long as each section, taken in isolation, remains understandable without depending on the previous paragraph.

## Common mistakes

- Using headings purely for their visual rendering (picking an `<h3>` because it "looks the right size"), ignoring the logical hierarchy.
- Putting the company name or a tagline in the H1 instead of the page's actual topic.
- Multiplying heading levels on short content, needlessly fragmenting the information and hurting overall readability.
- Letting a theme or plugin auto-generate empty or redundant H2s (e.g. "Summary", "Introduction" with no real content underneath).
- Phrasing two nearly identical H2s on the same page ("Our advantages" and "Our strengths"), which blurs the distinction between sections for an automated reader.
- Artificially stuffing keywords into headings at the expense of readability for a human reader — the structuring logic should never override the clarity of the sentence.

## Checking the hierarchy of your pages

A quick pass through the browser inspector (or a tool listing a page's heading tags) is enough to spot inconsistencies: multiple H1s, skipped levels, or empty headings. Fixing this structure usually requires no visual change if your CSS is already decoupled from the semantic HTML — see the lesson on [semantic HTML5](/en/academy/bases-seo-techniques/html5-semantique) to go further on that decoupling.

## The length and content of a good H2

An overly short H2 ("Benefits", "FAQ", "Pricing") works as a visual landmark for a human scanning the page, but conveys little information to a machine that has to judge, without context, the relevance of the block that follows. Conversely, an overly long H2 (a full 20-word sentence) loses readability and scannability for the human reader. The recommended balance generally sits between 5 and 12 words — enough to state a full question or claim, without becoming a disguised paragraph.

Another useful habit: reread only the list of your H2s and H3s, without the body text, the way a tool building an automatic table of contents would. If that list alone lets you guess the page's content with reasonable accuracy, the hierarchy is doing its job. If it's just a string of vague words, that's a sign it needs reworking.

### [The canonical tag](https://vurto.ai/en/academy/bases-seo-techniques/la-balise-canonical) (2026-06-01)

The canonical tag tells search engines which URL should be treated as the reference version when several addresses display identical or near-identical content. Set up wrong, it can make entire pages disappear from results.

## The duplicate content problem

The same content accessible through several URLs (with or without a trailing slash, with sorting parameters, in a "print" version, via HTTP and HTTPS simultaneously) creates ambiguity: the engine has to guess which of these URLs to index and display. Without a clear signal, it decides on its own — sometimes differently from one engine to another — which dilutes the page's authority across several addresses instead of concentrating it on one.

This dilution has a concrete cost: external links pointing to different variants of the same page (some to the version with a trailing slash, others without) split their value across several URLs instead of concentrating it on a single reference address. A properly set canonical consolidates that signal onto one URL, strengthening its standing compared to a situation where authority remains scattered.

## The common case of URL parameters

Filters, sorting, and campaign identifiers often generate dozens of variants of the same page:

```
https://example.com/shoes?color=red&sort=price
https://example.com/shoes?sort=price&color=red
https://example.com/shoes?utm_source=newsletter
```

These three URLs display the same underlying content. Each should point, via its canonical tag, to the "clean" version of the page:

```html
<link rel="canonical" href="https://example.com/shoes">
```

## Syntax and best practices

The tag goes in the document's `<head>`, with an absolute URL (never relative) pointing to the version you want indexed. A relative URL (`/product/product-name` instead of `https://example.com/product/product-name`) can be misinterpreted depending on how the browser or bot resolves the context, and is an easily avoidable source of error:

```html
<head>
  <link rel="canonical" href="https://example.com/product/product-name">
</head>
```

A few rules to follow:

- a page pointing canonically to itself is the correct default situation (it's not an error — it's actually recommended);
- the canonical URL must be reachable (no redirect, no 404) and consistent with the protocol and subdomain actually in use (https vs http, www vs no www);
- indicate only one canonical URL per page — a canonical tag is not a list.

## When to set a canonical (and when not to)

Set a canonical when several URLs display identical or near-identical content (technical pagination, sorting, tracking). Don't use it as a substitute for a [redirect](/en/academy/bases-seo-techniques/redirections-et-erreurs-http): if an old page is no longer meant to exist, a 301 redirect is the right tool, not a canonical that leaves both URLs active.

A common misconception is treating the canonical as a strict order given to the engine. In reality, it's a strong suggestion: if other signals clearly contradict your choice (massive external links pointing to the "non-canonical" version, for instance), some engines may still choose to index a different URL than the one indicated. This stays rare when the markup is consistent with the rest of the site, but it explains why a canonical poorly justified by the page's actual context is more likely to be ignored.

## Common mistakes

- A canonical pointing to a different page by configuration error (often after a migration or CMS change), which can accidentally deindex a page.
- Canonicals contradicting each other (page A points to B, which points to C) — known as a canonical chain, and best avoided just like a redirect chain.
- Forgetting the canonical on multilingual versions, which should be combined with the [hreflang tag](/en/academy/bases-seo-techniques/la-balise-hreflang) rather than pointing them at each other as duplicates.
- Setting a canonical to a page blocked by robots.txt, which prevents the engine from verifying content consistency.

## The case of pagination and near-duplicate content

Pagination (page 1, page 2, page 3 of a list of articles) deserves particular attention: each results page displays different content, even if the overall structure looks similar. Setting a single canonical to page 1 for all subsequent pages amounts to asking the engine to ignore the specific content of pages 2, 3, and so on — rarely desirable if that content has its own value. Best practice, in the vast majority of cases, is to leave each pagination page canonical to itself, and rely on clear internal linking rather than the canonical to signal the relationship between pages.

The case of "almost" duplicate content (two very similar product pages differing only by color or size variant) also calls for judgment: if the difference matters to the user (and therefore to the query they type), each variant generally deserves its own indexing rather than being canonicalized to a single version.

For a full overview of every technical point to check before publishing, see the [complete technical checklist](/en/academy/bases-seo-techniques/checklist-technique-complete).

### [Open Graph tags](https://vurto.ai/en/academy/bases-seo-techniques/les-balises-open-graph) (2026-06-02)

Open Graph tags control the preview of your page when it's shared on social media, in a messaging app, or cited in a conversation with a generative AI capable of showing a rich preview.

## What Open Graph tags are for

Without Open Graph markup, a platform displaying a link preview (social network, messaging app, some AI assistants) has to guess which title, image, and description to use — often awkwardly pulling from the page's content: a random first image found in the HTML unrelated to the topic, a title truncated mid-sentence, or a missing description. Open Graph tags take back control of that preview and guarantee a consistent presentation, no matter where the link gets shared.

This standard, originally designed by Facebook, has since been adopted as the de facto reference by nearly every platform that displays link previews, which makes it a low-effort, high-benefit piece of markup — a handful of tags cover most use cases.

## The essential tags

```html
<head>
  <meta property="og:title" content="Complete Guide to E-Invoicing">
  <meta property="og:description" content="Everything you need to know about formats, obligations, and e-invoicing platforms in 2026.">
  <meta property="og:image" content="https://example.com/images/invoicing-guide-og.jpg">
  <meta property="og:url" content="https://example.com/e-invoicing-guide">
  <meta property="og:type" content="article">
</head>
```

- `og:title`: the preview's title, often close to the `<title>` but can be reworded for a social-sharing context.
- `og:description`: summary shown under the title in the preview.
- `og:image`: the preview card's image — the element that most influences click-through on a share.
- `og:url`: the canonical URL of the shared page.
- `og:type`: the nature of the content (`article`, `website`, `product`...).

These five tags cover the vast majority of use cases. Other properties exist for more specific needs (`og:site_name` for the site's name, `article:author` or `article:published_time` for editorial content), but they remain secondary compared to the core foundation formed by the title, description, and image.

## Best practices for the Open Graph image

An incorrectly sized or missing image results in an empty or poorly cropped preview, which sharply reduces engagement on a share. The most common recommendations:

- recommended size around 1200 x 630 pixels (1.91:1 ratio);
- reasonable file size for a fast preview load;
- readable text even at small size if the image contains overlaid text;
- a centered main subject, since some platforms crop the image differently depending on the display format (wide preview, square thumbnail);
- a site-wide default image for pages that don't have a specific one, rather than leaving the field empty.

## An impact beyond social media

Some generative AI interfaces that display search results or source citations also use these metadata fields to build a visual preview of the cited source. Complete Open Graph markup consistent with the page's actual content therefore indirectly contributes to how your brand is presented in these contexts.

## Common mistakes

- Missing Open Graph tags, leaving each platform to guess a preview inconsistently.
- `og:image` pointing to a relative URL instead of an absolute one, which breaks the display on most platforms.
- `og:description` different enough to contradict the page's actual content.
- Forgetting to update these tags after a visual redesign or a major content change.

## Twitter Cards, and testing your markup before publishing

Some platforms use their own standard alongside Open Graph. X (formerly Twitter), for instance, reads its `twitter:card`, `twitter:title`, and `twitter:image` tags first if present, and falls back to Open Graph tags in their absence:

```html
<meta name="twitter:card" content="summary_large_image">
<meta name="twitter:title" content="Complete Guide to E-Invoicing">
<meta name="twitter:image" content="https://example.com/images/invoicing-guide-twitter.jpg">
```

In most cases, faithfully duplicating the Open Graph values into these complementary tags is enough; there's no need to write entirely distinct content for each platform, unless you're specifically targeting a different audience on one of them.

A misconfigured Open Graph preview usually isn't visible while browsing the site itself — the error only shows up at the moment of sharing, often after the link has already circulated. Debugging tools provided by the major social platforms let you simulate the preview before publishing and force a cache refresh if an image or title was changed after a first share: these platforms cache the generated preview and don't automatically regenerate it on every visit.

Open Graph tags naturally complement the [title tag](/en/academy/bases-seo-techniques/la-balise-title) and the [meta description](/en/academy/bases-seo-techniques/la-meta-description): together, the three define how your page presents itself, whether in a search engine or on a share. To go further on structured markup, the logical next lesson is on [implementing Schema.org structured data](/en/academy/bases-seo-techniques/schema-org-implementer).

### [The hreflang tag (multilingual)](https://vurto.ai/en/academy/bases-seo-techniques/la-balise-hreflang) (2026-06-03)

The hreflang tag tells search engines which language or regional version of a page to show a given user, when the same content exists in several languages or targets several countries.

## Use case: when hreflang becomes necessary

As soon as a site offers the same page (or an equivalent page) in several languages, or the same language adapted for several countries (French from France vs. French from Belgium, for example), the engine needs to know which one to show which user. Without hreflang, it risks showing the English version to a French-speaking user, or indexing the different versions as competing duplicate content rather than as complementary versions.

This perceived-duplication risk is the most misunderstood part of the tag: without hreflang, two very similar English and French pages can be read as two redundant translations of the same content rather than as two versions aimed at distinct audiences, which can hurt the ranking of both instead of reinforcing each other.

Conversely, a site offering only one language has no need for hreflang at all: adding the tag without real multilingual content brings nothing and needlessly complicates site maintenance.

## Correct syntax

The declaration usually goes in the `<head>` of each page, with a link to every available language version, including itself:

```html
<head>
  <link rel="alternate" hreflang="fr" href="https://example.com/fr/guide" />
  <link rel="alternate" hreflang="en" href="https://example.com/en/guide" />
  <link rel="alternate" hreflang="es" href="https://example.com/es/guide" />
  <link rel="alternate" hreflang="x-default" href="https://example.com/en/guide" />
</head>
```

The language code follows the ISO 639-1 standard (`fr`, `en`, `es`), and can be combined with an ISO 3166-1 country code to target a specific regional variant (`fr-BE` for Belgian French, `en-US` for American English). Hreflang can also be declared through the XML sitemap instead of in each page's `<head>`, an approach often preferred on sites with a large number of multilingual pages, since it centralizes every declaration in a single file that's easier to generate and maintain automatically.

## Reciprocity: the most commonly broken rule

Every page declaring an alternate version must itself be declared back by that alternate version. If the French page points to the English page via hreflang, the English page must point back to the French page. A non-reciprocal hreflang link is generally ignored by engines, which cancels the markup's effect without any error being clearly shown.

This rule becomes hard to maintain manually once a site passes three or four languages: adding a new language means updating the hreflang declaration on every existing page in every other language, not just on the newly created pages. This is one reason automatic generation through a content management system or a dedicated sitemap sharply reduces the risk of error compared to a manual, page-by-page declaration.

## The often-forgotten x-default

The `x-default` value indicates which version to show users whose language or location doesn't match any of the explicitly declared versions. Without it, those users get an arbitrary choice from the engine, sometimes inconsistent with the site's intent. `x-default` generally corresponds to the site's international or English version, but it can also point to an explicit language-selection page if the site has no obvious default version.

## Common mistakes

- Non-reciprocal hreflang links between language versions.
- Incorrect language or country codes (mixing up the language code and country code, e.g. writing `fr-FR` as `FR-fr`).
- hreflang pointing to a URL that redirects or returns a 404, instead of the final URL.
- Combining hreflang with a [canonical tag](/en/academy/bases-seo-techniques/la-balise-canonical) that points to a single "master" version, which contradicts the very purpose of hreflang — each language version should have its own canonical, pointing to itself.
- Forgetting x-default on a site with a large international audience.
- Declaring a language in hreflang without the target page's content actually being translated into that language (a duplicated but untranslated page, left in place temporarily).

## Checking your implementation

An incorrect hreflang deployment often goes unnoticed visually (the site works normally for a direct visitor) but can durably skew international rankings. Systematically checking link reciprocity, especially after adding a new language, avoids this kind of silent error.

A simple test is to open the source code of each language version and check, for every declared hreflang link, that the target page declares an equivalent link back. On a large site, this check is more realistic with an automated crawling tool that lists inconsistencies than done manually, page by page.

For the full set of points to check before going live, see the [complete technical checklist](/en/academy/bases-seo-techniques/checklist-technique-complete), and for managing old URLs during a multilingual reorganization, the lesson on [redirects and HTTP errors](/en/academy/bases-seo-techniques/redirections-et-erreurs-http).

### [Redirects and HTTP errors](https://vurto.ai/en/academy/bases-seo-techniques/redirections-et-erreurs-http) (2026-06-04)

HTTP status codes tell browsers, search engines, and AI crawlers what happened to a URL: redirected, not found, or server error. Poorly managed, they fragment crawling and erode the trust granted to a site.

## 301, 302, and 308: what's the difference

All three codes signal a redirect, but with different implications:

- **301 (permanent redirect)**: the URL has changed for good. It signals that the page's authority and history should transfer to the new address. Use it for any lasting URL migration.
- **302 (temporary redirect)**: the original URL remains the long-term reference (for example, a product page temporarily unavailable that redirects to the category page). An engine will generally keep the original URL in its index.
- **308 (permanent redirect, method preserved)**: equivalent to a 301 but guarantees the HTTP method of the request (GET, POST) is preserved during the redirect — useful in particular for form submissions.

```
HTTP/1.1 301 Moved Permanently
Location: https://example.com/new-url
```

Using a 302 for a permanent change is the most common mistake: the engine keeps treating the old URL as the reference, which delays transferring authority to the new page. This confusion is even more common because many frameworks and hosting platforms use 302 as the default redirect code when the desired type isn't explicitly specified — it's worth actually checking, with an HTTP header inspection tool, which code your configuration really returns rather than assuming it.

## Redirect chains to avoid

Every redirect adds a step the crawler must follow before reaching the final content. A chain of several successive redirects (A → B → C → D) slows down crawling, potentially dilutes the transferred signal, and some bots give up following after a limited number of hops. Best practice is to systematically redirect straight to the final destination (A → D), updating existing redirects with each new migration instead of stacking them.

These chains most often form gradually, migration after migration, with no single step being wrong in isolation: a page migrates once (A → B), then again months later (B → C), with nobody thinking to update the original redirect so it points straight to C. A periodic audit of existing redirects, especially on sites that have gone through several redesigns, helps spot and flatten these chains.

## 404 errors: normal, but worth monitoring

A page removed with no relevant replacement content should return a 404 error ("not found") — that's the correct HTTP behavior, not an anomaly in itself. The problem arises when:

- active internal links on the site still point to a 404 page (a broken link to fix);
- a large number of 404s come from pages that had value (backlinks, historical traffic) and would deserve a 301 redirect to equivalent content instead of a flat error.

A well-designed 404 page (clear message, link to internal search or main categories) still helps a human visitor who lands on a broken link, even though it changes nothing about how bots handle the 404 code itself.

## 5xx errors: a signal of degraded trust

A server error (500, 502, 503) means the page couldn't be delivered at all, regardless of its content. Frequent or repeated 5xx errors, especially at the moment a crawler tries to access the site, degrade the perception of the site's technical reliability and can reduce how often it comes back to crawl your pages. These errors typically occur during load spikes, poorly prepared server updates, or resource limits being hit on undersized hosting — causes often unrelated to the content itself, but that still affect the site's overall perception.

## How to check your site's HTTP statuses

A page's HTTP status isn't visible to the naked eye in a regular browser: a page returning a 500 error can display a generic message with no clear sign of the actual code on screen. The network inspection tools built into browsers (the "Network" tab in developer tools) show the exact status code returned for each request, letting you concretely verify that a redirect announced as permanent actually returns a 301, or that a page assumed to be working isn't quietly returning a 5xx error. On a large site, a periodic audit with a dedicated crawling tool remains the most reliable way to catch these anomalies at scale, rather than a manual, page-by-page check.

## Impact on crawling and trust

A site riddled with redirect chains, unhandled 404s, and intermittent 5xx errors sends a signal of poor technical maintenance. Crawlers, human and automated alike, generally place more confidence in a site whose URLs are stable and whose errors are rare and properly handled. Regularly checking these statuses is one of the points covered in the [complete technical checklist](/en/academy/bases-seo-techniques/checklist-technique-complete), alongside crawler access control covered in the lesson on [robots.txt](/en/academy/setup-technique/robots-txt-et-crawlers-ia).

Finally, don't confuse a redirect with a [canonical tag](/en/academy/bases-seo-techniques/la-balise-canonical): the former makes one URL disappear in favor of another, while the latter leaves both URLs active while indicating which one to prefer.

### [Semantic HTML5](https://vurto.ai/en/academy/bases-seo-techniques/html5-semantique) (2026-06-05)

Semantic HTML5 means choosing tags that describe the nature of the content (`<article>`, `<nav>`, `<main>`) rather than generic tags (`<div>`, `<span>`) that carry no meaning. This difference directly helps engines and AIs isolate useful content from the rest of the page.

## Why semantics matter

A `<div>` says nothing about its content: it could just as easily contain an article's main text, an ad block, or a navigation menu. A semantic tag, on the other hand, carries explicit information about the role of the block it wraps. For a machine that has to decide which part of a page is the main content worth indexing or citing, that information is directly usable, without having to guess from uncertain heuristics (position, CSS class, amount of text).

Before HTML5, developers already used naming conventions to give their tags meaning (`<div class="header">`, `<div class="nav">`), but these CSS classes carry no standardized semantic value: an engine can't be certain that a class named "header" actually corresponds to a header, since any developer can name their classes however they like. Semantic HTML5 tags, on the other hand, form a shared, unambiguous vocabulary, interpreted the same way by every system that reads HTML.

## The key semantic tags

```html
<body>
  <header>
    <nav><!-- main menu --></nav>
  </header>

  <main>
    <article>
      <h1>Article title</h1>
      <section>
        <h2>First part</h2>
        <p>...</p>
      </section>
      <section>
        <h2>Second part</h2>
        <p>...</p>
      </section>
    </article>
    <aside><!-- related, non-essential content --></aside>
  </main>

  <footer><!-- footer links, legal notices --></footer>
</body>
```

- `<header>`: page or section header (not necessarily the visual top of the page).
- `<nav>`: block of navigation links.
- `<main>`: the page's single main content — only one `<main>` tag per page.
- `<article>`: self-contained content that would make sense even isolated from the rest of the page (a blog post, a product sheet).
- `<section>`: a thematic grouping within an article or page — use it only when the section has its own logical heading (generally a heading tag), not as a mere visual split.
- `<aside>`: related but non-essential content (associated links, an ad block).
- `<footer>`: page or section footer.

## Isolating useful content from the noise

A web page rarely contains only main content: navigation menu, cookie banner, ad blocks, and a footer with legal links and social icons often make up a large share of the HTML. Without semantic structuring, a machine has to guess how much of the extracted text is actually relevant. Correct markup with `<nav>`, `<aside>`, and `<footer>` clearly separated from `<main>` and `<article>` reduces this noise and concentrates attention on the content that actually matters — an issue directly tied to how [an engine reads your page](/en/academy/bases-seo-techniques/comment-un-moteur-lit-une-page).

On some sites, navigation content and peripheral blocks make up a larger share of the HTML than the main content itself, particularly on short pages. In that case, precise semantic structuring becomes even more decisive so a machine doesn't overweight that noise relative to the text that actually matters.

## Common mistakes

- Coding everything in `<div>` with CSS classes, with no semantic tags at all — still common on sites generated by some visual builders.
- Using multiple `<main>` tags on the same page.
- Putting the main editorial content inside an `<aside>` by structural mistake.
- Using `<section>` as a plain styling container with no real thematic grouping.
- Nesting an `<article>` inside a `<nav>` or an `<aside>`, which contradicts the very logic of those tags.

## A benefit beyond SEO

Semantic HTML5 also improves accessibility: screen readers rely on these same tags to let a visually impaired user jump directly to the main content, without having to navigate the menu on every page. The same structuring effort therefore serves accessibility, classic SEO, and GEO simultaneously.

## Migrating an existing site to a semantic structure

Restructuring a site already built with generic `<div>`s doesn't necessarily require a full rebuild. In most cases, it means replacing existing wrapper tags with their semantic equivalent (`<div class="main-content">` becomes `<main>`, `<div class="site-footer">` becomes `<footer>`) without touching the CSS that already targets these elements by class. This gradual migration, template by template, improves the structure without risking breaking the existing visual layout — a realistic undertaking even on a site already in production.

To go further on what this structuring enables once the text is extracted, the lesson on [chunking](/en/academy/comprendre-le-geo/le-chunking-comment-le-llm-decoupe-le-contenu) explains how an LLM then splits this content into usable blocks, and the lesson on [Schema.org structured data](/en/academy/bases-seo-techniques/schema-org-comprendre) shows how to go even further in the clarity of the information conveyed.

### [Schema.org structured data: understanding it](https://vurto.ai/en/academy/bases-seo-techniques/schema-org-comprendre) (2026-06-06)

Schema.org is a standardized vocabulary that lets you describe precise, unambiguous facts in a page's code: a price, an average rating, an author, a publication date. This structured data gives engines and AIs information they no longer have to infer from free text.

## What Schema.org markup is

Schema.org is a shared vocabulary, jointly maintained by Google, Microsoft, Yahoo, and Yandex, that defines entity types (`Product`, `Article`, `Organization`, `FAQPage`...) and their properties (`price`, `author`, `datePublished`...). This vocabulary can be expressed in HTML through several syntaxes, the most widespread today being JSON-LD: a block of structured data in JSON format, inserted in a `<script>` tag, kept separate from the page's visual content.

Two other syntaxes exist — microdata and RDFa — which are inserted directly into the attributes of visible HTML tags rather than in a separate block. JSON-LD has established itself as the syntax recommended by most engines, mainly because it can be added, changed, or removed without touching the page's visual HTML, which considerably simplifies its maintenance over time.

```html
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "Understanding Schema.org structured data",
  "author": {
    "@type": "Organization",
    "name": "Vurto"
  },
  "datePublished": "2026-01-15"
}
</script>
```

## Why it gives AIs unambiguous facts

Text written in natural language can be ambiguous or require interpretation: "starting at €49" could mean a monthly price, an annual price, or a temporary promotional offer. An AI that needs to extract that price to answer a question then has to interpret the context, with a risk of error. The same problem applies to a date spelled out in words ("January 15th" without specifying the year) or a qualitative rating ("highly rated" with no numeric value): information a human interprets effortlessly thanks to context, but that a machine has to guess. The same price marked up in Schema.org removes that ambiguity:

```json
"offers": {
  "@type": "Offer",
  "price": "49.00",
  "priceCurrency": "EUR"
}
```

Here, the data is structured, typed, and open to no interpretation. This is exactly the kind of clarity generative AIs look for when they need to cite a precise fact (price, rating, availability, date) rather than paraphrase a passage of text.

## How it actually works

The principle rests on three elements: a **type** (`@type`, e.g. `Product`, `Article`, `FAQPage`) indicating the nature of the described entity, **properties** (`name`, `price`, `author`, `review`...) describing its attributes, and a **context** (`@context`, almost always `https://schema.org`) indicating which vocabulary is in use. This triad lets any system that understands the Schema.org vocabulary interpret the data in a strictly identical, unambiguous way.

## What Schema.org doesn't do

Structured markup doesn't replace quality content, and doesn't artificially improve the ranking of a page whose actual content is thin or misleading. Nor does it guarantee the appearance of rich snippets in search results: those depend on additional criteria and decisions specific to each engine. Schema.org's role is to make explicit information that's already present and accurate on the page — never to declare data that doesn't match the visible content, a practice engines penalize when detected.

This distinction is essential to grasp before starting implementation: Schema.org speeds up and makes more reliable the reading of content that's already solid, it never compensates for that content's absence. A site that added elaborate `AggregateRating` markup to a page with no real customer reviews would take on a penalty risk far greater than the hoped-for benefit.

## The link to entities and knowledge graphs

Beyond isolated facts (a price, a date), Schema.org also helps define entities and the relationships that connect them: which `Organization` authored which `Article`, which `Product` belongs to which `Brand`, which page belongs to which `BreadcrumbList`. This web of explicit relationships brings your content closer to how the knowledge graphs that engines and some AIs build behind the scenes connect entities, brands, and topics with each other. A site whose entities are clearly identified and linked through Schema.org makes this mapping work easier, rather than leaving the engine to infer those links from text alone.

## Toward implementation

Understanding the principle is the first step; actually implementing it — which types to prioritize, how to structure the JSON-LD, how to validate it — is covered in the next lesson on [implementing Schema.org structured data](/en/academy/bases-seo-techniques/schema-org-implementer). This factual clarity also connects to a central topic of the next module: how [citations, figures, and evidence](/en/academy/comprendre-le-geo/citations-chiffres-et-preuves) strengthen content's credibility in an AI's eyes, along with the [E-E-A-T](/en/academy/comprendre-le-geo/eeat-expertise-autorite-confiance) signals that partly depend on it.

### [Schema.org structured data: implementing it](https://vurto.ai/en/academy/bases-seo-techniques/schema-org-implementer) (2026-06-07)

Once [the principle of structured data is understood](/en/academy/bases-seo-techniques/schema-org-comprendre), the practical question remains: which Schema.org types to implement first, how to write the corresponding JSON-LD, and how to verify the markup is valid.

## Priority types

Five types cover the large majority of a site's needs:

- **`Organization`**: the company's identity (name, logo, social profiles, contact details) — usually set once, on the homepage or globally.
- **`Product`**: a product page, with price, availability, and reviews.
- **`FAQPage`**: a list of questions and answers, particularly useful for being picked up directly in a generative AI's answer.
- **`Article`**: editorial content (blog, news, guide), with author, publication date, and update date.
- **`BreadcrumbList`**: the breadcrumb trail, making a page's position in the site's hierarchy explicit.

Other types exist for more specific needs: `Event` for an event with a date and location, `LocalBusiness` for a business with a physical address and hours, `Review` for an individual review distinct from an aggregate rating, or `HowTo` for content structured as steps. It's better to correctly implement the five priority types before extending markup to more specialized types whose benefit depends heavily on the industry.

## Annotated JSON-LD example

```html
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Product name",
  "description": "Short, factual product description.",
  "brand": {
    "@type": "Brand",
    "name": "BrandName"
  },
  "offers": {
    "@type": "Offer",
    "price": "99.00",
    "priceCurrency": "EUR",
    "availability": "https://schema.org/InStock"
  },
  "aggregateRating": {
    "@type": "AggregateRating",
    "ratingValue": "4.6",
    "reviewCount": "128"
  }
}
</script>
```

Every property should reflect information actually present and verifiable on the page: the displayed price must match the marked-up price, and the average rating must correspond to real, viewable reviews.

For a frequently-asked-questions page:

```html
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "How long does setup take?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Setup generally takes between 2 and 5 business days depending on project complexity."
      }
    }
  ]
}
</script>
```

## Where to place the code

JSON-LD is usually inserted in the `<head>` or just before the closing `</body>`, with no dependency on the page's visual layout — unlike [semantic HTML5](/en/academy/bases-seo-techniques/html5-semantique) markup, which stays closely tied to the document's visual structure. This is one of JSON-LD's advantages over alternative syntaxes (microdata, RDFa) that require annotating the visible HTML tags directly. This independence from visual rendering also makes it possible to generate it dynamically from a database (a product catalog, a content management system) without manual work on each page, making it particularly well suited to sites with a large number of similar pages.

## Validation tools

Before publishing, two checks are essential: syntax validation (is the JSON well-formed, with no stray comma or missing quote) and semantic validation (do the types and properties used actually exist in the Schema.org vocabulary, and are they consistent with each other). Google's rich results testing tools and community Schema.org validators help catch both categories of errors before going live, and confirm that no required property is missing for the type in use.

This check deserves repeating after every change to the template that generates the markup, not just at initial implementation: a CMS update, a database structure change, or a theme update can silently break a dynamically generated JSON-LD block, with nothing visually flagging it on the page.

## Combining several types on one page

A page can legitimately carry several Schema.org blocks at once: a product page might combine `Product` for the listing itself, `BreadcrumbList` for the breadcrumb trail, and `Organization` for the identity of the brand selling it. These blocks can be declared separately in several `<script>` tags, or grouped into a single block via the `@graph` property, which lets you explicitly link several entities together within one JSON-LD document. This second approach becomes preferable as soon as entities need to reference each other (an article's author, a product's brand), since it avoids duplicating information across several independent blocks.

## Common mistakes

- Duplicated or contradictory markup across several JSON-LD blocks on the same page.
- Marked-up data that no longer matches the visible content after an update (price changed on the page but not in the JSON-LD).
- Using a Schema.org type that doesn't actually match the content (marking a blog page as `Product`).
- Malformed JSON that silently invalidates the entire block.
- Multiplying secondary types before correctly implementing and validating the priority types.

This lesson closes the technical markup portion of the module. The [complete technical checklist](/en/academy/bases-seo-techniques/checklist-technique-complete) that follows rounds up every point covered so far into an actionable list, before moving on to the fundamentals of GEO itself.

### [The complete technical checklist](https://vurto.ai/en/academy/bases-seo-techniques/checklist-technique-complete) (2026-06-08)

Before shifting focus to content production, this checklist rounds up the technical fundamentals covered in the previous modules into 20 points. Use it as a verification grid before any major publication or redesign.

## Structure and basic markup

1. Every page has a unique `<title>` tag, between 50 and 60 characters, built around intent and entity — see the [title tag](/en/academy/bases-seo-techniques/la-balise-title).
2. Every page has a unique meta description, informative and consistent with the actual content.
3. Every page contains a single `<h1>`, with no skipped level in the heading hierarchy.
4. H2s and H3s in in-depth content are phrased as concrete questions where relevant.
5. The HTML uses semantic tags (`<main>`, `<article>`, `<nav>`, `<header>`, `<footer>`) rather than generic `<div>`s everywhere.

## Content and rendering

6. Essential text content exists in the raw HTML, without depending exclusively on client-side JavaScript execution.
7. The order of the HTML in the source code matches the logical reading order, independent of the CSS.
8. Images carrying information have a descriptive `alt` attribute.

## URLs, duplication, and multilingual

9. Every page has a canonical tag pointing to an absolute, reachable, and consistent URL.
10. Pages with parameters (sorting, filters, tracking) canonically point to the clean version of the URL.
11. No canonical points to a page blocked by robots.txt or returning an error.
12. Every language version declares its hreflang tags, including one pointing to itself (if applicable).
13. Hreflang links are reciprocal across all versions.
14. An `x-default` is set if the site targets an international audience with no obvious default language.

## HTTP statuses

15. Permanent redirects use a 301 (or 308 if the HTTP method must be preserved), never a 302.
16. No chain longer than one redirect remains anywhere on the site.
17. Internal links never point to a URL returning a 404 or an error.

## Sharing and structured data

18. Open Graph tags (`og:title`, `og:description`, `og:image`, `og:url`) are present and consistent with the page's content.
19. Relevant Schema.org markup (`Organization`, `Article`, `Product`, `FAQPage`, `BreadcrumbList` depending on the page type) is in place and validated.
20. Data marked up in Schema.org matches exactly what's visible on the page (price, reviews, dates, availability).

## Using this checklist in practice

These 20 points aren't a one-shot exercise: revisit them at every redesign, migration, or new page type. Most of the errors listed here are silent — they break nothing visually for a visitor, but progressively degrade how engines and AIs read, trust, and cite your site.

Once these technical foundations are in place, the logical next step is working on the quality and credibility of the content itself: how an LLM [chunks your content](/en/academy/comprendre-le-geo/le-chunking-comment-le-llm-decoupe-le-contenu) into usable blocks, how to build the [E-E-A-T](/en/academy/comprendre-le-geo/eeat-expertise-autorite-confiance) signals that strengthen the trust granted to your pages, and why [citations, figures, and evidence](/en/academy/comprendre-le-geo/citations-chiffres-et-preuves) weigh so heavily in how an AI chooses to cite you instead of a competitor.

### Understanding GEO

### [The search landscape in 2026](https://vurto.ai/en/academy/comprendre-le-geo/paysage-de-la-recherche-en-2026) (2026-06-09)

In 2026, finding information online no longer means simply typing a query into Google and clicking a blue link. A growing share of searches now go through AI-generated summaries shown directly in the results page, or through chatbots like ChatGPT, Perplexity, Gemini, and Claude, which answer without ever displaying a list of sites to visit. Understanding this new landscape is the first step to knowing where to focus your visibility efforts.

## Classic search hasn't disappeared, it has changed shape

Google and Bing remain, by volume, the most-used entry points for finding information. But the results page itself has changed a great deal: featured snippets, knowledge panels, ads, and now generative summaries occupy a growing share of the visible space, especially on mobile. A classic organic link, even a well-ranked one, is now often pushed further down the page than a few years ago, below several automatically generated blocks. The habit of typing a short query into a search bar remains deeply ingrained for most users, but what they see in return has changed in nature: less a list of choices, more an already-digested answer.

This effect is reinforced by growing mobile usage, where the visible space before the first organic link is more limited than on desktop, and by voice search, which already pushes many assistants to formulate a single answer rather than a list of options to choose from.

## AI Overviews: search augmented by AI

For many informational queries, Google now displays a written summary at the top of the page, generated from several web sources retrieved and synthesized on the fly. This summary often lets the user get an answer without clicking through to a site at all — what's known as a "zero-click" search, where the query produces an answer rather than a visit. The underlying selection logic changes too: it's no longer about ranking ten competing links, but about choosing passages judged relevant and assembling them into a single, coherent answer. This mechanism, its implications for source sites, and its real weight in overall traffic are covered in the lesson [Google AI Overviews: the biggest AI channel by volume](/en/academy/comprendre-le-geo/google-ai-overviews).

## AI chatbots: a new entry point into information

Alongside traditional search engines, tools like ChatGPT, Claude, Gemini, Perplexity, and Copilot are increasingly used directly as a starting point for research, not just for drafting text or answering generic questions. Perplexity, for instance, was built from the ground up as an answer engine that consistently shows its sources. These tools work differently from one another: some lean heavily on real-time web searches before answering, others rely more on their internal knowledge, supplemented occasionally with content retrieved at the moment of the question. These differences directly affect what it takes to get cited, and are covered in the lesson [ChatGPT, Perplexity, Gemini, Claude: how each one chooses its sources](/en/academy/comprendre-le-geo/comparer-chatgpt-perplexity-gemini-claude). This conversational dimension also has a direct consequence on the type of queries handled: questions there tend to be longer, more precise, and closer to a natural conversation than the isolated keywords typed into a classic search engine. Content designed only to answer short keywords is therefore less prepared for this mode of questioning than for traditional search.

Another notable shift is the conversational nature of these tools: a user asks an initial question, then refines the request through follow-up questions within a single thread. This changes how a search intent is built, often over several steps rather than a single well-formed query.

## Where user attention is shifting

The change in habits is gradual rather than sudden. Classic search remains dominant in total query volume, but the share of answers obtained without clicking through to a third-party site — whether via an AI Overview or a standalone chatbot — is steadily growing. Certain categories of questions are more affected than others: product comparisons, recommendations, and complex questions that require pulling together several sources to answer properly. For these kinds of queries, a growing number of users turn directly to a chatbot rather than running a classic search followed by several clicks and open tabs. This isn't a wholesale replacement of one channel by another: classic search, AI-generated summaries embedded in search, and standalone chatbots coexist, and the same user may move between them depending on the type of need. Take a concrete example: a buyer comparing accounting software might today type a classic query into Google, read an AI Overview summarizing several reviews, then open Perplexity to ask a precise follow-up question about a tool's tax compliance. Three different channels, three different selection logics, for a single buying journey.

## What this shift means for your visibility

For a brand or a website, this means you now need to think about your presence across several channels at once: classic organic results, generative summaries embedded in search, and answers produced by standalone AI chatbots. The content-selection mechanisms are not the same from one channel to the next, which explains why a strong SEO position guarantees nothing about being cited in an AI answer, and vice versa. The lesson [SEO vs. GEO: the key differences](/en/academy/comprendre-le-geo/seo-vs-geo-les-differences) covers these differences in detail, while [Plain LLM vs. retrieval-augmented LLM (RAG)](/en/academy/comprendre-le-geo/llm-classique-vs-llm-rag) explains the technical mechanism that lets a chatbot fetch up-to-date information before formulating its answer, rather than relying solely on what it learned during training.

The key takeaway: the search landscape in 2026 is no longer a single line but a network of channels. Each has its own selection rules, and a complete visibility strategy now has to account for all of them, rather than betting on a single entry point.

### [Plain LLM vs. retrieval-augmented LLM (RAG)](https://vurto.ai/en/academy/comprendre-le-geo/llm-classique-vs-llm-rag) (2026-06-10)

A plain LLM generates its answer solely from what it memorized during training, frozen at a given cutoff date. An LLM with search, known as "RAG" (retrieval-augmented generation), first looks up relevant recent documents, then relies on those documents to build its answer. This distinction is central to understanding why your content can, or cannot, influence what an AI says about your brand.

## What a plain LLM can and can't do

A large language model is trained on a huge volume of text, up to a cutoff date. Once that training is complete, its "knowledge" is frozen in the model's weights: it doesn't read anything in real time, it generates text by relying on statistical regularities it learned. Asked about a stable, well-documented topic from the time of its training, a plain LLM can answer accurately. But asked about recent news, a product released after its cutoff date, or a brand that barely appeared in its training data, it simply doesn't have the information — and risks "hallucinating" a plausible-sounding but false answer. A concrete example illustrates this limitation well: if you launch a new brand or a new product after a model's cutoff date, that model simply has no internal knowledge of it. Asked directly about it, it may either honestly admit as much, or, in the worst case, invent a plausible-sounding but entirely false description based on similar things it learned elsewhere.

## The principle of RAG: search before answering

RAG adds an information-retrieval step before the answer is generated. Concretely, in a typical RAG architecture, the system turns the user's question into one or more queries, retrieves relevant documents (via a web search engine, a vector database, or both), then feeds those documents to the language model as extra context for writing its answer. The model no longer just recites what it memorized: it reads excerpts retrieved at the very moment of the question and uses them as raw material. This is the mechanism that lets a chatbot cite a source, a recent figure, or an article published the day before. This retrieval step isn't free in terms of computation or response time: it's a deliberate trade-off made by AI providers between the speed of a purely generated answer and the reliability of an answer checked against recent sources. That's also why some tools only trigger a search when the question clearly calls for it, such as current events or a specific topic, rather than systematically for every exchange.

## Why this distinction changes everything for your visibility

If an AI operates in purely classic mode, without search, your only chance of being mentioned is having been sufficiently present in the model's training data — largely outside your control in the present moment. If, on the other hand, the AI operates in RAG mode, content you publish today can be retrieved, read, and cited tomorrow, provided it's accessible to crawlers and well structured. It's this second situation that opens up real room for action in GEO: producing content that's retrievable, understandable in isolated excerpts, and well-enough substantiated to be judged trustworthy at selection time. The lesson [How a RAG-based LLM builds an answer](/en/academy/comprendre-le-geo/comment-un-llm-rag-fabrique-une-reponse) breaks down the concrete steps of this process, from the initial query to the final citation. This guarantees nothing automatically: being retrievable doesn't mean being kept, but it's an indispensable precondition, without which none of the following steps even get a chance to happen.

## Not every AI tool is RAG in the same way

In practice, most consumer chatbots combine both approaches: they rely on internal knowledge for stable topics, and trigger a web search when the question concerns current events, a specific topic, or when the user explicitly asks for it. How often and how each tool triggers this search varies noticeably from product to product: some search almost systematically, others only in certain cases. These nuances are explored in the lesson [ChatGPT, Perplexity, Gemini, Claude: how each one chooses its sources](/en/academy/comprendre-le-geo/comparer-chatgpt-perplexity-gemini-claude), which compares how the main tools on the market operate. In some tools, users can even manually turn this search mode on or off, by explicitly selecting an option like "search the web," which shows that plain and RAG aren't two separate categories of tools, but often two modes of operation within the same product.

## What this means concretely for your content

In a world dominated by RAG, the question to ask is no longer just "am I well ranked in search results" but "can my content be found, understood, and judged trustworthy within seconds by a system that only reads an isolated excerpt of my page." That requires content accessible to crawlers, self-contained passages that don't need the rest of the article to make sense, and concrete evidence that reassures the system about the reliability of the information. These requirements are covered in more depth in the lessons on chunking and on structuring content for GEO.

The essential takeaway: a plain LLM recites, a RAG-based LLM searches then synthesizes. None of this is about gaming a system: it's simply about recognizing which mode an AI is operating in before deciding where to invest your time. It's this second logic that makes GEO possible, because it leaves room, at every new question a user asks, for content you publish today. In practice, this means checking, tool by tool, whether an answer engine is drawing on fresh retrieval or on frozen training data before drawing conclusions about why your brand was or wasn't mentioned.

### [How a RAG-based LLM builds an answer](https://vurto.ai/en/academy/comprendre-le-geo/comment-un-llm-rag-fabrique-une-reponse) (2026-06-11)

When you ask a chatbot like ChatGPT or Perplexity a question and it answers while citing sources, that's not a simple copy-paste. The system follows a chain of steps: reformulate the question, retrieve documents, select the most relevant passages, synthesize them into a coherent answer, then cite its sources. Understanding this chain lets you know precisely where and how your content can fit into it. It's worth keeping in mind that this chain isn't perfectly linear or identical from one tool to another: some systems merge several of these steps, others add extra ones, but the general logic, search then synthesize, remains common to the vast majority of current RAG architectures.

## Step 1: the query gets rewritten before it's sent out

Your question, exactly as you typed it, is almost never used as-is to search for information. The system generally rewrites it into one or more more precise search queries, sometimes breaking it down into several distinct sub-questions. This is the fan-out query mechanism, covered in the lesson [Fan-out queries: the key concept](/en/academy/comprendre-le-geo/les-fan-out-queries): a single question from the user can trigger several parallel searches, each covering a different angle of the topic. The number of sub-queries generated varies with how complex the question is perceived to be: sometimes two or three reformulations are enough, sometimes the breakdown is much broader. This number isn't fixed and depends on the tool and the topic at hand.

## Step 2: retrieving relevant documents

Once the queries are formulated, the system looks for documents likely to answer them: results from a web search engine, content from an internal database, or a combination of both depending on the tool. This step generally produces a broader set of candidates than what will ultimately be used, somewhat like a first list of results before filtering. In practice, this step often relies on a combination of keyword search and semantic similarity search, the latter making it possible to find relevant passages even when they don't use exactly the same words as the question asked.

## Step 3: selecting and ranking passages (reranking)

Among the retrieved documents, not all are equal. A ranking mechanism, often called reranking, evaluates how relevant each passage is to the question asked, and only keeps the excerpts judged most useful. This filtering generally happens at the passage level, not the whole-document level: a page can contain one highly relevant paragraph in the middle of otherwise unhelpful content, and only that paragraph will be kept. This mechanism is explained in more detail in the lesson [How the LLM merges results (reranking)](/en/academy/comprendre-le-geo/comment-le-llm-fusionne-les-resultats).

## Step 4: synthesis into a coherent answer

The language model receives the selected passages as context, alongside the initial question, and writes an answer that synthesizes them. This isn't a matter of stitching quotes together: the model rephrases, summarizes, sometimes compares several sources that contradict each other, and produces a coherent piece of natural-language text. This is the step where the writing quality of the source content matters: a clear, self-contained passage is easier to integrate correctly than an ambiguous paragraph that depends on surrounding context. When several selected sources contradict each other, for instance on a price or a technical specification, the model also has to arbitrate: it may present both versions, favor the source it judges most recent or reliable, or sometimes stay vague rather than pick a side. This is one reason why the freshness and consistency of content matters as much as simply being available online.

## Step 5: citation, when it exists

Depending on the tool, the final answer may or may not include an explicit citation to the source used: a link, a domain name, a footnote. This step isn't systematic or uniform across tools, and it also depends on the nature of the question. One thing holds true across most architectures: to be cited, you first have to survive steps 2 and 3 — that is, be retrieved and then judged sufficiently relevant. It isn't unusual for content to shape the wording of an answer without ever being explicitly cited, which makes measuring its real impact more complex than simply counting cited links.

## What this chain of steps means for your content

Each step in this chain is a filter. Your content first needs to be accessible and indexable to survive step 2. It then needs self-contained, clear passages to survive step 3. Finally, it needs to be written so that an automatic synthesis renders it faithfully at step 4. It's this whole set of constraints that defines what's called "GEO-ready" content, a notion developed in the lesson [What makes content "GEO-ready"?](/en/academy/comprendre-le-geo/contenu-geo-ready). A concrete example sums up the stakes well: two articles can cover exactly the same topic with the same level of accuracy, and yet one will get cited three times more often than the other simply because its structure makes each of these steps easier, while the other's makes them harder.

Keeping this chain in mind changes how you write: rather than optimizing for a single ranking algorithm, the goal is to produce content that stays useful and understandable at every step an AI puts it through before answering someone else. None of these steps is optional from the content's point of view, and skipping the analysis of even one of them tends to produce blind spots in how you evaluate why a piece of content did, or didn't, get used. Thinking through each step separately, rather than treating the whole process as one opaque black box, makes it far easier to pinpoint where a specific piece of content is falling short.

### [Fan-out queries: the key concept](https://vurto.ai/en/academy/comprendre-le-geo/les-fan-out-queries) (2026-06-12)

When a user asks a generative AI a question, that single question frequently gives rise to several distinct searches, launched in parallel, each covering a different angle. This is what's called the fan-out query mechanism: a query that "fans out" into several sub-queries before the system synthesizes a single answer. It's probably the most useful concept to understand for anyone thinking about content for GEO, because it shifts the question from "what's THE right keyword" to "what are ALL the questions related to this topic." This shift in perspective often requires more upfront organizational effort: mapping out the likely questions around a topic before writing, rather than starting from a single keyword and unspooling content from there.

## The principle: one question, several searches

Take a simple example: a user asks "what's the best invoicing software for a small business in France." Rather than running a single search with those exact words, a RAG system may break the question down into several distinct sub-queries: "invoicing software for small businesses in France," "invoicing software comparison for SMBs," "invoicing software compliant with anti-VAT-fraud law," "reviews of invoicing software for small business." Each of these sub-queries is processed separately, produces its own batch of results, before everything is merged to build the final answer. The exact number of sub-queries generated isn't fixed: it depends on how complex the question is perceived to be and on the tool used, ranging from as few as two or three reformulations to a much broader breakdown for a topic perceived as complex.

## Why AI systems work this way

This decomposition allows several facets of an intent that's often implicit in a natural-language question to be covered. A single literal query would poorly capture the diversity of real needs behind the question: price, legal compliance, ease of use, other users' reviews. By multiplying search angles, the system increases its chances of retrieving documents that, put together, genuinely cover the topic rather than just one aspect of it. Take another example: a question like "how can I reduce my social contributions as a freelancer" can break down into sub-queries about the most suitable legal status, existing exemption schemes, common mistakes to avoid, or other freelancers' experiences. A site that only addresses one of these angles will only capture a fraction of the fan-out generated by that single question. This is consistent with how a RAG-based LLM builds its answer in multiple steps, as described in the lesson [How a RAG-based LLM builds an answer](/en/academy/comprendre-le-geo/comment-un-llm-rag-fabrique-une-reponse).

## What this means for how you cover your content

The direct consequence of this mechanism is that a single article, however well written, has little chance of answering all the sub-queries generated by a complex question. A site that only addresses a topic from one angle — say, only price, without ever covering legal compliance or user reviews — will only be retrieved for a fraction of the fan-out's sub-queries, and so will be mentioned partially, or not at all, in the final synthesis. Conversely, a site that covers a topic from several complementary angles, across multiple well-structured pages or sections, statistically has a better chance of appearing in several of the generated sub-queries, and so of being more strongly represented in the final answer. This doesn't mean everything needs to be crammed into one overloaded article: it's often better to have several solid pages, each focused on a precise angle, rather than a single page that tries to cover everything superficially and loses clarity on every point.

## Thinking in "topical coverage" rather than a single keyword

This logic brings GEO closer to a discipline of topical coverage than precise keyword targeting. It's no longer about aiming for an exact query, but about covering the full range of questions a user might implicitly be asking around a topic: price, alternatives, use cases, limitations, reviews, comparisons. This approach ties directly into the logic of the question-and-answer format, covered in the lesson [The question-and-answer (FAQ) format and GEO](/en/academy/comprendre-le-geo/format-question-reponse-faq), which is precisely what lets you anticipate several of these potential sub-queries within a single well-structured piece of content.

## A mechanism that also explains the weight of brand presence

Fan-out also explains why brands already mentioned in multiple places across the web have a structural advantage: the more a brand is associated with related topics across different sources, the more likely it is to surface in several of the sub-queries generated by a question, even one phrased unexpectedly. This link between a brand's repeated presence across the web and its likelihood of being cited is developed in the lesson [Brand mentions: the number-one factor in AI visibility](/en/academy/comprendre-le-geo/brand-mentions-facteur-numero-un). In other words, fan-out doesn't just play out at the level of a single page, but also at the level of all the signals available about a brand across the web, which directly ties this mechanism to a broader visibility strategy than content production alone.

The essential takeaway: a question asked to an AI is almost never treated as a single query. It breaks down into several parallel searches, and your content needs to cover enough angles of a given topic to appear in several of those sub-queries, rather than betting on a single phrasing. Practically, this means auditing your existing content less by keyword ranking and more by the range of angles it actually addresses on a given topic. Over time, this kind of mapping tends to reveal gaps in a site's content that keyword-based planning alone would never have surfaced.

### [How the LLM merges results (reranking)](https://vurto.ai/en/academy/comprendre-le-geo/comment-le-llm-fusionne-les-resultats) (2026-06-13)

After the initial retrieval, an AI system typically ends up with far more candidate passages than it can use in an answer. A sorting mechanism, called reranking, evaluates and prioritizes these passages to keep only the most relevant ones. This discreet but decisive step largely determines who gets cited and who doesn't. Unlike an SEO ranking algorithm, which stays broadly stable over time for a given query, the outcome of reranking can vary from one question to another even on a related topic, which makes it even trickier to observe.

## Why sorting is needed after retrieval

The first document-retrieval step, whether through a classic search engine or a vector database, generally produces a fairly broad list of potentially relevant results, without much fine-grained precision. This first list is deliberately wide so as not to miss useful content, but it also contains noise: passages barely related to the topic, redundant with one another, or of uneven quality. Reranking exists to refine this list, evaluating each passage more precisely against the specific question the user asked. This step remains largely invisible to a website: unlike a position in Google's results, there's no "page 2" of reranking you can check to understand why a passage got dropped. That's what makes the topic hard to diagnose without a dedicated measurement tool.

## The general principle behind reranking

In a typical RAG architecture, a specialized model, often smaller and faster than a full LLM, re-evaluates each candidate passage by comparing it directly to the question asked, rather than through simple keyword matching. This model assigns a relevance score to each passage, and only the top-scoring passages are passed on to the language model for drafting the final answer. Some systems also combine several sources of results, for instance keyword search and semantic similarity search, then merge the two rankings into one; researchers often describe this kind of merging as "reciprocal rank fusion," a method that favors passages ranked well across several lists at once rather than those that dominate a single list. This choice illustrates a fairly general principle of reranking: a passage's robustness, meaning its ability to stay relevant across several different evaluation methods, often matters more than a perfect but isolated match on a single criterion.

## What makes a passage more likely to rank well

Without claiming to know the exact, weighted criteria of each proprietary system, some general characteristics can be observed that tend to favor a passage during reranking: a clear, direct match with the intent of the question, phrasing that answers head-on rather than circling the topic, and enough standalone clarity to be understood without having read the rest of the page. Conversely, a passage buried in generic content, one that requires outside context to make sense, or one that vaguely restates a question without precisely answering it, statistically has less chance of being kept at this stage. Take an example: for the question "what's the legal withdrawal period for an online purchase," a passage that opens directly with "the legal withdrawal period is 14 days from receipt of the product" has a better chance of ranking well than a paragraph that starts with "the question of deadlines is complex and depends on several factors" before arriving, a few sentences later, at the same information. The gap between the two isn't about accuracy, since both can be equally correct, but about how quickly the useful information surfaces for a system that has no patience for a slow build-up.

## The link with chunking and content structure

Reranking operates on passages, not whole pages. That's why how your content is split up beforehand, a process called chunking, has a direct influence on what makes it to this sorting step. An article where each section forms a self-contained unit of meaning gives the system more "good candidates" to evaluate, rather than one long block of text where the useful information gets diluted. This splitting mechanism is explained in detail in the lesson [Chunking: how the LLM splits up your content](/en/academy/comprendre-le-geo/le-chunking-comment-le-llm-decoupe-le-contenu). In other words, quality content that stays poorly segmented can end up penalized not on substance, but on form: reranking simply doesn't have the right units to evaluate.

## What this means for how you write your content

In practice, this encourages writing that gets straight to the point: answering a question within the first sentences of a section, clearly structuring each subsection with an explicit subheading, avoiding paragraphs that require having read the whole article to make sense. These structuring principles are detailed in the lesson [Structuring an article for GEO](/en/academy/comprendre-le-geo/structurer-un-article-pour-le-geo), which offers a template designed to maximize the chances that as many of your passages as possible survive the reranking step.

Keep in mind that reranking is an invisible but decisive filter: your content can be perfectly retrieved at the first step and still never appear in the final answer, simply because it wasn't judged relevant, clear, or self-contained enough at sorting time. This observation makes the case for regularly reviewing your own content, section by section, honestly asking whether each passage, taken in isolation, would give a clear answer to someone who had read nothing else. None of this requires guessing blindly: publishing content, checking whether it gets cited for related questions over time, and adjusting its structure based on what does and doesn't get picked up is a workable, iterative way to learn what a given reranking system tends to favor.

### [Chunking: how the LLM splits up your content](https://vurto.ai/en/academy/comprendre-le-geo/le-chunking-comment-le-llm-decoupe-le-contenu) (2026-06-14)

Before a RAG system can search for and retrieve your content, that content has to be split into smaller pieces, called "chunks." This splitting process, chunking, determines what an AI actually sees from your page when answering a question: rarely the whole article, almost always one or a few isolated excerpts. This reality often surprises people used to thinking in terms of a "page" rather than a "passage," and yet it's the basic unit most current RAG systems operate on. Understanding this mechanism fundamentally changes how you should write for GEO. This isn't a peripheral technical detail: it's literally the unit of work from which a generative AI builds its perception of your content.

## Why content gets split up rather than processed whole

An LLM can't reread the entire web for every question asked, and even in a RAG architecture, it's neither efficient nor necessary to process each page as a whole. Systems therefore index content ahead of time as smaller blocks, often a few hundred words each, stored and made quickly retrievable. When a question comes in, the system isn't asking "which page matches best" but "which chunks, taken individually, match best." This is a fundamental difference from classic search ranking logic, which generally evaluates a whole page. The size of these chunks varies across systems: some splits stay fairly short, from a few sentences to a paragraph, others group several related paragraphs together. Either way, the goal stays the same: produce units small enough to be precise, and large enough to remain understandable.

## How this splitting is generally done

In a typical RAG architecture, splitting often relies on the document's existing structure: headings and subheadings (Hn tags), paragraphs, sometimes a maximum length set in advance. Well-structured content, with clear subheadings that announce what each section covers, makes for more coherent splitting, where each chunk corresponds to a complete unit of meaning. The lesson [Hn tags (H1, H2, H3)](/en/academy/bases-seo-techniques/les-balises-hn) explains how this heading hierarchy, already useful for classic SEO, also becomes a landmark for automatic content splitting by AI. A concrete example: an article that alternates long, subheading-free paragraphs running several hundred words offers far fewer landmarks for clean splitting than an article where each H2 subheading introduces a precise idea across three or four short paragraphs.

## The problem of passages that depend on context

The main risk with chunking is that an excerpt pulled out of its context can lose all its meaning. A paragraph that starts with "however, this method has one significant limitation" is incomprehensible if the chunk doesn't include the explanation of "this method" mentioned just before, in another block. Likewise, a paragraph that uses "it," "this," or "that product" without ever reintroducing the subject risks being misread, or simply ignored, once isolated from its surroundings. Here's a simple illustration of the problem. A bad, context-dependent passage: "This approach is more expensive, however. It remains recommended in most cases nonetheless." A good, self-contained passage: "Migrating to a new invoicing tool is generally more expensive in the short term than sticking with the existing setup, but it remains recommended for most small businesses because of the time savings it delivers over the long run." The second reintroduces the subject, "migrating to a new invoicing tool," whereas the first assumes the reader already knows what "this approach" refers to. Content written for a human reader who reads the article in order, start to finish, isn't always suited to a system that will only ever see an isolated fragment.

## Writing sections that stand on their own

The practical consequence is that you need to write each section, each subheading, as though it could be read independently of the rest of the article. That means: reintroducing the subject you're discussing even if it creates slight redundancy with the overall title, avoiding ambiguous pronouns that refer back to a distant sentence, and making sure each section clearly answers the question posed by its own subheading rather than continuing an idea started earlier. This requirement ties directly into the principles detailed in the lesson [Structuring an article for GEO](/en/academy/comprendre-le-geo/structurer-un-article-pour-le-geo), which offers an article structure designed for this kind of fragmented reading.

## Chunking and structured data: two complementary mechanisms

Chunking focuses on visible text, but there's a complementary mechanism to help systems understand the meaning of your pages: Schema.org structured data, which explicitly describes the nature of a piece of content (an article, a question-and-answer, a product) in a machine-readable format. The lesson [Schema.org structured data: understanding it](/en/academy/bases-seo-techniques/schema-org-comprendre) explains how this complementary markup helps automated systems better interpret content, even once it's been split into fragments.

The essential takeaway: your content will almost never be read in full by an AI when it's answering a question. It will be seen in fragments. Writing self-contained sections that can stand on their own is therefore one of the most concrete skills to develop for GEO. Unlike a stylistic exercise, this isn't a purely technical skill either: it's above all a writing discipline, similar to what's asked of a journalist writing an article where each paragraph could function as an independent brief. Reviewing a draft with this lens in mind, section by section, is a fast way to spot the paragraphs most likely to be misread once separated from their neighbors. Over time, applying this discipline consistently tends to improve overall clarity even for readers who never interact with an AI system at all.

### [Why AIs sometimes ignore your site](https://vurto.ai/en/academy/comprendre-le-geo/pourquoi-les-ia-ignorent-votre-site) (2026-06-15)

A site can be perfectly well-ranked on Google and yet be completely absent from answers generated by ChatGPT, Perplexity, or Gemini. This is almost never random: there's a limited set of recurring causes, often compounding one another, that explain why an AI ignores a site that's otherwise legitimate on its topic. The good news is that these causes are, for the most part, identifiable and fixable, once you know precisely where to look rather than simply noting the absence of results.

## The crawler simply doesn't have access to the content

The most basic, and yet common, cause is a technical block. Some sites explicitly block, via their robots.txt file, the crawlers used by generative AIs to index content, sometimes out of unfamiliarity, sometimes as a deliberate choice. Others make their content inaccessible unintentionally, for instance by rendering it entirely client-side in JavaScript with no version accessible to bots that don't execute it. In both cases, the result is the same: the content is never retrieved at the search step, and so can never appear in an answer. The lesson [The robots.txt file and controlling AI crawlers](/en/academy/setup-technique/robots-txt-et-crawlers-ia) explains how to check and adjust these settings. A slow site, with high load times, or one whose crawl budget is already largely consumed by low-value pages, may also see its most relevant content crawled less often, or later, than a technically leaner site. These technical performance factors, often overlooked once content is judged satisfactory on substance, deserve regular checking just as much as editorial quality does.

## Content too generic to be selected

Even when accessible, content that just rephrases what already exists elsewhere, with no distinct angle, no hard data, no concrete example, has little chance of being kept during passage sorting. Faced with several sources saying roughly the same thing in a similar way, a reranking system generally favors the one that provides more precise or more directly useful information for the question asked. Content that stays on the surface, describing a topic without ever answering it in an operational way, often gets overtaken by more concrete sources. Duplicate, or near-duplicate, content across several pages of the same site poses a similar problem: when several passages say roughly the same thing in an almost identical way, a sorting system has no reason to favor one over the other, and may simply discard both in favor of a more distinct third-party source.

## Lack of evidence and trust signals

Generative AI systems, like human readers, need reasons to trust a piece of information before repeating it. Content with no identified author, no date, no verifiable data, no links to external sources, sends weaker trust signals than content that clearly states who wrote it, with what expertise, and on what basis. This isn't a matter of content length: a short but precise article, with an identifiable author and a verifiable source, often inspires more trust than a long but anonymous article that never takes a real position. This bundle of signals is grouped under the acronym E-E-A-T (experience, expertise, authoritativeness, trustworthiness), covered in the lesson [E-E-A-T: experience, expertise, authority, trust](/en/academy/comprendre-le-geo/eeat-expertise-autorite-confiance). The importance of concrete evidence and sourced figures is also covered in the lesson [Citations, figures, and evidence: why they matter](/en/academy/comprendre-le-geo/citations-chiffres-et-preuves).

## A structure that prevents good fragment-level reading

As seen in the lesson on chunking, an AI rarely reads a whole page: it retrieves fragments of it. Content written as one long continuous text, with no clear subheadings, no self-contained paragraphs, is harder to split correctly, and some of its most useful passages may simply never get isolated or kept. Well-structured content, with subheadings that explicitly pose a question followed by a direct answer, is mechanically easier for these systems to process. The same goes for content locked away in formats that are poorly machine-readable, such as essential information presented only in an image, a video, or a scanned PDF with no extractable text: the content can be visible to a human and yet completely invisible to a system that only reads text.

## A brand barely mentioned elsewhere on the web

Finally, a less visible but equally decisive cause: your brand's presence outside your own site. A generative AI doesn't rely solely on what you say about yourself; it also takes into account, directly or indirectly, how your brand is discussed elsewhere — on third-party sites, comparison articles, reviews. A brand that's rarely or never mentioned outside its own site is structurally less likely to be recognized as a relevant answer. It's to pinpoint exactly this kind of gap, and know on which engines and which topics a brand is being ignored, that tools like Vurto measure citation rates per AI engine rather than leaving it to guesswork.

Keep in mind that invisibility in AI answers is almost never a uniform or mysterious phenomenon: it generally breaks down into one or more specific blockers, technical or editorial, that can be identified and fixed one at a time. Addressing these causes one by one, rather than searching for a single sweeping explanation, remains the most effective way to restore lost visibility. Working through this list methodically, rather than assuming a single root cause, is usually the fastest way to move from invisibility to a first citation. A quarterly check across these five areas is usually enough to keep a growing site from quietly drifting back into invisibility. Fixing even one of these blockers can be enough to unlock a first meaningful round of citations.

### [SEO vs. GEO: the key differences](https://vurto.ai/en/academy/comprendre-le-geo/seo-vs-geo-les-differences) (2026-06-16)

SEO tries to get a link to appear as high as possible in a list of results. GEO tries to get a piece of content selected as the source of a generated answer. These two goals look similar on the surface, but rest on different mechanisms, with different success metrics. Understanding precisely where they overlap and where they diverge helps you avoid misdirecting your efforts. This distinction isn't just a theoretical exercise: it has direct consequences for how you prioritize the time spent producing, revising, and structuring content.

## Two different units of selection: the link vs. the passage

Classic search ranking evaluates and ranks whole pages against each other, in an ordered list of results. The user then decides which link to click. GEO, on the other hand, operates at the passage level: a RAG system retrieves and evaluates fragments of content, not whole pages, and it's the system, not the user, that decides which passages get incorporated into the final answer. A page can therefore be invisible in Google's top ten and still provide the passage cited in a ChatGPT answer, or the reverse. This difference in unit also explains why the same page can have an excellent SEO score thanks to generally solid content, while containing, at its core, no passage precise or self-contained enough to be picked out on its own by a reranking system.

## Two different success metrics: the click vs. the citation

In SEO, the reference metric remains the click: a user sees a link, decides to click, lands on the site. The whole classic measurement chain, from click-through rate to conversion rate, rests on that actual visit. In GEO, the metric changes in nature: what matters is being cited or mentioned in an answer, whether or not the user then clicks a link. Many AI answers generate no click at all to the source site, which doesn't mean they have no value: being named as a reference in an answer affects brand perception, even without an immediate visit. It's to capture this new form of visibility that tools like Vurto measure citation scores per AI engine, alongside classic traffic metrics.

## What remains common to both disciplines

Despite these differences, SEO and GEO share a foundation of fundamentals that would be a mistake to pit against each other. Clear, well-structured content that genuinely answers a question and rests on verifiable evidence remains an advantage under both logics. A site that's technically accessible to crawlers, fast, and properly marked up is also a necessary condition in both cases: without technical accessibility, neither SEO nor GEO can work. Likewise, how fresh a piece of content is, how regularly it's updated, and how consistent the information it presents is from page to page remain useful criteria under both logics, even if their relative weight isn't necessarily identical. This common foundation is covered in the lesson [What makes content "GEO-ready"?](/en/academy/comprendre-le-geo/contenu-geo-ready), which shows that content designed for GEO generally remains good SEO content too, though the reverse isn't always true.

## Where the two logics diverge the most

The clearest divergence concerns the unit of optimization: classic SEO still largely thinks in terms of keywords and target pages, whereas GEO thinks in terms of topical coverage and self-contained passages, a principle covered in the lesson [Fan-out queries: the key concept](/en/academy/comprendre-le-geo/les-fan-out-queries). Another important divergence: the weight of a brand's external reputation. In SEO, inbound links (backlinks) play a fairly well-documented authority role. In GEO, what seems to matter more is how often a brand is mentioned, cited, or recommended elsewhere across the web, even without a hyperlink — a factor covered in the lesson [Brand mentions: the number-one factor in AI visibility](/en/academy/comprendre-le-geo/brand-mentions-facteur-numero-un). A third, more subtle divergence concerns the stability of results: an SEO position generally shifts gradually from one week to the next, whereas a GEO citation can appear or disappear between two answers to a nearly identical question, which makes tracking AI visibility more demanding in terms of measurement frequency than classic SEO tracking.

## A question of priority, not exclusivity

It would be a mistake to treat GEO as a replacement for SEO, or vice versa. The two channels coexist, as seen in the lesson [The search landscape in 2026](/en/academy/comprendre-le-geo/paysage-de-la-recherche-en-2026), and many efforts benefit both at once. The real question isn't "SEO or GEO," but "how much of my attention should go to each channel," based on how your users actually search for information today. For a small team with limited resources, that can mean prioritizing the fundamentals shared by both disciplines first, such as clarity and technical accessibility, before investing in optimizations more specific to one channel or the other. Conversely, a team that already has substantial editorial resources can afford to handle both channels in parallel, simply adapting its production process rather than duplicating it.

Keep in mind that SEO and GEO aren't two competing disciplines but two selection logics that coexist, with partly different units, metrics, and levers, resting on a largely shared foundation of content quality. In both cases, neglecting this shared foundation in favor of optimizations too specific to a single channel remains the most common trap. Treating the two as complementary lenses on the same underlying content, rather than as rival checklists, tends to produce better results than optimizing for either one in isolation. Framing the two this way also makes it easier to explain the value of GEO work to stakeholders used to thinking purely in SEO terms.

### [What makes content "GEO-ready"?](https://vurto.ai/en/academy/comprendre-le-geo/contenu-geo-ready) (2026-06-17)

"GEO-ready" content is content designed to be easily retrieved, understood, and cited by a generative AI system, without hurting the reading experience for a human. It's not some hidden technique, but a set of concrete editorial choices that directly respond to the mechanisms covered in earlier lessons of this module: retrieval, splitting, sorting, synthesis. None of these choices stands alone: it's their combination, applied consistently across an entire article, that distinguishes genuinely GEO-ready content from content that applies one or two good practices in isolation without reaping the compounded benefit. Take an example of this transformation. A classic opening line, more oriented toward SEO or storytelling: "Choosing the right project management tool can prove complex given how vast the market is today." A GEO-ready opening line for the same article: "The best project management tool mostly depends on your team's size and the complexity of your projects: small teams generally do better with a simple tool, larger teams with a more feature-complete one for reporting." The second version already answers, in one sentence, part of the question.

## A direct answer within the first sentences

The most recognizable feature of GEO-ready content is that it answers the question posed by its title or subheading within the very first sentences, with no introductory detour. This habit, which might seem counterintuitive to anyone taught to "build suspense" in a piece of writing, is exactly what a reranking system needs to quickly judge how relevant a passage is. A busy human reader benefits too: they get the essentials before deciding whether to read further. This principle applies just as much to a blog article as to a product page or an FAQ page, regardless of the format or length of the final content.

## A clear structure, with subheadings that pose a question

GEO-ready content is organized into clearly delimited sections with explicit subheadings, ideally phrased as questions or clear statements rather than vague titles. This structure makes it easier for content to be automatically split into coherent chunks, as seen in the lesson [Chunking: how the LLM splits up your content](/en/academy/comprendre-le-geo/le-chunking-comment-le-llm-decoupe-le-contenu), and helps a system quickly associate a section with the question it addresses. A vague subheading like "A few things to consider" says far less than a precise one like "What budget should you plan for a project management tool in 2026?", which clearly announces the section's content and makes both human reading and automatic splitting easier.

## Concrete evidence rather than vague claims

Content that asserts without demonstrating is harder to judge trustworthy than content that backs its claims with sourced figures, concrete examples, or verifiable data. This isn't unique to GEO: it's also what distinguishes quality content for a demanding human reader. The lesson [Citations, figures, and evidence: why they matter](/en/academy/comprendre-le-geo/citations-chiffres-et-preuves) explains how to bring in this kind of evidence credibly, without falling into an artificial pile-up of unverified statistics. The goal isn't to pile up figures for the sake of it, but to choose, for each important claim, the most relevant evidence available: a study result, an anonymized customer example, an official data point, or failing that, explicit reasoning that shows how you arrived at that conclusion.

## Self-containment of each section

As seen in the lesson on chunking, a RAG system rarely reads an article in full. GEO-ready content is therefore written so that each section can be understood in isolation, without requiring the rest of the article: the subject is clearly reintroduced in each section, ambiguous pronouns are avoided, and each paragraph carries a complete piece of information rather than depending on reasoning built up earlier in the text.

## Technical accessibility you can't see but that everything depends on

None of these editorial principles matter if the content isn't technically accessible to the crawlers used by generative AIs. GEO-ready content therefore also rests on solid technical foundations: a robots.txt file that doesn't exclude legitimate AI crawlers, clean HTML structure with consistent Hn tags, and possibly Schema.org structured data that makes the nature of the content explicit. A misconfigured robots.txt file alone can undo all the editorial effort described above, since the content won't even be retrieved in the first place. These technical basics are covered in the lesson [Schema.org structured data: understanding it](/en/academy/bases-seo-techniques/schema-org-comprendre).

## A concrete template for applying these principles

Put together, these principles define a fairly precise article template, developed in the lesson [Structuring an article for GEO](/en/academy/comprendre-le-geo/structurer-un-article-pour-le-geo): a short answer up front, development in self-contained thematic sections, possibly a question-and-answer section, and a conclusion that stays actionable even read on its own.

Keep in mind that GEO-ready content isn't content written differently to "trick" an automated system: it's content that's clearer, better structured, and better substantiated, which also happens to be easier for an AI to process. The two goals, human readability and machine readability, converge far more than they conflict. The best test to check whether content is GEO-ready is often the simplest one: have someone unfamiliar with the topic skim it, and see whether they walk away with the essentials after a few seconds of scanning. None of these principles requires specialized tools to get started: a careful, honest rereading of an existing article is often enough to spot where it falls short of this standard. Applied consistently, this standard tends to raise the overall quality bar of a site, well beyond the specific articles it was first applied to.

### [Structuring an article for GEO](https://vurto.ai/en/academy/comprendre-le-geo/structurer-un-article-pour-le-geo) (2026-06-18)

An article structured for GEO generally follows a precise template: a short answer right in the introduction, a body organized into self-contained thematic sections, possibly a question-and-answer section, and a conclusion that stays useful even read in isolation. This template isn't an artificial constraint: it flows directly from how a RAG system retrieves, splits, and selects content, as seen in earlier lessons of this module. Each of its components responds to a specific step in that process, which makes it less a stylistic recipe than a structural answer to an observable technical mechanism. This template also isn't so rigid that it excludes any variation: depending on the length and complexity of the topic, some sections can be merged or split further, as long as the overall spirit — direct answer, then self-contained development — is respected.

## Open with a direct answer, not with context-setting

The first section of a GEO-ready article should answer the question posed by the title in one to three sentences, before any elaboration. This direct answer serves a double purpose: it immediately gives a busy human reader the essentials, and it forms the passage most likely to be retrieved in isolation by a search system, since it concentrates the densest information on the title's exact topic. Avoid long context-setting, opening anecdotes, or historical background before getting to the point. A simple test to check this rule: if you deleted the whole article except its first two or three sentences, would a reader already get a useful answer to the question posed by the title? If not, the opening is worth reworking.

## Break the body into self-contained thematic sections

After the opening, the body of the article is split into several sections, each introduced by an explicit H2 subheading, covering complementary angles of the topic rather than developing a single idea in a linear fashion. Each section should be understandable on its own: the subject is clearly reintroduced there, without depending on what was said in the previous section. This logic of covering multiple angles ties directly into the fan-out query mechanism, covered in the lesson [Fan-out queries: the key concept](/en/academy/comprendre-le-geo/les-fan-out-queries): the more distinct angles of a given topic your article covers, the more likely it is to appear in several of the sub-queries generated by a user's question. An example template for an article on choosing a tool: a section on selection criteria, a section on pricing and pricing models, a section on alternatives and cases where a competing tool fits better, a section on common mistakes to avoid when choosing.

## Include a question-and-answer section when it's relevant

For topics that raise recurring, well-identified questions, a section organized as explicit questions and answers is particularly effective: each question becomes a subheading, followed by a direct, self-contained answer. This format matches almost exactly how users phrase their questions to a chatbot, which makes it easier to match a query to the right passage. The lesson [The question-and-answer (FAQ) format and GEO](/en/academy/comprendre-le-geo/format-question-reponse-faq) explains how to build this section without it sounding artificial or redundant with the rest of the article. To avoid redundancy, this section is best used for questions that complement those already covered in the body of the article, rather than repeating the same points in a slightly different form.

## Back up each section with concrete evidence

A clear structure isn't enough if the content of each section stays vague. Every claim benefits from being backed by a concrete example, a sourced figure, or verifiable data, rather than restated generically. This principle, covered in the lesson [Citations, figures, and evidence: why they matter](/en/academy/comprendre-le-geo/citations-chiffres-et-preuves), strengthens both the credibility perceived by an automated sorting system and the actual value of the content for a human reader.

## End with something actionable, not a summary

The conclusion of a GEO-ready article avoids a plain recap of what's already been said, which adds little value once isolated from the rest. It's better spent on a concrete recommendation, a next step, or a synthesis point not yet stated elsewhere in the article. Like the other sections, it should stay understandable if retrieved on its own, without assuming the reader — human or system — has read everything that came before.

## A template that still serves the human reader

This template is nothing like a technical gimmick disconnected from editorial quality: each of its components — direct answer, self-contained sections, concrete evidence, actionable conclusion — also improves the experience of a busy human reader who scans an article rather than reading it line by line. Structuring for GEO is, first and foremost, structuring for clarity.

Remember this five-part template: short answer, self-contained thematic sections covering multiple angles, optional FAQ, concrete evidence woven in throughout, actionable conclusion. It's the practical synthesis of every mechanism covered in this module. This template isn't reserved for long content or complex topics: even a short article benefits from following this logic, if only for its opening and conclusion. Applied consistently, this template gradually reshapes how an entire site is perceived by generative AI systems, article after article, well beyond just the content initially reworked. Revisiting older content through this template, section by section, is often a faster path to improved visibility than producing new articles from scratch. Treated as a checklist rather than a rigid formula, it can be adapted to most formats without losing its underlying logic. Most teams find it easier to apply this template to one important page first, then extend it gradually across the rest of the site.

### [The Q&A / FAQ format and GEO](https://vurto.ai/en/academy/comprendre-le-geo/format-question-reponse-faq) (2026-06-19)

The question-and-answer format almost mirrors word for word how a user types a query into ChatGPT, Perplexity, or Gemini: a question, followed by a direct answer. That's what makes it especially easy for a generative engine to extract and cite, since it has to summarize or rephrase a source in a matter of seconds. But mechanical or excessive use of the FAQ format can also hurt a page's readability and credibility.

## Why AI engines favor the Q&A format

A language model answering a query doesn't "read" an entire page the way a human does: it often works from fragments of text (passages, paragraphs, sometimes isolated sentences) retrieved because they match the question semantically. A paragraph that opens with a question close to the user's, followed by a self-contained, complete answer, reduces the interpretive work the model has to do. There's less ambiguity about what the passage "answers," and therefore less risk that the AI distorts or truncates the information when citing it. The retrieval systems these engines rely on, often built on semantic similarity between the query and the available passages, structurally favor text segments that already resemble a question-answer pair, because the semantic distance between the query and the passage is mechanically smaller than in ordinary narrative text.

This isn't unique to generative AI: Google has relied for years on question-answer excerpts for its featured snippets. GEO simply extends that same logic to engines that, instead of displaying a link, produce a synthesized sentence with or without attribution.

## Identifying the right questions to address

The most common mistake is inventing generic questions ("What is [product]?", "Why choose [product]?") that don't correspond to any real query. Questions worth addressing come from concrete sources: related searches in search engines, questions asked by prospects to the sales or support team, threads on specialized forums, customer reviews that contain an implicit question.

The lesson on [prompts and AI queries](/en/academy/comprendre-le-geo/prompts-et-requetes-ia) details the main query families (informational, comparative, transactional). The useful takeaway here: a good FAQ question almost always maps to a precise informational or comparative query, not a broad transactional intent.

## Structuring an answer that can be cited as-is

A citable answer starts by answering, not by setting the scene. If the question is "How long does it take to see results with GEO?", the first sentence should offer a direct, even cautious, element of answer ("There is no universal timeframe, but the first effects on visibility in AI answers are generally observed after several weeks of regular publishing"), before nuancing further in the following sentences.

Avoid answers that don't stand on their own once taken out of context: "it depends on several factors," without specifying which ones, helps neither the reader nor the engine trying to cite something concrete. Content that is truly [GEO-ready](/en/academy/comprendre-le-geo/contenu-geo-ready) treats every question as a self-contained unit of meaning. A useful test before publishing: read the answer alone, stripped of its surrounding page, and check whether it still makes sense and still answers the question fully. If it needs the paragraph before or after it to be understood, it isn't yet self-contained enough to be lifted and cited on its own.

## The right length for a Q&A answer

An answer that's too short lacks substance and pushes the engine to prefer a more complete source; an answer that's too long dilutes the useful information in filler the AI has to sift through before it can cite it. In practice, a direct answer often fits in two to four sentences, followed if needed by a longer development for the human reader who wants to dig deeper. This funnel structure — short answer first, context and nuance after — serves both audiences at once: the generative engine can extract the core of the answer without having to sort through several paragraphs, and the reader who stays on the page finds the detail they're looking for right below.

Avoid, conversely, answers that open with a long setup before getting to the point: the engine risks citing that preamble instead of the useful information, or ignoring the passage altogether for lack of a clear answer in the first lines.

## The FAQ format doesn't replace substantive content

Stacking ten generic questions at the bottom of a page in hopes of gaining visibility is a practice Google has long recognized and penalized, explicitly listing "low-value FAQ content" among its examples of spam created to manipulate rankings. Generative engines face the same problem: a block of redundant questions without substantial answers offers nothing worth citing.

The Q&A format works better woven into the natural flow of an article — a subsection answering a precise question within a substantive paragraph — than as a generic accordion tacked on at the end of a page just to "do FAQ." Each question should bring information the rest of the page doesn't already cover.

## FAQ and structured data

On the technical side, Schema.org's FAQPage markup can help traditional search engines identify your questions and answers, as covered in the lesson on [Schema.org structured data](/en/academy/bases-seo-techniques/schema-org-comprendre). For generative AI, this markup isn't the main driver of citation: what matters first is the clarity and self-sufficiency of the text itself. Markup is a complementary signal, not a substitute for a well-written answer.

In short, the question-and-answer format is powerful because it mirrors the very mechanics of AI queries, but its effectiveness depends entirely on the quality and specificity of the answers it contains — not on how many questions are asked.

### [Citations, numbers and evidence: why they matter](https://vurto.ai/en/academy/comprendre-le-geo/citations-chiffres-et-preuves) (2026-06-20)

A generative engine rephrasing an answer tries to limit the risk of saying something false. Content that backs its claims with precise, sourced, verifiable data gives the AI a concrete reason to cite it rather than a page that offers only general assertions.

## Why AI engines look for evidence

Generative language models are partly evaluated on their ability to avoid hallucinations — invented or incorrect information. This pushes systems that rely on web retrieval (Perplexity, ChatGPT with browsing, Gemini, Google's AI Overviews) to favor sources presenting verifiable facts over unsupported opinions. A sentence like "our customers are very satisfied" offers nothing to cite; a sentence reporting a precise, sourced figure, even an approximate one, gives the model something useful to point to.

That doesn't mean every published number gets picked up: the engine also has to judge the reliability of the source, which ties directly back to the [E-E-A-T](/en/academy/comprendre-le-geo/eeat-expertise-autorite-confiance) criteria covered in another lesson. This mechanism also explains why a claim repeated across several independent sources has better odds of being cited than one appearing on a single site only: convergence across several independent sources acts as a form of cross-validation, both for a human reader and for a system trying to judge a piece of information's reliability before repeating it. This doesn't mean copying a competitor's claim just to make it "appear" more often: it means understanding that, if your data point is accurate and well sourced, the best way to reinforce its perceived credibility is to let other serious players in the field cite it and pass it along in turn.

## What data to include, and how to source it

The most useful data for GEO-ready content comes from an identifiable primary source: a study you conducted, aggregated and anonymized internal statistics, data from a public body or a recognized industry report. In every case, the data point should be accompanied by its source (study name, organization, year) directly in the text or in a note, not only in a general bibliography. An internal data point, even a modest one, gains credibility once you specify the sample size and the measurement period: "across 1,200 support tickets analyzed between January and June" is more citable than "our data shows that."

If you cite a statistic produced by a third party, link it to its primary source rather than to an article that repeats it without a reference. Generative engines, like careful readers, are more inclined to trace back to the origin of a data point when it's traceable.

## Making a claim "citable"

A citable claim is specific, self-contained, and attributable. It avoids vague phrasing ("many companies see an improvement") in favor of wording that names the source and the nature of the measurement ("according to [study/organization], companies that adopted this practice report a measurable improvement, though the exact magnitude is not consistently agreed upon"). This caution doesn't weaken the point — it makes it more credible, because it matches what can actually be demonstrated. Systematically stating the period and geographic scope of a data point ("in France, in 2025" rather than a timeless, universal claim) further reduces the risk that it gets repeated out of context or applied to a situation it doesn't actually describe.

## A concrete before / after example

Take a simple example. A sentence like "AI search visibility is growing" offers no verifiable anchor: no engine can cite it as proof of anything specific. A revised version might read: "several publishers and analytics vendors report, over recent quarters, an increase in referral traffic coming from generative AI, though no consensus has yet formed on its exact scale" — the nuance is owned, the source is identifiable even in generic terms, and the claim remains useful.

This kind of rewrite takes more editorial effort than a vague opening line, but that extra effort is exactly what separates content an AI can safely cite from content it has to ignore for lack of a way to verify it. A simple habit to build: whenever a sentence contains a qualifier like "many," "significant," or "sharply rising," ask yourself whether you can replace it with a data point or a source, even a partial one.

## The risk of inventing numbers

Inventing or rounding up a percentage to make it more impressive ("GEO increases traffic by 47%") is a practice to avoid entirely, for two reasons. First, because once picked up by a generative AI and then by other sites citing it in turn, that figure spreads as an established fact despite having no verifiable basis. Second, because engines and readers who trace an invented figure back to its source lose trust in the content as a whole, which damages brand credibility over the long run.

The lesson on [GEO numbers](/en/academy/comprendre-le-geo/chiffres-du-geo-marche-roi) applies this same caution to market data itself: it's better to state a trend with the nuance it deserves than to cite an unverified precise percentage.

## Where to find reliable data for your industry

Public sources (national statistical bodies, sector regulators, openly accessible academic publications) remain the most robust. Reports produced by analyst firms or industry players can be useful, provided you check the methodology used and avoid presenting an estimate as a certain measurement. When data is missing, it's perfectly legitimate to write "there is, to our knowledge, no consolidated public measurement on this point" rather than fill the gap with an estimate presented as fact.

Ultimately, what makes content citable by an AI isn't the quantity of numbers it contains, but the traceability and honesty of each one.

### [E-E-A-T: experience, expertise, authority, trust](https://vurto.ai/en/academy/comprendre-le-geo/eeat-expertise-autorite-confiance) (2026-06-21)

E-E-A-T (Experience, Expertise, Authoritativeness, Trust) is the framework Google formalized in its guidelines to evaluate the quality of content and its author. It wasn't originally designed for generative AI, but the criteria it describes — genuine lived experience, demonstrated skill, recognition by peers, perceived reliability — are precisely what pushes a generative engine to prefer one source over another when several contain similar information.

## Where the E-E-A-T framework comes from

E-E-A-T comes from Google's Search Quality Rater Guidelines, a document meant for human raters who score the quality of search results. It isn't a direct ranking factor in the sense of an algorithm that literally "scores" E-E-A-T, but a set of signals Google tries to approximate through other means (links, mentions, site structure, author information). Generative engines, which also have to judge a source's reliability before citing it, rely on very similar signals, whether or not they explicitly call them E-E-A-T in their own internal methods. The fact that this framework was designed for human raters rather than an automated algorithm is actually an asset for GEO: it describes what a demanding reader would try to verify before trusting a piece of content, which is very close to what a language model has to approximate statistically.

## Experience: the genuine, lived-through content

The first E, added in 2022, distinguishes direct experience from mere theoretical expertise. An article about using a tool, written by someone who actually used it, stands apart from a summary article compiled without hands-on practice. Concretely, this shows up as details that would be impossible to produce without having done the thing: real screenshots, observed results, limitations encountered in practice rather than generalities. A generative AI can't "feel" this lived experience, but it can detect, through the language used and the specificity of the details, that a text stems from experience rather than paraphrase. In terms of writing, this lived experience also shows up in the vocabulary used: a text that recounts a test run over a specific period, mentions a software version tested or a date of experimentation, or describes a difficulty encountered and how it was resolved, sends different signals than a purely descriptive text written in the conditional ("this should work," "it is likely that"). These linguistic markers, taken individually, may seem trivial, but their accumulation helps distinguish content grounded in real practice from content compiled from other sources without direct experimentation.

## Expertise and authority: who is speaking, and with what legitimacy

Expertise concerns the writer's competence on the subject at hand; authority concerns the source's reputation in its field, independent of any single article. An identifiable author name, with a bio detailing their background and area of competence, is a simple, high-value signal to put in place. At the site level, authority is built through external recognition: being cited by other established sources in the field, contributing to third-party publications, being mentioned in independent comparisons or studies. This topic is covered in more detail in the lesson on [off-site signals that matter](/en/academy/strategie-seo-geo/signaux-off-site-qui-comptent).

In sensitive sectors (health, finance, law), often grouped under the label YMYL ("Your Money or Your Life") in Google's guidelines, the bar for demonstrable expertise is even higher: a factual error there carries potentially more serious consequences for the reader, which makes engines, human or generative, especially cautious before citing an unidentified source or one with no apparent qualification.

## Trust: the common denominator

Google presents trust as the most important component of the framework: content can display experience and expertise, yet if it's perceived as unreliable (unverified information, no transparency about who runs the site, internal contradictions), it loses its value. Trust is built through concrete, verifiable elements: clear legal notices, a policy for correcting errors, sourced citations for factual claims, consistency between what's announced and what's demonstrated. This connects directly to the lesson on [citations, numbers and evidence](/en/academy/comprendre-le-geo/citations-chiffres-et-preuves): a sourced claim is one of the most direct ways to build trust. Trust also isn't confined to a single page: a site whose different pages contradict each other on the same topic (different figures for the same metric, inconsistent recommendations from one article to the next) sends a negative signal that can affect perception of the whole domain, not just the offending page. Periodically reviewing all published content for such inconsistencies is part of the editorial hygiene that sustains trust over time.

## Applying E-E-A-T concretely to your content

In practice, improving a piece of content's E-E-A-T comes down to simple, cumulative actions: byline articles with a real, identifiable author, add a detailed "about" page covering the company and its experts, systematically source figures and statistics, date and update content as information evolves, and avoid claims that a third party couldn't verify. None of these elements alone guarantees a citation by a generative AI. But together, they reduce the uncertainty the engine has to manage when deciding whether your page deserves to be the source it chooses to cite over another — a stake directly tied to what sometimes leads an AI to [ignore a site](/en/academy/comprendre-le-geo/pourquoi-les-ia-ignorent-votre-site) that doesn't offer these assurances.

E-E-A-T is therefore not a box to tick, but a coherent set of editorial practices that, taken together, make content more trustworthy — for a human reader as much as for a generative engine that has to decide, in a matter of seconds, whether it can rely on what you wrote. Treat it as an ongoing editorial standard rather than a one-time checklist to complete before publishing.

### [Prompts and AI queries: how users actually ask](https://vurto.ai/en/academy/comprendre-le-geo/prompts-et-requetes-ia) (2026-06-22)

Users don't ask ChatGPT or Perplexity questions the way they used to type keywords into Google. AI queries map more closely to three broad families — informational, comparative, transactional — and each calls for a different kind of content if you want a chance at being cited.

## A conversation rather than a list of keywords

A classic Google search often boils down to two or three words ("invoicing software small business"), because the user knows they'll have to filter the results themselves. Facing a generative AI, the same intent tends to be phrased as a full sentence, with context ("I'm looking for simple invoicing software for a 10-person small business, with automatic accounting export"). This difference has a direct consequence for content: a text that only answers the generic keyword, without addressing contextual variants and precise use cases, captures these longer, more specific queries less effectively.

## Informational queries

These are the most common: the user is trying to understand a topic, a concept, or how something works ("How does two-factor authentication work?", "What is free cash flow?"). Unlike a classic Google search, where the user often had to visit several pages to piece together an answer, generative AI condenses information into a single response and expects a source that explains the topic clearly and completely on its own. Content that answers this type of query well is structured like a full explanation, with a clear definition up front — the approach described in the lesson on [GEO-ready content](/en/academy/comprendre-le-geo/contenu-geo-ready).

## Comparative queries

The user is trying to choose between several options ("What's the difference between X and Y?", "Which tool should I pick for Z?"). These queries are particularly sensitive for a brand, because the AI has to decide which alternatives to mention and in what order. Content that handles a comparison honestly — laying out decision criteria, the use cases where each option makes sense, without excessively bashing competitors — has a better chance of being seen as a reliable source and cited for the comparison as a whole, not just for the mention of your own product. This is also where the way each engine selects its sources diverges the most, a topic explored in the lesson [ChatGPT, Perplexity, Gemini, Claude: how each picks its sources](/en/academy/comprendre-le-geo/comparer-chatgpt-perplexity-gemini-claude).

## Transactional queries

The user is ready to act: sign up, buy, book ("Where to buy this product," "What's the price of this service"). Mainstream generative AI tools today remain more cautious with this type of query than with the previous two, partly because they avoid committing to volatile information like a price or availability. The content most likely to be picked up here presents stable factual information (a public pricing grid, terms and conditions, clear terms) rather than vague sales copy.

## What this typology means for your content

This classification isn't just an academic exercise: it should shape your site's editorial structure. A product page generally answers a transactional intent but often fails to address the informational and comparative intents that precede a purchase decision in the user's actual journey. Producing content dedicated to each query type — an explanatory article, an honest comparison, a clear factual page — multiplies the entry points through which a generative AI might cite your brand, at different moments of the user's reasoning. In practice, the same topic often deserves three distinct pieces of content rather than a single article trying to cover everything superficially: depth on each intent matters more than breadth in one piece. It's worth mapping your existing content against these three families before producing anything new: list your current pages, identify which intent each one primarily answers, then spot the intents left uncovered. It's not uncommon to find an entire site structured around transactional intent (product pages, pricing, contact forms) while most of the AI queries asked earlier in the decision process are informational or comparative. Filling these blind spots, rather than multiplying transactional pages, is often the most cost-effective lever in the short term for widening your presence in AI answers.

## Anticipating rephrasing

The same information need can be expressed in dozens of different ways depending on the user, their familiarity with the topic, or the engine used. Rather than targeting a single phrasing, it's more effective to cover a subject from several logical angles (definition, method, comparison, limits, concrete cases): this increases the odds that your content semantically matches one of the many possible rephrasings of the same intent. It's also what makes content resilient as query habits evolve — and they're evolving faster with generative AI than they ever did with classic search, as shown by the rapid shift in the [research landscape in 2026](/en/academy/comprendre-le-geo/paysage-de-la-recherche-en-2026). Simple tools help identify these rephrasings: related search suggestions, the questions listed in "people also ask" sections of classic search engines, or directly asking the same question to several generative AI tools and comparing the terms they use in their answers. This last method has the advantage of revealing the vocabulary the engines themselves use to synthesize a topic, a useful clue for aligning your own phrasing.

Understanding these three query families doesn't replace directly observing the questions actually being asked about your industry, but it provides a framework for organizing that observation and prioritizing which content to produce first. Revisit the mapping periodically, since the balance between these three intents for a given topic tends to shift as a market matures and buyers become more informed.

### [ChatGPT, Perplexity, Gemini, Claude: how each picks its sources](https://vurto.ai/en/academy/comprendre-le-geo/comparer-chatgpt-perplexity-gemini-claude) (2026-06-23)

The four major mainstream generative engines don't select their sources the same way, because they rely on different architectures and partnerships. These differences are observable in practice, but they keep evolving: what follows describes trends, not fixed rules.

## Why these differences exist

Before going engine by engine, it helps to understand where these gaps come from. Two factors carry particular weight: the share of "internal" knowledge (learned during the model's training) versus real-time web search, and whether or not commercial partnerships exist with content publishers that can influence how often certain sources appear. An engine that relies mostly on internal knowledge reacts slowly to newly published content; an engine that relies mostly on real-time search reacts fast, but then depends heavily on the technical quality of indexing. Another factor worth considering is each model's training knowledge cutoff date: an engine whose training stopped several months before it went into production may miss recent changes in your industry, unless it compensates with active web search. This date varies from one model to the next and shifts with every new version, which makes regular re-checking all the more useful compared to a one-time, frozen assessment.

## Perplexity: the conversational search engine

Perplexity is built around real-time web search: each answer draws on a set of search results retrieved at the moment of the query, then synthesized with numbered citations linking back to the source pages. This architecture makes it especially dependent on the freshness and technical accessibility of content (see the lesson on [llms.txt and the XML sitemap](/en/academy/setup-technique/llms-txt-et-sitemap-xml) for the technical side). In practice, Perplexity tends to cite a wide range of sources, including forums, news sites, and documentation pages, making it one of the engines with the highest number of visible sources per answer.

## ChatGPT (with browsing): between trained knowledge and live search

ChatGPT combines knowledge learned during the model's training with, when web browsing is enabled, results retrieved live. This duality means your brand's presence in ChatGPT answers depends both on how it was mentioned in the content used for training (past mentions, hard to influence retroactively) and on its current visibility on the web when search is active. OpenAI has content partnerships with several publishers, which can influence how often certain sources appear, without this being exhaustively and publicly documented.

## Gemini: integration with the Google ecosystem

Gemini benefits from direct access to Google's search infrastructure, which naturally brings it closer to the logic already at work in classic Google ranking and in AI Overviews (covered in a dedicated lesson on [Google AI Overviews](/en/academy/comprendre-le-geo/google-ai-overviews)). Content that is already well-ranked and technically solid for traditional SEO therefore tends to have more favorable ground to appear in Gemini answers than in other engines built on different logic — without this guaranteeing systematic pickup. Because Gemini answers are surfaced to an enormous default audience through Google Search and the Android ecosystem, even a modest improvement in how it treats your content can reach a disproportionately large number of users compared with a similar improvement on a smaller, opt-in chatbot.

## Claude: caution and limited sourcing

Claude, developed by Anthropic, has historically taken a more cautious stance on web browsing and citing external sources than its competitors, tending to rely more on the model's internal knowledge and to state explicit caveats when information can't be verified in real time. This caution can reduce how often it directly cites recent web sources compared to an engine like Perplexity, but it makes it all the more important for your brand to have a presence in widely distributed, widely referenced content that has a better chance of having been folded into the model's knowledge. That includes content picked up by major outlets, indexed documentation, and reference sites likely to be swept into future training data, rather than obscure, low-traffic pages.

## What to remember despite the differences

These architectural differences shouldn't lead to a scattered, engine-by-engine strategy: the fundamentals that make content citable (clarity, sourcing, self-contained information, perceived brand authority) hold everywhere. What varies is the relative weight of each factor and the freshness of the sources consulted. A good practice is to regularly observe how your brand appears (or doesn't) in each of these engines' answers on queries representative of your industry, rather than relying on generalities that quickly go stale given how fast these products evolve. Concretely, this means submitting the same panel of questions to Perplexity, ChatGPT, Gemini, and Claude at regular intervals, noting which sources are cited and how your brand is presented within them, then comparing these readings over time rather than relying on a single, isolated observation. This discipline of repeated, engine-by-engine observation is more revealing than a one-off audit: it lets you spot whether a change in your content or in a distribution partnership actually moved the needle on your presence.

Keep in mind too that these behaviors shift with every product update: a snapshot frozen at one point in time ages fast, which is why it's worth following the broader evolution of the [research landscape](/en/academy/comprendre-le-geo/paysage-de-la-recherche-en-2026) rather than treating any rule about a given engine as final. Treat everything above as a starting map, not a fixed territory: the safest habit is to re-test your own representative queries across all four engines every so often, since a description that felt accurate six months ago may already be out of date.

### [The goal: being everywhere (AI share of voice)](https://vurto.ai/en/academy/comprendre-le-geo/objectif-etre-partout-share-of-voice) (2026-06-24)

AI share of voice refers to how often a brand appears in generative engine answers across a set of queries representative of its industry, compared to its competitors. Unlike classic SEO ranking, where a single position matters at a time, the same AI query can surface several brands within the same answer: the goal isn't to be first, but to be present as often as possible.

## What share of voice means applied to AI

In traditional marketing, share of voice measures the portion of the conversation a brand occupies relative to its competitors, across all channels. Applied to GEO, the same principle means tracking, across a representative panel of prompts ("best software for X," "Y vs Z comparison," "how to solve this problem"), the proportion of answers where your brand is cited, alone or alongside others. A brand can very well be cited without being the top pick in a comparison: being mentioned regularly, even in a secondary position, builds cumulative presence in the mind of a user running multiple searches.

## Why a single position no longer describes visibility

In classic SEO, visibility often boils down to a position in a list: first page, first result. That binary model doesn't work the same way for a generated answer, which can mention three, four, or five brands within the same paragraph, with nuances of emphasis (order of citation, length of the mention, tone used) that a simple ranking doesn't capture. Share of voice replaces this single-position logic with a cumulative one: out of a hundred representative queries put to an engine, in how many answers does your brand appear, and with what quality of mention?

## Why being cited elsewhere matters as much as your own site

An often underestimated point: a generative AI doesn't only cite your own site to talk about your brand. It can just as easily cite a third-party comparison article, a user review, a case study published by a partner, or a discussion forum that mentions your product. This means your share of voice doesn't depend solely on what you publish yourself, but on the sum of your brand's mentions scattered across the web — a phenomenon detailed in the lesson on [brand mentions](/en/academy/comprendre-le-geo/brand-mentions-facteur-numero-un). Optimizing only your own site leaves out a significant, often majority, share of the opportunities to be cited.

## How to measure your share of voice

Measuring your AI share of voice means regularly submitting a representative panel of queries to several engines and noting the presence (or absence) of your brand, as well as that of your direct competitors, in the resulting answers. Done manually, this exercise is time-consuming and hard to reproduce consistently over time, since AI answers shift from one week to the next. Tools like Vurto automate this tracking by calculating a visibility score per engine, per brand, and per keyword, which makes it possible to track share-of-voice trends objectively rather than relying on one-off, non-comparable observations. Measurement frequency matters too: a single reading gives you a snapshot, but it's repetition at regular intervals (monthly or quarterly depending on how fast your industry moves) that turns that snapshot into an actionable trend. A one-off spike in citations after publishing a piece of content says nothing about how durable that presence is; repeated tracking is what lets you tell a temporary blip apart from a lasting gain in presence.

## Building a multi-source presence

Building your share of voice requires a presence spread across several types of sources: your own site (substantive articles, clear product pages), third-party sources that talk about your brand (trade press, independent comparisons, customer reviews), and activity that naturally generates mentions (partnerships, public appearances, content picked up by others). None of these sources is enough on its own: it's their accumulation, over time, that increases the odds a generative engine encounters your brand often enough to consider it a relevant answer worth citing. Take a concrete example: a company that publishes a substantive article on its own site, earns a mention in an independent comparison published by an industry outlet, and gets its product cited in a thread on a specialized forum statistically has better odds of being picked up by a generative AI on a given query than a company that pours all its effort into its own site alone, even if that site has better editorial quality taken in isolation.

## Tracking competitors to gauge your progress

Share of voice only makes sense when compared to that of your direct competitors on the same queries: being cited in 20% of answers means something very different depending on whether the leading competitor is cited in 15% or 70% of cases. That's why tracking AI share of voice naturally goes hand in hand with analyzing competitive presence, a topic covered in depth in the lesson on [analyzing competitors in AI answers](/en/academy/avec-vurto/analyser-ses-concurrents-dans-les-reponses-ia).

Being everywhere, in the end, doesn't mean chasing omnipresence at all costs, but methodically building a coherent, repeated presence across the sources generative engines actually consult when answering questions about your industry. Treat share of voice as a long-term editorial and PR metric to revisit quarterly, not a number to chase week to week.

### [Brand mentions: the #1 factor in AI visibility](https://vurto.ai/en/academy/comprendre-le-geo/brand-mentions-facteur-numero-un) (2026-06-25)

A brand mention is any reference to your brand's name on a third-party site, with or without a hyperlink pointing back to your site. For a generative engine, that simple mention already acts as a signal linking your brand to the topic at hand, which makes it one of the most frequently cited factors explaining why a brand shows up in AI answers.

## What counts as a brand mention (with or without a link)

In classic SEO, an inbound link (backlink) serves both as a recommendation and as a technical signal search engines use to assess a site's authority. A mention without a link historically had only indirect value. For a generative engine, the logic shifts: what matters first is the textual association between the brand name and a given context (a problem solved, a product category, a comparison), whether or not that association comes with a link. An article that says "for this use case, tools like [brand] and [competitor] are often recommended" creates that association even without a hyperlink.

## Why AI engines rely on them

Generative language models learn in part from statistical regularities present across enormous volumes of text: the more frequently and consistently a brand is mentioned in a given context, the more the model associates that brand with that context. For engines that rely on real-time web search, every new mention on an indexed page is one more occurrence the search algorithm can find and cite. In both cases, the frequency and consistency of mentions matter more than their raw count at any single point in time.

## Not all mentions are equal

A mention buried in the middle of a long, undifferentiated list of ten brands doesn't carry the same weight as a mention accompanied by qualitative context ("for small teams, [brand] stands out because..."). Similarly, a mention on a source the engine itself considers reliable and relevant (a specialized site, a recognized publication) carries more weight than an isolated mention on a barely visited site. This means the quality and context of a mention matter almost as much as its mere existence: a handful of qualitative mentions on relevant sources beats a large number of generic, interchangeable ones.

What counts as an "authoritative" source varies a great deal from one industry to another. In some highly specialized B2B fields, a technical forum frequented by professionals can carry more weight in a generative engine's perception than a general-interest press article that only skims the topic, simply because that forum is consulted and cited more often for that specific type of question. In consumer-facing sectors, conversely, trade press and widely distributed comparisons often play a more decisive role. Before concentrating your mention-generation efforts on a single type of source, it's worth checking, query by query, which sources are actually being cited today by the engines for questions in your industry, rather than applying a generic recipe that works elsewhere.

## Where to generate brand mentions

The most effective sources are the ones generative engines actually consult when answering queries in your industry: trade press, independent comparisons, directories and listings recognized in your field, forums and communities where users discuss solutions to their problems, case studies published by partners or customers. A PR strategy, an active and genuinely useful presence in relevant communities, or content partnerships with other players in the space generate this type of mention more durably than simply asking for a link.

## Mentions without links vs. backlinks: what changes

This doesn't mean links become useless: they remain a strong authority signal for traditional SEO and for engines that, like Gemini, lean heavily on existing search infrastructure. But the gap in importance between a link and a plain mention narrows in the context of GEO. Concretely, this widens the range of possible actions: a mention in a forum post well-regarded by the community, a citation in a press article without a link, or a useful answer given on a Q&A platform can all contribute to AI visibility, whereas they would have had limited SEO value in an approach focused solely on backlinks.

## Tracking and growing your brand mentions

Tracking your brand mentions means identifying where and how it appears across the web, including in contexts that don't generate directly measurable traffic through classic analytics tools. This connects directly to the [off-site signals that matter](/en/academy/strategie-seo-geo/signaux-off-site-qui-comptent) for an overall GEO strategy, and it's closely tied to the broader goal of [share of voice](/en/academy/comprendre-le-geo/objectif-etre-partout-share-of-voice): the more mentions there are, spread across varied sources, the higher the cumulative odds of appearing in an AI answer. Tools like Vurto let you track these mentions and the sources carrying them, rather than guessing where your brand shows up across the web.

In short, brand mentions function as distributed social proof: each mention, however modest, is an implicit vote of confidence that a generative engine can find and turn into a citation. This is groundwork measured in months rather than weeks, but its effect, once established, tends to hold up more durably than a one-off classic SEO action. It also compounds: each new mention makes the next one slightly more likely, as the surrounding conversation about your brand thickens across independent sources.

### [Google AI Overviews: the largest AI channel by volume](https://vurto.ai/en/academy/comprendre-le-geo/google-ai-overviews) (2026-06-26)

Google AI Overviews is the AI-generated summary displayed at the top of Google search results for a growing share of queries. Simply by being built directly into the world's most used search engine, it reaches a volume of searches that far exceeds all dedicated AI chatbots combined, even if the exact scale of that gap varies depending on the source and the measurement method.

## What Google AI Overviews is

AI Overviews generates a synthesis written directly into the results page, with links to a limited number of cited sources, appearing even before the classic organic results. Unlike ChatGPT or Perplexity, where the user starts a dedicated conversation, someone who sees an AI Overview generally didn't set out to get an AI-generated answer: they simply typed a query into Google the way they've done for years, and Google decided the query lent itself to an automatic summary.

## The scale: why this channel outpaces dedicated chatbots in volume

This native integration into Google search explains the scale gap with dedicated chatbots: by default it reaches a massive user base that has never opened an AI chatbot app and, for many, doesn't even realize the summary displayed is AI-generated. This is an important point for prioritizing GEO efforts: optimizing to appear in an AI Overview can affect a much larger volume of queries than optimizing exclusively for ChatGPT or Claude, particularly for common informational queries in your industry.

## How Google selects the sources it cites

Google states that AI Overviews rely on the same ranking and quality systems as classic organic search: content that isn't well-ranked for a given query has little chance of being cited in the corresponding AI Overview. This means technical SEO fundamentals remain a baseline requirement, not a secondary factor: clean indexing, clear structure, relevant [Schema.org markup](/en/academy/bases-seo-techniques/schema-org-comprendre), and fast load times still play a role. What then distinguishes cited content from ignored content within the AI Overview itself falls back on the broader criteria for [GEO-ready content](/en/academy/comprendre-le-geo/contenu-geo-ready): a clear, self-contained, verifiable answer to the question asked.

Google doesn't show an AI Overview for every query: certain categories, especially those considered sensitive (health, finance, very recent news), are still handled more cautiously, with rarer display or more caveats built into the summary. Conversely, simple, frequent informational queries on stable topics are the ones that most regularly trigger an AI Overview. Understanding which queries in your industry currently trigger one, and which don't, helps prioritize which pages to rework first.

## Improving your odds of appearing

Beyond the technical fundamentals, a few specific levers increase the odds of appearing in an AI Overview: covering a topic completely on a single page rather than spreading it across several pieces of content that compete with each other, placing a direct, extractable answer near the beginning of the content, keeping information up to date as the topic evolves (AI summaries appear to favor recent sources for time-sensitive subjects), and avoiding an overly promotional tone, which tends to be picked up less than a factual, neutral one. Explicitly signaling a content's last-update date, when it's genuine and not merely cosmetic, can also work in your favor on topics where currency matters: a guide updated last week inspires more confidence, all else equal, than an article whose publication date goes back several years with no visible revision. None of these levers guarantees an appearance, but their absence noticeably reduces the odds.

## AI Overviews aren't identical everywhere

The rollout of AI Overviews isn't uniform across countries and languages: Google has extended this feature progressively, market by market, and the behavior observed (how often it appears, which query types trigger it, which sources it favors) can differ noticeably between an English-speaking market, where the feature is generally more mature, and a market like France, where its rollout and behavior are still evolving. This means an observation made in one market doesn't automatically transfer to another, and it's better to directly observe how AI Overviews behave on your own queries, in your own language and country, rather than relying on observations made elsewhere.

## AI Overviews and the rest of the AI ecosystem

Appearing as a cited source in an AI Overview doesn't always translate into a click: several publishers and traffic analytics tools report a downward trend in click-through rate when an AI summary already answers the question directly on the results page itself, a phenomenon sometimes called "zero-click." That doesn't make the citation worthless, though: being identified as a source by Google remains a credibility signal and a form of brand exposure, even without a measurable visit. AI Overviews don't replace the other AI channels, they add to them, within a [research landscape](/en/academy/comprendre-le-geo/paysage-de-la-recherche-en-2026) that has become more fragmented than ever. Tracking your presence in both AI Overviews and dedicated chatbot answers gives a more complete picture of your overall visibility than monitoring just one of these channels. Given the volume it commands, even incremental improvements in AI Overview appearances are often worth prioritizing over chasing a smaller, more visible chatbot audience alone.

### [GEO numbers: market size, ROI and conversion](https://vurto.ai/en/academy/comprendre-le-geo/chiffres-du-geo-marche-roi) (2026-06-27)

The GEO market is too young and too fast-moving for any single figure to reliably sum up its size or return on investment. What can be said with more confidence are the directional trends observed by several independent players, and the methodological limits that call for caution around the statistics that circulate.

## A market still taking shape, hard to measure precisely

There's no stable, agreed-upon definition of what "the GEO market" covers: some analysts include AI visibility tracking software, others add consulting and agency services on top, and others still frame it as a shift of existing advertising and SEO budgets toward new practices. Without a shared definition, market-size figures published by different sources are generally not comparable to one another, even when presented in the same unit.

## Why comparing sources can be misleading

Two reports can show very different growth figures for the same period without one being "wrong" and the other "right": they often measure different scopes (global market versus a single country's market, software spend versus total spend including consulting), on different samples (large enterprises only, or all sizes combined), using different projection methods. Before adopting a growth or market-size figure, it's better to check exactly what it includes rather than comparing it as-is to another figure that appears contradictory.

## What public trends suggest about AI traffic

Several publishers and traffic analytics vendors report a directional increase in referral traffic coming from AI chatbots and AI Overviews over recent quarters. This observation comes up often enough to be taken seriously as an underlying trend. That said, the precise magnitude of that increase varies significantly depending on the measurement method used, the period observed, and the industry involved, and for most sites this traffic still remains a minority in absolute terms compared with classic search. It's therefore safer to note the direction of the movement than to repeat the exact percentage put forward by any one source.

## GEO ROI: why it's still hard to isolate

Measuring a return on investment requires linking an action (publishing a piece of content, earning a mention) to a measurable outcome (a visit, a conversion, a sale). But a large part of GEO's effect happens without a measurable click: a user who gets a satisfying answer directly inside ChatGPT or an AI Overview may remember your brand without ever visiting your site right away — an awareness effect that, if it converts at all, does so much later and through a different channel. Attribution tools capable of cleanly isolating this effect from the rest of marketing activity (classic SEO, PR, word of mouth) are still young and incomplete. A useful parallel here is classic SEO in its early days, where direct ROI was similarly hard to isolate from other marketing channels before multi-touch attribution methods became widespread over time. GEO is going through a comparable phase today: measurement tools exist but remain young, and a portion of the real effect will likely escape direct measurement for a while yet. That's not a reason to ignore the topic while waiting for perfect tools, but it does call for treating the available indicators as directional signals rather than definitive accounting proof.

## How to evaluate your own return, without borrowed numbers

Rather than importing a growth percentage or an average ROI calculated on an industry or sample that doesn't necessarily match yours, especially since two businesses in the same industry can see very different results depending on their starting visibility, their content maturity, and how competitive their specific queries already are, it's more reliable to build your own baseline: track how your visibility score evolves per engine and per query over time, watch incoming traffic identifiable as coming from AI referrers where that data is available, and cross-check it against qualitative feedback from prospects ("I found your offer through ChatGPT"). A tool like Vurto, by measuring your before/after progress on specific queries, gives a trend indication specific to your brand, which is a far more actionable indicator than a generic market statistic.

## Staying cautious about the statistics in circulation

A great many statistics circulating about GEO trace back to a single survey, often run on a small sample or an unpublished proprietary methodology, then repeated from article to article without the primary source ever being checked. Before repeating a figure, it's worth tracing it back to the original study, checking the sample size and method, and asking whether the context (market, language, industry) matches your own. One extra habit specific to market data: check the publication date of the cited study, since a market as young as GEO's can see its orders of magnitude shift significantly from one year to the next, which makes a two- or three-year-old statistic particularly fragile to reuse as-is. This need for traceability, covered more broadly in the lesson on [citations, numbers and evidence](/en/academy/comprendre-le-geo/citations-chiffres-et-preuves), applies especially to market-size and ROI figures, which are among the most frequently distorted as they spread. The lesson on [what academic research says](/en/academy/comprendre-le-geo/ce-que-dit-la-recherche-academique) extends this same caution to the more rigorous, but equally limited, ground of academic studies. When in doubt, a simple rule holds up well: the more precise and dramatic a GEO statistic sounds, the more scrutiny it deserves before you repeat it in your own content, and a round, headline-friendly number is often the first sign that it has been simplified somewhere along the way.

### [What academic research says about GEO](https://vurto.ai/en/academy/comprendre-le-geo/ce-que-dit-la-recherche-academique) (2026-06-28)

The most frequently cited reference paper on this topic is titled "GEO: Generative Engine Optimization," published in 2024 by researchers affiliated with Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi. It doesn't offer a magic formula for getting cited by an AI, but a methodology for testing and comparing different content optimization techniques against simulated generative engines.

## The founding paper: "GEO: Generative Engine Optimization"

This work is generally regarded as one of the first to formalize GEO as a discipline distinct from SEO, by asking the following question: since generative engines synthesize answers from multiple sources rather than simply ranking them, which characteristics of a piece of content increase the likelihood it gets picked up and featured in the generated answer? To answer this, the authors built a set of queries spanning varied domains, meant to serve as a reproducible benchmark for comparing different versions of the same content.

## Why this paper mattered for the field

Before this kind of work, discussion around visibility in AI answers stayed largely anecdotal: individual observations, non-reproducible tests, intuitions borrowed from traditional SEO and applied by analogy. The paper's main contribution isn't really any single technique it tested, but the fact of proposing a systematic, shared evaluation framework in the first place, one that let other researchers reproduce, challenge, or refine the results. It's this kind of open methodology, more than any single figure it contains, that became a reference point for subsequent publications on the topic.

## How the researchers evaluated generative visibility

The method involves submitting the same query to a generative engine, observing which sources are cited in the produced answer and with what relative weight, then modifying a source's content according to a given technique (adding citations, rephrasing, etc.) before repeating the experiment to see whether its visibility in the generated answer changes. This approach borrows from methods used in information retrieval research to evaluate result relevance, adapted to a case where the final output is no longer a list of links but a synthesized text.

## The main families of techniques tested

The researchers studied several categories of content modifications: adding citations and statistics, inserting quotes attributed to recognized figures, simplifying language, enriching text with relevant keywords, and generally improving fluency and structure. The central finding of this research is that some of these techniques have a measurable effect on a source's visibility in the generated answer, but that this effect is neither uniform nor guaranteed: its magnitude varies by the domain tested (health, law, technology...), the engine used, and the nature of the query. In other words, the existence of measurable optimization levers is established, but no technique works universally and consistently.

Since this initial work was published, other research teams and several industry companies have released their own studies or benchmarks on related questions, sometimes with different methodologies and conclusions. This diversity is actually healthy for the field: it signals that no single study, however rigorous, should be treated as the final word on the subject. A serious researcher or practitioner cross-checks several sources against each other rather than relying on a single reference, no matter how often it gets cited.

## The limits of this research

This work has important limitations to keep in mind. It reflects the generative engines and models as they existed at the time of publication in 2024: commercial systems evolve rapidly, and results obtained on a test configuration don't necessarily transfer identically to current versions of ChatGPT, Perplexity, Gemini, or Claude. The number of domains and queries tested, while designed to be representative, remains a limited sample compared to the real diversity of use cases. Finally, a controlled research environment differs from a commercial production system, which may incorporate partnerships, safety filters, or business logic that isn't publicly documented and is absent from the experimental setup.

## What this means for your content strategy

This kind of academic research should be read as directional confirmation rather than a precise instruction manual: it establishes that optimizing for generative engines isn't a random phenomenon and that measurable levers exist, without providing a formula along the lines of "this technique increases visibility by this percentage" that would hold for your particular case. The right practice is to lean on the broad principles it confirms — the importance of sourcing, covered in the lesson on [citations, numbers and evidence](/en/academy/comprendre-le-geo/citations-chiffres-et-preuves), and the importance of a source's perceived credibility, covered in the lesson on [E-E-A-T](/en/academy/comprendre-le-geo/eeat-expertise-autorite-confiance) — while directly observing, on your own content and queries, what actually works, rather than relying on a single figure, even one from a serious research paper. As with the market data covered in the lesson on [GEO numbers](/en/academy/comprendre-le-geo/chiffres-du-geo-marche-roi), the best approach remains checking the primary source before repeating a conclusion. Before citing a result from a research paper, make it a habit to read at least the abstract and the methodology section rather than just the title or a press article about it: that's usually where the nuances and limitations live, the ones media coverage tends to smooth over.

### SEO/GEO strategy

### [Understanding search intent](https://vurto.ai/en/academy/strategie-seo-geo/comprendre-intention-de-recherche) (2026-06-29)

Search intent is the goal a person is pursuing when they type a query into a search engine or a prompt into an AI assistant like ChatGPT. Guessing it correctly determines the format, tone, and substance of the content — or the answer — that will actually satisfy that person.

## The four main intent categories

Most queries fall into one of four categories.

**Informational.** The person wants to understand a topic, learn a concept, resolve a doubt. “What is GEO”, “how does a generative search engine work”, “why is my traffic dropping” are informational queries. This is the broadest category, and the one where long, educational content performs best.

**Navigational.** The person already knows where they want to go and uses the engine as a shortcut. “Google Search Console login”, “Semrush pricing”: there's little value in creating content here, unless you are the brand being searched for.

**Transactional.** The person is ready to act: buy, sign up, download. “Free trial SEO tracking tool”, “buy a .com domain name”. The content should remove the last barriers to a decision, not re-explain the basics.

**Comparative.** An intermediate intent, between informational and transactional: the person is weighing several options and looking for arguments to decide. “Semrush vs Ahrefs”, “best CMS for a professional blog”. These queries are particularly likely to be cited by generative engines, which tend to produce comparison tables.

## Why this typology matters as much for GEO as for SEO

In classic SEO, misreading intent shows up as poor rankings: Google favors pages that match the expected result type for a given query. If the top ten results are guides and you publish a product page, you won't rank, no matter how good your content is.

In GEO, the stakes are different but just as structural. A generative engine doesn't rank ten links: it builds a single answer, usually drawing on three to eight sources. If it interprets the query as informational and your page reads like a sales brochure, it simply won't be cited — not because it's bad, but because it doesn't answer the right question. This is also why a mismatch in intent costs more in GEO than in classic SEO: there is no “page eleven” in a generated answer.

## How to guess the intent behind a query

Three methods complement each other.

**Look at the results already ranking.** Type the query into Google and see what shows up at the top. Blog posts? The intent is probably informational. Product pages or e-commerce category pages? Transactional. Comparison tables? Comparative. The engine has already done part of the classification work for you, based on the behavior of millions of users.

**Analyze the wording.** Question words (“how”, “why”, “what is”) almost always signal informational intent. Action verbs (“buy”, “download”, “sign up”, “try”) signal transactional intent. Two brand names, or words like “vs”, “or”, “best”, signal comparative intent. A single brand name with nothing else signals navigational intent.

**Ask a generative engine directly.** Put the query to ChatGPT or Claude and observe the type of answer it produces spontaneously: a structured explanation, a comparison table, a step-by-step list. This is a concrete indicator of how the engine itself has categorized the request — useful for GEO, since that engine is exactly the one you're trying to satisfy.

Combined on a real example, these three methods reinforce each other: for the query “CRM for small business”, the Google results mix guides and comparison pieces, the phrasing contains no action verb but two generic terms, and a generative engine typically answers with a list of criteria followed by software examples. The diagnosis becomes clear: a mainly comparative intent, with an informational undertone.

## The case of single-word, ambiguous queries

A query like “GEO” or “SEO”, reduced to a single generic word, is particularly hard to classify: it can just as easily reflect an informational intent (“what is this”) as a navigational one toward a specialized site. In that case, it's better to rely on the context of the site publishing the content rather than the query alone: a site already positioned around teaching the subject has every interest in treating this query as informational, then guiding the reader toward more precise topics once the definition is in place.

## A single query can carry more than one intent

Many queries are mixed. “Best invoicing software for freelancers” is both comparative (several options at play) and transactional (the person is close to choosing). In that case, the best answer often combines a comparison table near the top with clear action links for each option, rather than forcing a choice between the two formats.

This is also why a generic query like “GEO” deserves to be broken down into sub-topics handled separately rather than a single catch-all article: this logic of splitting a topic up is at the heart of the [Hub & Spoke architecture](/en/academy/strategie-seo-geo/architecture-hub-and-spoke), which organizes content around a broad intent split into more precise ones.

## Matching content to the intent you detect

Once the intent is identified, the content's structure follows almost mechanically:

- Informational intent → long, educational content, with clear definitions and concrete examples. This is the territory of the [pillar page](/en/academy/strategie-seo-geo/la-page-pilier).
- Comparative intent → tables, objective criteria, the pros and limits of each option stated without excessive bias.
- Transactional intent → short content, focused on the decision, with social proof and a clear call to action.
- Navigational intent → little editorial effort needed; the real work is technical (a fast, well-indexed page).

Before even searching for keywords with tools like [Google Search Console](/en/academy/strategie-seo-geo/mots-cles-avec-google-search-console) or [free tools](/en/academy/strategie-seo-geo/mots-cles-avec-outils-gratuits), mentally sorting each candidate query into these four families keeps you from producing content that will never find its audience — human or generative.

### [Finding keywords with Google Search Console](https://vurto.ai/en/academy/strategie-seo-geo/mots-cles-avec-google-search-console) (2026-06-30)

Google Search Console is the most reliable free source of keywords there is, because it doesn't guess anything: it shows the queries that have already generated impressions or clicks toward your site. Unlike a suggestion tool, you're working from real behavioral data, not hypotheses.

## Why start with Search Console rather than a third-party tool

Paid keyword research tools (Semrush, Ahrefs, and the like) estimate search volumes from panels and statistical models. Search Console gives you the exact queries typed by real visitors who saw or clicked on one of your pages. The information is free, exhaustive for your own domain, and directly actionable: every row in the report corresponds to an opportunity that's already started, not a guess. No sign-up beyond verifying your site is required, which makes it the natural first stop before spending on anything else.

The limitation is symmetrical: Search Console only shows queries for which you already appear, even weakly. It's ideal for developing an existing topic, less so for discovering a theme completely absent from your site — for that, [free tools like Autocomplete or Google Trends](/en/academy/strategie-seo-geo/mots-cles-avec-outils-gratuits) take over.

## Getting to the right report

In Search Console, open the “Performance” tab, then “Search results”. By default the tool shows clicks from the last three months: widen the range to 16 months (the maximum available) to capture seasonal variation. Then check the “Queries” box in the table below: that's the raw list of everything people typed to land — or almost land — on your site.

## The “high impressions, low clicks” method

Sort the table by descending impressions, then look for queries that combine a high number of impressions with few clicks, at an average position between 5 and 15. These are topics Google already considers you relevant for, but not relevant enough to convince people to click, or not yet ranking high enough to be seen.

Two levers apply here: improve the existing content (structure, a more direct answer in the first lines, examples), or create a dedicated page if the topic deserves better than a paragraph buried inside a broader article.

A concrete example: a query with 4,000 monthly impressions, 40 clicks, and an average position of 8 has a click-through rate of about 1% — well below the average expected at that position (often 3 to 5%). That's a clear signal the title or meta description isn't compelling enough, or that the page doesn't fully answer the query once clicked.

## Filtering by page to consolidate what already works

The other useful view: filter by URL instead of by query. Pick a page you consider strategic and look at every query that surfaces it. You'll often find that a page attracts traffic on phrasings you hadn't anticipated when writing it — a strong signal to enrich that page rather than create a new one that would duplicate it. This filter is also the right place to check that none of your strategic pages is missing from results for queries that should logically surface it — a possible sign the content isn't well aligned with visitors' real intent.

## Refining with additional filters

Search Console also offers filters by country, device, and search type. If you're targeting an English-speaking market alongside a French one, compare queries by country: the wording, and sometimes the topics themselves, often differ noticeably. The device filter helps spot queries that perform well on mobile but not on desktop — a sign that the expected answer format (quick, consumable on the go) differs from what your current page offers. The search type filter (Web, Image, Video) is worth a separate look too: a topic that draws meaningful Image or Video traffic signals a format preference that a text-only page cannot fully satisfy on its own.

## Grouping queries into topics, not isolated keywords

A common mistake is treating every Search Console query as a potential page. In practice, dozens of different phrasings (“how to find keywords”, “keyword research method”, “free keyword finder”) cover the same intent and should feed a single piece of content, not three pages that would cannibalize each other. Export the queries, sort them by theme in a spreadsheet, and keep only one topic per cluster — this logic of grouping by entity rather than by isolated keyword is developed in the lesson on [thinking in entities](/en/academy/strategie-seo-geo/penser-en-entites).

## Exporting the data for long-term tracking

Search Console only retains a rolling 16 months of data: beyond that, the history disappears. To keep a longer record and compare year over year, export the reports regularly (via the export button or a connection to Google Sheets) and archive them in a dated spreadsheet. This simple habit makes it possible, after two or three years, to spot topics whose interest is structurally growing and those that are stagnating — information no free tool provides directly.

## Turning the data into a content plan

Once the groupings are done, sort them by intent (see the lesson on [search intent](/en/academy/strategie-seo-geo/comprendre-intention-de-recherche)) and by impression volume. High-volume topics with a clear intent become candidates for a pillar page; narrower variants feed spoke pages that enrich the internal linking around that pillar.

This work is best redone every quarter: queries evolve, new phrasings appear, and Search Console remains the best free compass for tracking that movement over time. A recurring calendar reminder is usually enough to keep the habit alive once the first pass is done.

### [Finding keywords with free tools](https://vurto.ai/en/academy/strategie-seo-geo/mots-cles-avec-outils-gratuits) (2026-07-01)

Without a budget for a paid SEO tool, three free resources are enough to build a solid list of topics: Google Autocomplete, the “People Also Ask” box, and Google Trends. Combined, they reveal both the exact language people use and how their interest shifts over time.

## Google Autocomplete: capturing the exact language people use

Type the start of a query into the Google search bar without hitting enter, and look at the suggestions offered. They're generated from the most frequent real searches associated with that beginning of a phrase — a direct source of authentic phrasing, often closer to everyday language than what a writer would come up with on their own.

A concrete example: for a site selling invoicing software, typing “invoicing software” into Google surfaces suggestions like “free invoicing software”, “invoicing software for freelancers”, “invoicing software legal requirements”. Each of these variants corresponds to a different user profile (freelancer, small business, compliance concern) and can justify a dedicated piece of content rather than a single generic page.

To go further, add a letter after your base query (“how to find keywords a”, then “b”, “c”...): each letter triggers a new batch of suggestions, letting you cover the full alphabet of possible variants. This technique, sometimes called the “alphabet soup” method, is tedious to do by hand but stays free and remarkably thorough.

## People Also Ask: digging into sub-questions

On the Google results page, the People Also Ask box lists questions related to your query. Clicking one reveals new ones, generated dynamically: a simple way to map out every sub-question a reader might have around a given topic.

This box is especially valuable for GEO: generative engines often structure their answers as question-and-answer pairs, and the questions appearing there are a good indicator of which ones a generative engine will consider relevant to round out an answer.

## Google Trends: comparing and dating topics

Google Trends doesn't give an absolute search volume, but a relative interest curve over time, comparable across several queries. Two practical uses: detecting a topic's seasonality (useful for planning publication) and avoiding spending writing time on a theme whose interest is in structural, not temporary, decline.

Trends also lets you compare several phrasings of the same topic (“GEO” vs “generative engine optimization” vs “AI search optimization”) to see which dominates in your target region — useful information for choosing a piece's main title. Another use: comparing your main topic to a related one to prioritize your editorial effort. If “generative AI” is climbing noticeably faster than “chatbot” in Trends over the past twelve months, that's a signal to lean your titles and subheadings toward the growing term, without necessarily dropping the older one if your current traffic still depends on it. Trends also breaks results down by region within a country, which can matter for a business whose customer base is geographically concentrated rather than spread evenly across the market.

## Bonus: YouTube suggestions, forums, and social platforms

The YouTube search bar works on the same autocomplete principle as Google, but reflects somewhat different intents, often more tutorial- or demonstration-oriented. If your topic lends itself to video, it's worth checking separately.

Forums (Reddit, niche communities) and platforms where questions are asked in natural language, with context and nuance, are another valuable source — particularly because generative engines increasingly cite this type of content as a source, provided it offers a clear, verifiable answer.

## Keeping the collection in a single spreadsheet

To avoid losing track across the three sources, centralize everything in one spreadsheet: a column for the exact phrasing, one for the source (Autocomplete, People Also Ask, Trends), one for the likely intent, one for status (to do, in progress, published). This spreadsheet becomes the backbone of your editorial calendar and keeps you from redoing the same research twice a few weeks apart.

## Combining the three sources without duplicating work

The efficient method: start from a broad query identified in [Search Console](/en/academy/strategie-seo-geo/mots-cles-avec-google-search-console) or a topic strategic to your business, run it through Autocomplete to collect phrasings, through the questions box to collect sub-questions, then through Trends to confirm interest isn't declining. Group everything by intent (see the lesson on [search intent](/en/academy/strategie-seo-geo/comprendre-intention-de-recherche)) rather than treating each suggestion as a separate piece of content. A realistic pace for a small team: dedicating thirty minutes a week to this collection is enough, over several months, to build up a reserve of dozens of topics ready to be prioritized in the editorial calendar.

## The limits to keep in mind

These free tools give no reliable search volume data or competition level: they tell you what to search for, not exactly how many people search for it or whether the topic is already saturated. One simple way to partly offset this gap: cross-check the rough volume suggested by Google Ads Keyword Planner (accessible for free with an ad account, even without an active budget) against the trends observed in Trends. The figure remains a wide range, but it gives enough of a sense of scale to prioritize between several candidate topics. These tools should be seen as a starting point for mapping a topic, to be cross-checked with your own performance data and, budget permitting, a paid volume tool.

This mapping of topics connected by meaning rather than by the mere presence of a keyword ties directly back to the logic developed in the lesson on [thinking in entities](/en/academy/strategie-seo-geo/penser-en-entites).

### [Thinking in entities, not just keywords](https://vurto.ai/en/academy/strategie-seo-geo/penser-en-entites) (2026-07-02)

An entity is an identifiable object in the real world — a person, a brand, a place, a product, a concept — that search engines and generative AI recognize and connect to one another, regardless of the exact words used to describe them. Thinking in entities rather than isolated keywords changes how you organize content, and becomes central once that content also needs to be understood by a generative engine.

## What an entity looks like in practice

Vurto is an entity (a brand). Paris is an entity (a place). Google Search Console is an entity (a product). Generative engine optimization is an entity (a concept), just like its more common abbreviation, GEO. A keyword, by contrast, is just a string of characters: “GEO” and “generative engine optimization” are two different keywords, but they can point to the same entity if the engine manages to establish the equivalence.

This ability to connect several phrasings to a single object is what distinguishes keyword-based search from entity-based search: Google (through its Knowledge Graph, since 2012) and the language models behind ChatGPT or Claude no longer look only for textual matches — they try to identify what is actually being talked about.

## The knowledge graph: connecting entities to one another

A knowledge graph associates each entity with attributes (a brand has an industry, a founding date, products) and with relationships to other entities (a brand belongs to an industry, competes with other brands, is based in a place). When an engine processes a query, it doesn't just look for the words it contains: it tries to map it to one or more entities in its graph, then pull in related entities that can enrich the answer.

This is the mechanism behind why searching for a niche brand often surfaces, alongside the results, its competitors, its industry, or related products: the engine is navigating the graph, not just an index of words. The same logic applies to generative engines: when asked to compare two brands, a model draws on whatever relationships it has been able to establish between the corresponding entities, not solely on text it happens to have seen mentioning both names together.

Take an example: a search for “Claude” can refer to a first name, a Roman emperor, or an AI model built by Anthropic. The engine resolves this ambiguity using context — the other words in the query, browsing history, the relative popularity of each entity — and maps the query to the most likely entity before even looking for pages to display. Content that never explicitly states which entity it's talking about makes this disambiguation harder, at the cost of its own visibility.

## Why keywords alone no longer cut it

A content strategy built solely on lists of isolated keywords often produces redundant pages: several similar pages targeting variants of the same word, without real added value between them. A strategy built on entities naturally groups these variants around a single topic covered in depth, which reduces cannibalization and makes it easier for the engine to understand the subject.

It's also a shift in framing: instead of asking “which keyword should I target”, the question becomes “which entities should my content cover, and how do I connect them explicitly”. A concrete example: a site that publishes twelve different pages around variants of the word “GEO” without linking them together dilutes its authority across twelve competing URLs instead of concentrating it on one clearly established entity — a common cannibalization scenario for sites that still think purely in keywords.

## What this changes for writing

In practice, thinking in entities means:

- Naming important entities explicitly rather than relying on pronouns or vague paraphrases throughout an article (“that kind of tool” instead of repeating “Google Search Console”).
- Clearly stating the relationships between entities: that a tool belongs to a category, that a method is an alternative to another, that a concept is a subset of a broader one.
- Staying consistent in the terminology used to name a given entity across an entire site, rather than alternating synonyms that blur the association.
- Structuring content so a central entity is covered in depth in one place, with links to related entities — exactly the principle behind the [Hub & Spoke architecture](/en/academy/strategie-seo-geo/architecture-hub-and-spoke).

## An example applied to internal linking

In practice, across a content cluster dedicated to SEO, the entity “Google Search Console” should be named consistently on every page that mentions it, with a link pointing to the page that covers it in depth rather than a different definition each time. Conversely, calling the tool “Search Console” in one place, “the Google search console” in another, and “GSC” elsewhere, without ever explicitly establishing the equivalence, makes the matching work harder for an engine that has to decide whether these three mentions really refer to the same thing.

## Entities and GEO: an even more direct stake

For a classic search engine, misidentifying an entity costs you a ranking. For a generative engine, it can cost you an entire citation: if the model can't connect your content to the entity it's trying to describe in its answer, it will simply look for that information elsewhere. [Structuring an article for GEO](/en/academy/comprendre-le-geo/structurer-un-article-pour-le-geo) and producing [GEO-ready content](/en/academy/comprendre-le-geo/contenu-geo-ready) both rely heavily on this entity-by-entity clarity.

The full vocabulary of these notions — entity, knowledge graph, and the other terms of GEO — is gathered in the [glossary of GEO and generative AI terms](/en/academy/strategie-seo-geo/glossaire-geo-aeo-ia-generative), useful for checking a definition while writing.

### [Site architecture: the Hub & Spoke model](https://vurto.ai/en/academy/strategie-seo-geo/architecture-hub-and-spoke) (2026-07-03)

The Hub & Spoke model organizes a site around a central page — the hub, or pillar page — that covers a broad topic, surrounded by secondary pages — the spokes — each covering a sub-topic in depth and consistently linking back to the hub. This architecture, popular in SEO for years, serves visibility in generative engines just as well.

## The principle, in one image

Picture a bicycle wheel. The hub at the center is the pillar page: a page that introduces an entire topic (say, “GEO” or “keyword research”) without going deep into every detail. The spokes are pages that cover a specific sub-topic (“thinking in entities”, “finding keywords with Google Search Console”) in depth, each linked to the hub and, ideally, to other relevant spokes.

A hub with no spokes stays shallow on every sub-topic. Spokes with no hub remain isolated pages, with no overall structure or coherence signal for engines to pick up on. It's the combination of both that creates value.

## Why this model beats a set of isolated pages

Without an explicit structure, a site that publishes content regularly often ends up accumulating similar pages with no clear relationship between them, competing against each other instead of reinforcing one another. The Hub & Spoke model imposes the opposite discipline: before publishing a new page, you ask which hub it belongs to and which existing spokes it should reference. Content organizes itself into coherent clusters rather than a flat list of articles.

A concrete example: a site that publishes ten independent articles on aspects of SEO, with no hub and no cross-links, will statistically struggle more to rank any one of them than a site that organizes those same ten articles into a single cluster around a shared pillar page.

## What star-shaped linking brings to classic SEO

Three well-documented benefits. First, the distribution of link equity: every spoke that earns external links passes some of that value to the hub through internal linking, reinforcing the cluster's most strategic page. Second, a signal of topical depth: an engine that sees a hub surrounded by ten specific, well-linked pages understands the site covers the subject seriously, not just on the surface. Third, easier crawling: a star-shaped architecture limits the number of clicks needed to reach any page from the hub, which favors faster, more complete indexing.

## What it brings specifically to GEO

For a generative engine that has to answer a precise question, a well-built Hub & Spoke cluster offers exactly what it's looking for: a page that addresses the exact sub-topic of the question, without dilution, and a broader context (the hub) that confirms this page belongs to a site that genuinely knows the subject — a trust signal that matters when the engine picks which sources to cite.

It's also easier to maintain over time: when a topic evolves, updating one precise spoke is simpler than reworking a monolithic article that mixed every level of detail together. And it scales more gracefully with a team: different writers can own different spokes in parallel without stepping on each other's territory, as long as everyone agrees on what the hub already covers.

## Building your own Hub & Spoke, step by step

1. **Identify the broad topic** that deserves a hub: a theme you have enough content, or enough material, for to justify five to ten distinct sub-topics.
2. **List the sub-topics** using data you've already gathered — queries from [Search Console](/en/academy/strategie-seo-geo/mots-cles-avec-google-search-console), suggestions from free tools, recurring questions from your customers.
3. **Write or identify the hub**: a [pillar page](/en/academy/strategie-seo-geo/la-page-pilier) that introduces every sub-topic and points to its corresponding spoke.
4. **Write the spokes** one at a time, each linked to the hub from the introduction onward, and supplemented with links to neighboring spokes where relevant — this work of linking pages to one another is covered in the lesson on [internal linking](/en/academy/strategie-seo-geo/maillage-interne).
5. **Check the consistency of entities** used from page to page: the [glossary](/en/academy/strategie-seo-geo/glossaire-geo-aeo-ia-generative) helps keep terminology stable across the whole cluster.

Expect anywhere from several weeks to a few months for a complete cluster to show visible results in rankings or citations, the time it takes engines to recrawl and reassess the whole set of linked pages.

## How many spokes does a hub need

There's no fixed number, but a cluster starts producing measurable effects from around five to seven spokes, well linked to each other and to the hub. Below that, the signal of topical depth stays too weak to stand out from a site that only covers the subject on the surface. Beyond around fifteen spokes, it's often worth asking whether the topic should be split into two separate hubs, each with its own, tighter set of spokes.

## Common mistakes to avoid

The first mistake is creating spokes that repeat the hub's content instead of deepening it: every spoke should add information the hub, being more general by design, can't develop. The second is forgetting links back to the hub once spokes are published, or adding them only once at the bottom of the page with no context, where they go largely unclicked. The third, more subtle, is building a hub too early: without at least three or four spokes already written, a hub remains an empty shell that proves nothing. A fourth, less obvious mistake is linking every spoke to the hub but never to each other, which leaves the cluster looking more like a set of spare spokes than an interconnected wheel.

### [The pillar page](https://vurto.ai/en/academy/strategie-seo-geo/la-page-pilier) (2026-07-04)

A pillar page is a site's reference content on a broad topic: it covers the topic's major dimensions with a structured overview, then links out to more specific pages — the spokes — for each sub-topic covered in depth.

## Its role in a Hub & Spoke cluster

In the [Hub & Spoke architecture](/en/academy/strategie-seo-geo/architecture-hub-and-spoke), the pillar page plays the role of the hub: it isn't meant to say everything about every sub-topic, but to give the reader — human or engine — a clear map of the subject, with an entry point to each deeper piece. It's the page that receives the most incoming internal links, and often the most external links too, which makes it the most strategic page in the cluster to push up in search results.

## Typical structure

An effective pillar page usually follows this pattern:

- An introduction that directly answers the topic's implicit question, in one or two sentences, before any elaboration.
- A table of contents or a series of H2 sections, each matching a sub-topic, with a few summary paragraphs per section — just enough for the section to stand on its own without duplicating what the corresponding spoke says.
- An explicit, visible link to the [spoke page](/en/academy/strategie-seo-geo/la-page-spoke) dedicated to each sub-topic, placed within the flowing text rather than relegated to the end of the article.
- A conclusion or closing section that puts the whole topic back into perspective, useful especially for readers who never click through to a spoke.

## How long should a pillar page be

There's no magic word count, but a pillar page is almost always longer than the spokes surrounding it, because it has to cover several sub-topics even briefly. In practice, that often works out to 1,500 to 3,000 words for a genuinely broad topic. Below that, the topic is likely not covered with enough depth to justify pillar-page status; above it, the risk is duplicating what the spokes should be covering in detail.

## What a pillar page should not do

Two traps come up often. The first is turning the pillar page into an exhaustive article that covers every sub-topic in depth: it becomes so long that it competes with its own spokes instead of showcasing them, and [internal linking](/en/academy/strategie-seo-geo/maillage-interne) loses its purpose. The second is the opposite: a pillar page so short it's just a list of links with no real standalone reading value — a generative engine will then have no reason to cite it on its own merits.

## Letting a pillar page evolve over time

A pillar page is never finished. Every time a new spoke is published on the topic, the pillar page should be updated to include it: a new summary section, a new link, possibly a reorganized table of contents if the topic has grown enough to deserve a different breakdown. It's also the natural place to check that the content stays [GEO-ready](/en/academy/comprendre-le-geo/contenu-geo-ready): direct answers at the start of each section, consistent terminology, a structure readable by a human and by a generative engine pulling a summary from it alike.

### [The spoke page](https://vurto.ai/en/academy/strategie-seo-geo/la-page-spoke) (2026-07-05)

A spoke page covers one precise sub-topic in depth, with a level of detail the pillar page can't afford, and consistently links back to it to place the topic in its broader context.

## Its role in the cluster

Where the [pillar page](/en/academy/strategie-seo-geo/la-page-pilier) gives an overview, the spoke answers one precise question with all the depth it needs: concrete examples, a step-by-step method, edge cases, nuances the pillar has no room to develop. In the Hub & Spoke architecture, it's the spoke that captures traffic on long, precise queries, while the hub captures traffic on the broad topic.

## What sets a good spoke apart from a standalone article

A spoke isn't just another blog post: it exists in explicit relation to a hub and, ideally, to the other spokes in the same cluster. Three things distinguish it from an article published with no overall structure: a clear link to the hub from the introduction or first section, one or more links to neighboring spokes when the topic calls for it, and a deliberately narrow scope — a spoke that spills into another spoke's territory creates redundancy and blurs the signal sent to engines.

## Typical structure

- A direct answer to the question posed by the title, in the first few lines.
- A body organized into sections that each dig into one aspect of the sub-topic, without revisiting generalities already covered by the hub.
- Concrete examples, numbered steps, or figures where relevant — this level of detail is often what justifies the page's existence in the first place.
- A link to the hub, and to other relevant spokes, woven naturally into the text rather than bolted on.

## Length and level of detail

A spoke is generally shorter than a pillar, but not necessarily short in absolute terms: it all depends on the sub-topic's complexity. A very operational how-to guide can fit in 600 words; a topic requiring nuance, edge cases, or figures can easily justify 1,200 to 1,500 words. The criterion to keep in mind isn't a target word count, but the question: have I covered this sub-topic deeply enough that a reader won't need to look for the answer elsewhere?

## Spokes: the preferred playing field for GEO

Generative engines most often answer precise questions, not broad topics — exactly the spokes' territory. A well-built spoke, one that directly answers a question in its first lines and then develops the useful detail, statistically stands a better chance of being extracted and cited than a general pillar page, which does more to build the site's overall credibility around the topic.

## Producing spokes regularly without sacrificing depth

The main obstacle to a successful Hub & Spoke cluster is often writing time: a hub without enough spokes stays a plan with no content. To keep a steady publishing pace without letting quality slip, some teams rely on writing-assistance tools to produce a structured first draft of each spoke, leaving a person to verify facts, adjust the tone, and add examples specific to the business. This is the principle covered in the lesson on [generating GEO-ready content with Vurto](/en/academy/avec-vurto/generer-du-contenu-geo-ready-avec-vurto).

Once several spokes are published, it's still essential to rework the [internal linking](/en/academy/strategie-seo-geo/maillage-interne) between them, and to consider [link building for a small business](/en/academy/strategie-seo-geo/netlinking-pour-une-tpe-pme) to strengthen the cluster from outside.

### [Internal linking](https://vurto.ai/en/academy/strategie-seo-geo/maillage-interne) (2026-07-06)

Internal linking is the set of links that connect a site's pages to one another. Done well, it helps visitors navigate, distributes authority across your pages, and gives search engines and generative engines alike a map of how your content relates.

## Why internal linking matters for both SEO and GEO

In traditional SEO, internal linking has always served two purposes: making pages easier to crawl (bots discover pages by following links) and passing on some authority from one page to another through the links pointing to it. A page that's well linked from other pages on the site typically gets indexed faster and ranks better.

For GEO, the logic is similar but the stakes differ. Language models and the systems that power them (through direct crawling or real-time content retrieval) also rely on a site's structure to understand how topics connect. Clear internal linking signals that a pillar page is authoritative on a topic and that the pages linked from it are its sub-topics. This helps generative AIs place a piece of content in its thematic context rather than reading it as an isolated page.

## Best practices for anchor text

The anchor is the clickable text of a link. It's the most overlooked part of internal linking, even though it carries a large share of the information about the destination page.

- Use descriptive anchors close to the actual topic of the destination page, rather than generic phrases ("click here", "learn more").
- Vary the wording from one page to another instead of systematically repeating the exact match of your main keyword: it stays more natural and avoids overly mechanical optimization.
- Place the link inside a sentence that still makes sense without it, so the surrounding context confirms the destination page's topic.
- Avoid misleading anchors: a link should lead exactly where its text suggests.

## Click depth: an often underrated factor

Click depth measures how many clicks it takes to reach a page from the homepage. A page three or four clicks away from the homepage is harder to discover, for visitors and crawlers alike, than one reachable in one or two clicks.

There's no universal rule for the maximum acceptable depth — it depends on site size — but the goal stays the same: bring important pages closer to the site's root, notably through the main menu, pillar pages, or a summary page. A site organized around a [hub & spoke architecture](/en/academy/strategie-seo-geo/architecture-hub-and-spoke) naturally supports this, since every spoke page links directly back to its pillar page.

## How internal linking distributes authority across pages

Every inbound link to a page can be seen as a vote of confidence. A page that receives links from many well-linked pages on the site accumulates more internal authority than an isolated page. That's why [pillar pages](/en/academy/strategie-seo-geo/la-page-pilier) usually receive the most internal inbound links: they act as hubs that redistribute this authority to the spoke pages they reference.

This distribution isn't fixed. Adding a link from a high-traffic or high-authority page to a newer or less visible one is one of the fastest ways to give the latter a boost, without waiting for new external links.

## Internal linking and thematic understanding by AI

Beyond distributing authority, internal linking also clarifies topics. When several pages covering related subjects link to each other with consistent anchors, they form a set that engines can recognize as a coherent whole — which ties back to the idea of [GEO-ready content](/en/academy/comprendre-le-geo/contenu-geo-ready): content that's easy to place in context is more likely to be picked up correctly in a generated answer.

Conversely, a site where pages exist without any logical link between them — each optimized in isolation — makes this reading harder. The risk isn't just weaker visibility, but also poorly contextualized content if it does get cited: an AI that doesn't perceive the connection between your product page and its related buying guide risks presenting one without the other, or citing a competitor who made that link more explicit.

## A simple method to audit your internal linking

You don't need sophisticated tools to get started:

1. List your most important pages (pillar pages, strategic product pages, conversion pages).
2. For each one, check how many internal pages link to it and with which anchors.
3. Identify orphan pages — those receiving no internal links — and add links to them from thematically related pages.
4. Spot generic or repetitive anchors and rewrite them so they actually describe the destination page.
5. Check that your pillar pages link out to all their spoke pages, and vice versa.

This audit is worth repeating regularly, especially after publishing new content: that's when adding a few well-chosen internal links has the most impact, both for visitors and for how well search and AI engines can read your site.

### [Link building for a small business](https://vurto.ai/en/academy/strategie-seo-geo/netlinking-pour-une-tpe-pme) (2026-07-07)

Link building refers to all the external links pointing to your site from other domains. For a small or medium business, with limited budget and time, the goal isn't to compete on volume with large accounts, but to build a small number of relevant, durable links.

## Why link building still matters at a small scale

An external link is still a trust signal: another site agrees to point to yours, which search engines interpret as a sign of reliability. For GEO, the effect is indirect but real: external links contribute to a site's overall reputation, and that reputation often translates into [brand mentions](/en/academy/comprendre-le-geo/brand-mentions-facteur-numero-un) on other pages, forums, or media outlets — and it's precisely these mentions, linked or not, that weigh most heavily on visibility within AI-generated answers.

For a small business, it's better to think in terms of quality rather than quantity: ten links from sites recognized in your industry or region carry more weight than a hundred links from generic directories unrelated to your business.

## Local press and media relations

Local press (regional newspapers, local news sites, community radio) remains one of the most accessible levers for a small business. Unlike national press, local newsrooms readily cover economic news from their area: an opening, hiring, innovation, community involvement, or an event organized by a local company.

A few concrete approaches:

- Build a list of journalist contacts covering your industry and geographic area.
- Offer an angle useful to their readers, not just a commercial announcement: a figure, a trend you've observed in your market, a testimonial.
- Respond to local news when your expertise is relevant, rather than waiting only for your own announcements.

Media relations don't systematically guarantee a link, but they often generate citations and mentions of your brand, which are valuable in themselves for GEO.

## Partnerships and co-produced content

Partnerships with non-competing businesses in your ecosystem (suppliers, clients, professional associations, chambers of commerce) are a natural, durable source of links. A co-written article, a joint case study, a mutual interview, or a simple "partners" page mentioning your business create links anchored in a real relationship, which makes them more solid than a link bought or negotiated cold.

Professional associations and federations in your industry also often publish member directories or studies featuring their members: appearing in them is generally simple and relevant.

## Qualified directories: which to choose, which to avoid

Not all directories are equal. Generic, poorly moderated directories unrelated to your industry bring little value and can even hurt your image if the site is low quality. Conversely, a few categories of directories remain relevant for a small business:

- Sector directories recognized in your field of activity.
- Review and local listing platforms (business listings, chambers of commerce, tourism offices where relevant).
- Directories linked to your status or certifications (craftsmanship, social and solidarity economy, quality certifications).

The goal isn't to multiply listings, but to make sure the few directories you choose are consistent with your industry and with the information shown on your Google Business Profile — a point covered in the [local SEO](/en/academy/strategie-seo-geo/seo-local-fondamentaux) lesson.

## Content that naturally attracts links

The most durable method, though the slowest, is still to produce content other sites have an incentive to cite: a complete practical guide on a technical topic in your trade, original data or first-hand experience, a simple free tool, or a well-argued position on a topic that matters to your industry.

This type of content doesn't generate links overnight, but each new publication of this kind increases the chances of being picked up — by a specialized blog, a forum, or a press article — without any active outreach on your part.

## What to absolutely avoid

Some practices carry more risk than benefit, especially for a small business that doesn't have the room to absorb a penalty:

- Buying links in bulk through automated platforms, often detectable and penalized by search engines.
- Systematic reciprocal link exchanges ("I'll link to you if you link to me"), which stop looking natural beyond a small volume.
- Networks of sites created solely to link to each other.

These practices target traditional SEO, but they bring nothing for GEO: an artificial link generates neither real reputation nor a credible brand mention.

## A step-by-step method for a small business

Without a dedicated team, it's better to move forward in simple, repeatable steps:

1. List partners, suppliers, and clients with whom a cross-mention would make sense.
2. Identify three to five local or industry media outlets relevant to your business.
3. Check your presence on the two or three directories most relevant to your industry.
4. Regularly publish useful content that's naturally citable (guide, first-hand account, case study).
5. Track results over time with tools that identify who's already citing your brand, as covered in the lesson on [analyzing your citation sources](/en/academy/avec-vurto/analyser-ses-sources-de-citations).

Link building for a small business is about consistency more than volume: a few relevant links earned each quarter are worth more than a one-off, intense campaign that leaves no lasting trace.

### [The off-site signals that matter](https://vurto.ai/en/academy/strategie-seo-geo/signaux-off-site-qui-comptent) (2026-07-08)

Off-site signals cover everything said and built about your brand outside your own site: customer reviews, forum discussions, press articles, social media, directories. For GEO, these signals often carry more weight than your site's own content, because generative AIs rely on a range of external sources to assess a brand's legitimacy.

## Why off-site signals matter even more for GEO

In traditional SEO, off-site signals (links, reviews, mentions) complement on-site work that's already decisive: the quality of your pages, their structure, their keywords. In GEO, this balance shifts. Language models don't just read your site: they also rely on what other sources say about you, because a brand mentioned by independent third parties is perceived as more trustworthy than a brand that only talks about itself. This is the whole point of [brand mentions](/en/academy/comprendre-le-geo/brand-mentions-facteur-numero-un), considered the single most decisive factor in AI answer visibility.

In practice, this means a company with a technically flawless website but absent from reviews, forums, and the press will struggle more to appear in generated answers than a company with a more modest site that's widely cited elsewhere.

## Customer reviews: volume, recency, content

Customer reviews (Google, industry platforms, comparison sites) are one of the most visible off-site signals and among the most frequently used by generative AIs, particularly for comparative questions ("what's the best X for Y?").

Three dimensions matter in particular:

- Volume: too few reviews make a signal insignificant, regardless of quality.
- Recency: reviews that accumulate steadily over time are perceived as more reliable than an old spike followed by silence.
- Content: detailed reviews mentioning specific use cases or concrete features are more useful for an engine trying to understand what you offer than a simple numeric rating.

Responding to reviews, including negative ones, is also a signal of seriousness that can be picked up in certain contexts.

## Forums and specialized communities

Reddit, professional forums, and specialized communities (Discord, publicly indexed private groups, trade forums) occupy a growing place among the sources cited by generative engines, particularly Perplexity and ChatGPT's search features. These platforms are perceived as sources of "unsponsored" opinion, which makes them especially credible in the eyes of AIs looking to answer a comparative question or make a recommendation.

Being mentioned positively in a relevant Reddit thread, or having your brand recommended by a user in a specialized discussion, carries a weight few other channels match — precisely because this kind of mention can't be bought directly. That doesn't mean posting under disguise yourself: communities quickly spot self-promotion, which can hurt a brand's image more than help it. The realistic strategy is to produce a product or service satisfying enough to be mentioned spontaneously, and to participate honestly in discussions when relevant.

## Press and specialized media

A press article, even a modest one, remains a highly credible source for a generative engine: media outlets are generally well indexed, regularly updated, and considered relatively reliable. An article in your industry's trade press, in particular, can carry more weight than a broader mainstream piece, because it directly targets the vocabulary and questions users ask AIs about your market.

As with link building, press access for a small business mostly comes through local and trade media relations — a topic covered in the lesson on [link building for a small business](/en/academy/strategie-seo-geo/netlinking-pour-une-tpe-pme).

## Social media: an indirect but real signal

Social media isn't always directly crawled or cited by generative engines, depending on the platform and the constantly shifting landscape. Its influence remains real, though, in an indirect way: regular social activity generates brand mentions elsewhere (shared by third parties, screenshots reused on other sites, discussions triggered), which in turn feeds the signals described above.

So the direct effect of social media on GEO shouldn't be overestimated, but it also shouldn't be dismissed as a trigger for other, more directly used signals.

## Directories: a reduced but non-zero role

Generic directories have lost much of their SEO weight in recent years, and their role in GEO remains limited. They still hold real value in two specific cases: consistency of business information (name, address, phone — a central point in [local SEO](/en/academy/strategie-seo-geo/seo-local-fondamentaux)) and presence in recognized industry directories, which remain consulted by some real-time search tools.

## How to prioritize these signals with limited resources

With limited time, the realistic priority order for most small businesses is:

1. Customer reviews, especially on Google and platforms specific to your industry.
2. Local and trade media relations.
3. Organic — unsolicited — presence on forums and communities relevant to your market.
4. Consistency of information across the most consulted directories in your industry.
5. Regular social activity, as support rather than the main driver.

To find out which of these signals actually work in your favor, the best method remains observing what AIs already cite when they talk about you, as detailed in the lesson on [analyzing your citation sources](/en/academy/avec-vurto/analyser-ses-sources-de-citations).

### [Local SEO: the fundamentals](https://vurto.ai/en/academy/strategie-seo-geo/seo-local-fondamentaux) (2026-07-09)

Local SEO covers the practices that let a business appear in search results and on maps for geolocated queries. It's an essential foundation before working on local GEO, since generative AIs rely heavily on the same structured data and review ecosystem as traditional search engines.

## What local SEO covers

Local SEO aims to appear in results tied to a specific geographic area: a search like "plumber in Lyon" or "italian restaurant near me" triggers specific results (the local pack, maps) distinct from standard organic results. It primarily concerns businesses with a physical presence or a defined service area: shops, tradespeople, healthcare professionals, restaurants, agencies.

Local SEO rests on three main pillars: the business listing (Google Business Profile above all), the consistency of business information across the web, and customer reviews. These three pillars also form the base that local GEO builds on, covered in the dedicated lesson on [getting recommended by AI near you](/en/academy/strategie-seo-geo/geo-local-etre-recommande-pres-de-chez-vous).

## Google Business Profile: the foundation of everything

The Google Business Profile listing (formerly Google My Business) is the reference point for any business with a local presence. It centralizes the information shown on Google Maps and in the local pack of search results, and serves as a source for many other services that sync their local data.

The elements to prioritize:

- The primary and secondary categories, chosen precisely rather than by default.
- Hours, kept up to date including for holidays and one-off exceptions.
- Photos, regularly refreshed, showing the premises, the team, or the products.
- The business description, written without keyword stuffing but covering the terms your customers actually use.
- Q&A and posts, often neglected but still visible and consulted.

An incomplete or poorly maintained listing isn't just a missed opportunity: it can also get filled with incorrect information from other sources, which can be hard to correct afterward.

## NAP consistency (name, address, phone)

NAP stands for "Name, Address, Phone". The principle is simple: these three pieces of information must be identical, down to the character, across every platform where your business appears — website, Google listing, directories, social media, press. An address written differently from one place to another ("Street" vs. "St." for example), or an old phone number still showing on a directory, blurs the signal sent to engines and can slow down recognition of your business as a reliable, unique entity.

A periodic audit of NAP mentions is recommended, especially after a move or a phone number change: outdated details tend to linger for a long time on third-party sources if not actively corrected.

## Reviews: the most visible signal

Customer reviews, along with the business listing, are the most decisive signal in local SEO. Their number, average rating, and recency directly influence ranking in the local pack. Responding to reviews — both positive and negative — is also factored in and shows active management of your online reputation.

This signal extends well beyond traditional local SEO: it's also one of the [off-site signals that matter](/en/academy/strategie-seo-geo/signaux-off-site-qui-comptent) for visibility in generative AI answers, which rely on the same review platforms to assess a business's reputation.

## Local pages: when and how to create them

For a business operating across several geographic areas (multiple branches, multiple cities served), creating a dedicated page per area can make sense — provided each page includes content genuinely specific to that location (address, local team, local market specifics, reviews from customers in that area) rather than a mere duplicate of generic text with the city name swapped.

A local page without differentiating content adds little value and can even hurt, by giving the impression of a site built purely to multiply indexed pages. Conversely, a page that genuinely answers the questions of a customer in that area (precise service area, turnaround times, dedicated team) has real value, for visitors and engines alike.

## The link between local SEO and local GEO

Generative AIs answering local questions ("what's the best X near me?") rely heavily on the same sources as local SEO: business listings, reviews, information consistency, and increasingly on partnerships between AI platforms and mapping or review services. A business absent or poorly filled in on these basic sources starts at a disadvantage that's hard to overcome, regardless of the quality of its web content.

That's why local SEO isn't a step to skip before working on local GEO: it's the foundation any AI recommendation is built on. The levers specific to local GEO — how AIs handle proximity queries, what actually influences the recommendation — are covered in the next lesson on [local GEO](/en/academy/strategie-seo-geo/geo-local-etre-recommande-pres-de-chez-vous). Local link building, through partnerships and community press, also completes this foundation — a topic explored further in the lesson on [link building for a small business](/en/academy/strategie-seo-geo/netlinking-pour-une-tpe-pme).

### [Local GEO: getting recommended by AI near you](https://vurto.ai/en/academy/strategie-seo-geo/geo-local-etre-recommande-pres-de-chez-vous) (2026-07-10)

When a user asks a generative AI "what's the best Italian restaurant near me?", the answer doesn't come from a simple ranking of search results but from a combination of local sources, reviews, and sometimes the user's stated location. Understanding this mechanism helps identify the levers a local business can actually act on.

## How AIs handle a local query

A local query usually contains an explicit signal ("near me", "in Lyon", "in the neighborhood") or an implicit one (the user mentions their city earlier in the conversation). Generative engines with access to real-time search — increasingly the case for ChatGPT, Gemini, and Perplexity — handle the query differently from a general question: they trigger a targeted search of local sources rather than relying solely on the model's training data.

This changes things for local businesses: unlike a general question about a national market, where overall brand awareness carries a lot of weight, a local query places more value on genuine, recent presence on geolocated sources — business listings, reviews, community press — than on the company's size or age alone.

## The role of declared or inferred location

A user's location can be known in several ways: device geolocation (on a mobile app), a city explicitly mentioned in the conversation, or inferred from context (language, time zone, conversation history). The precision of this location varies significantly by platform and by the user's privacy settings, which means there's no universal method guaranteeing an AI will correctly place a business within its trade area.

What can be said with more confidence is that the more consistent and explicit a business's location information is across its various online sources, the more likely it is to be correctly associated with a geographic area by the systems powering AIs — a point that ties directly back to the fundamentals of [local SEO](/en/academy/strategie-seo-geo/seo-local-fondamentaux).

## The sources used for a local recommendation

To answer a local query, generative AIs generally combine several types of sources: standard geolocated search results, review platforms, sometimes mapping services through partnerships, and the indexed content of local sites themselves (regional press, local blogs, neighborhood forums). The exact weighting between these sources isn't public and varies from one engine to another, but their combined presence remains the most observable common factor.

A business well rated across several review platforms, cited by local press, and present in local community discussions (neighborhood groups, city forums) thus accumulates several independent signals that converge on the same recommendation — which ties back to the central role of [brand mentions](/en/academy/comprendre-le-geo/brand-mentions-facteur-numero-un) in AI visibility generally.

## What actually influences the recommendation

Based on the sources involved, several elements stand out as concrete levers:

- The consistency and completeness of business information across the main platforms.
- The volume and recency of reviews, especially those mentioning concrete details (type of service, neighborhood, specifics).
- Mentions in local press or neighborhood publications, which provide third-party validation that's hard to obtain otherwise.
- Presence in local community discussions, on forums or social media with local reach, which is a particularly relevant [off-site signal](/en/academy/strategie-seo-geo/signaux-off-site-qui-comptent) at the scale of a neighborhood or city.

Conversely, a polished website isolated from any local ecosystem — no reviews, no external mentions, no presence in local press or directories — has little chance of being recommended, even if its content is high quality.

## Current limits of local GEO

Local GEO remains a fast-evolving field, and certain limits should be kept in mind. The availability and reliability of real-time search vary by engine and change regularly with product updates. Geographic coverage isn't uniform: large cities generally benefit from a richer ecosystem of sources (reviews, press, forums) than rural areas or small towns, which can mechanically disadvantage businesses located there, regardless of the quality of their work.

There's also no guarantee an AI will systematically cite several options for a local query: some answers are limited to one or two recommendations, which makes competition for those top spots particularly fierce.

## Concrete actions for a local business

1. Check and complete your Google Business Profile listing and other relevant business listings for your industry.
2. Encourage detailed, recent reviews by making the process easy for satisfied customers.
3. Look for opportunities to be mentioned in local press and neighborhood publications.
4. Identify local forums and groups relevant to your business and participate honestly when useful.
5. Regularly track what AIs say about you on typical local queries, to identify which sources they actually cite — an approach detailed in the lesson on [analyzing your citation sources](/en/academy/avec-vurto/analyser-ses-sources-de-citations).

Local GEO rewards the density and consistency of a business's presence in the local ecosystem more than its size — a field where a well-organized small business can compete with much larger players.

### [Multimodal content: images, video, audio](https://vurto.ai/en/academy/strategie-seo-geo/contenu-multimodal) (2026-07-11)

Multimodal content refers to everything on a site that isn't plain text: images, videos, podcasts, infographics. For a generative AI to make use of this content, it needs to come with a sufficiently precise textual description — without one, it remains largely invisible to an engine.

## Why non-textual content remains largely invisible to AI

The language models powering generative AIs are, at their core, trained on text. Some recent models include image understanding or automatic description capabilities, but this capability remains uneven across models and contexts, and it doesn't replace a clear, explicit accompanying text. For a traditional search engine as much as a generative one, an image or video without a text description remains, to a large extent, content it can neither index precisely nor cite reliably.

This creates a frequent blind spot: a company might invest in a polished product demo video, or an infographic that perfectly summarizes its expertise, without that content ever being picked up in a generated answer — simply because no accompanying text lets an engine understand what it contains.

## Alt text: the foundation for images

Alt text is the textual description attached to an image in a page's code. Originally designed for accessibility (screen readers for visually impaired users), it has also become an essential signal for search engines and, by extension, for GEO.

A few simple principles:

- Describe what the image actually shows, factually, without keyword stuffing.
- Adjust the level of detail to the image's importance: a decorative photo can stay minimal, a technical diagram or product shot deserves a fuller description.
- Avoid generic text ("image1.jpg", "product photo") that provides no usable information.
- For purely decorative images, an empty alt text is preferable to an artificial one that adds nothing.

Well-written alt text lets an engine understand that an image illustrates a specific point on the page, and therefore consider it part of coherent content — a criterion tied to the principles of [GEO-ready content](/en/academy/comprendre-le-geo/contenu-geo-ready).

## Captions: valuable additional context

Unlike alt text, invisible to a human visitor, a caption is displayed below or next to the image. It provides additional context — a photo's source, a figure tied to a chart, an explanation of a diagram — which helps both visitors and engines understand visual content without needing to interpret the image itself.

Captions are especially useful for charts and infographics containing data: restating the key figures shown visually in a caption, or ideally in the surrounding text, lets an engine pick them up correctly, whereas data present only within an image remains very hard to extract reliably.

## Transcriptions for video and audio

For video and audio (podcasts, webinars, interviews), a full text transcript remains the most reliable way to make this content usable by an engine, generative or otherwise. A well-structured transcript also makes it possible to locate and cite a specific passage, which ties into the chunking logic used by many search systems (see the [glossary](/en/academy/strategie-seo-geo/glossaire-geo-aeo-ia-generative)).

A few best practices:

- Publish the full transcript on the same page as the video or audio, rather than on a separate, hard-to-link page.
- Structure the transcript with subheadings or timestamps for longer content, to make it easier to locate a specific passage.
- For subtitles embedded in the video, check their accuracy: an unreviewed automatic transcript often contains errors that can distort the meaning of a cited passage.
- Add a summary at the start of the transcript for longer content, useful both for time-pressed visitors and to give quick access to the main topic.

## Structured data applied to multimodal content

Beyond accompanying text, specific structured data types exist for images and videos ("ImageObject" and "VideoObject" in schema.org), which let you specify metadata such as a video's duration, publication date, or description. This technical markup doesn't replace good text description, but reinforces it by making it readable in a structured way by the systems that analyze it automatically.

## A practical method to catch up

Few sites handle all their multimodal content correctly from the start. Here's a step-by-step method to close existing gaps:

1. Inventory your most strategic non-textual content: product demo videos, infographics with key data, high-value podcasts or webinars.
2. Prioritize checking the alt text of images on your most visited or commercially important pages.
3. Add transcripts to videos and audio content that don't have them yet, starting with those containing the most useful information.
4. Restate in text form the key figures and information that currently exist only in visual form.
5. Track whether this content starts getting picked up in generated answers, notably through [analyzing your citation sources](/en/academy/avec-vurto/analyser-ses-sources-de-citations).

Multimodal content remains an often-overlooked axis of GEO, even though it frequently represents a site's richest and most differentiated content. Making it readable for engines, beyond its visual or audio quality, is a step not to skip.

### [Glossary: every GEO and generative AI term](https://vurto.ai/en/academy/strategie-seo-geo/glossaire-geo-aeo-ia-generative) (2026-07-12)

## GEO (Generative Engine Optimization)
The set of practices aimed at optimizing content and digital presence to be cited as a source by generative AI engines (ChatGPT, Claude, Gemini, Perplexity) when they build an answer.

## AEO (Answer Engine Optimization)
Optimization aimed at becoming THE direct answer to a specific question, whether in an AI chatbot or in a traditional search engine's enriched results (AI Overviews, voice assistants).

## AI visibility
The fact, for a brand, product or piece of content, of being mentioned in the answers generated by a generative AI — the equivalent of "ranking" in SEO, but at the scale of a single answer rather than a results page.

## AI answer / generative answer
The text produced by an LLM in response to a question, built from the data it was trained on and, in some cases, real-time search.

## LLM (Large Language Model)
A large language model, like the ones powering ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google) or Perplexity.

## Visibility score
A summary indicator that reflects a brand's, category's or keyword's presence across all tested prompts and engines.

## Share of voice
The proportion of mentions a brand gets relative to its competitors, across the same set of tested questions.

## Prompt
A question or query sent to an LLM to test whether it mentions a given brand, category, keyword or product.

## Persona
A fictional user profile (age, location, language, interests, buying behavior) used to simulate different types of questions and detect visibility gaps across audiences.

## Semantic scoring
The evaluation of how "AI-friendly" a product page or piece of content is, based on criteria like structure, structured data, the presence of an FAQ, or out-of-context readability.

## Structured data (schema.org / JSON-LD)
Technical markup added to a web page to explicitly describe its content (product, price, reviews, FAQ...), making it easier for search engines and AIs to understand.

## Detailed analysis
An in-depth report generated from visibility results: a summary, dominant and absent players, competitive comparison, risks and opportunities, prioritized recommendations.

## Credits
The unit of consumption for analyses on Vurto; each analysis consumes a variable number of credits depending on its type and how many engines are queried.

## RAG (Retrieval-Augmented Generation)
Retrieval-augmented generation is a technique that lets an LLM fetch information from a database or the web at the moment it answers, rather than relying solely on its training data. This is the mechanism that lets tools like Perplexity, or ChatGPT's search features, cite recent sources.

## Fan-out query
A technique by which a generative engine breaks a complex question into several sub-queries sent in parallel, then synthesizes the results into a single answer. A single question from a user can therefore trigger several distinct searches behind the scenes.

## Reranking / RRF (Reciprocal Rank Fusion)
Reranking refers to reordering a set of already-retrieved results based on their actual relevance to the question asked. RRF (Reciprocal Rank Fusion) is a common method for merging several result rankings from different sources or search methods into a single final ranking.

## Chunking
Splitting a long document into smaller fragments ("chunks") processed and indexed independently. This technique, central to RAG systems, explains why a well-structured page with self-contained sections is more likely to be used correctly than a long text without clear breaks.

## E-E-A-T
Stands for Experience, Expertise, Authoritativeness, Trustworthiness — a set of criteria Google uses to assess content quality, notably through its quality rater guidelines. While not a direct, measurable ranking factor, these criteria influence how trustworthy content is judged to be, by both search engines and generative AIs.

## Essential HTML tags (Title, Meta description, Hn, Canonical, Hreflang, Open Graph)
A set of technical elements placed in a page's code to describe its content to engines and social networks: the Title (displayed in results), the Meta description (summary shown under the title), Hn tags (heading hierarchy, H1 to H6), the Canonical tag (indicating the reference version of a duplicated page), the Hreflang tag (signaling the language versions of the same page), and Open Graph tags (controlling how a page's preview looks when shared on social media).

## Core Web Vitals
A set of metrics defined by Google to measure a page's user experience: how fast the main content loads, how responsive it is to interactions, and how visually stable it is while loading. These metrics are used as a ranking factor in SEO and indirectly contribute to how well a site can be crawled.

## robots.txt
A file placed at the root of a site that tells bots (search engines, AI crawlers) which parts of the site they are and aren't allowed to browse. It doesn't guarantee content will never be used — a bad-faith bot can ignore it — but it remains the standard way to communicate crawling preferences.

## llms.txt
A file proposed as an emerging standard, placed at the root of a site, meant to give language models a structured, condensed overview of the site's content — somewhat like a sitemap designed specifically for AIs rather than traditional search engines. Its adoption remains uneven across engines as of today.

## XML Sitemap
A file listing a site's pages intended for indexing, along with additional information such as their last modification date. It makes pages easier for bots to discover, especially on large sites or ones with a complex internal linking structure.

## AI crawlers (GPTBot, ClaudeBot, PerplexityBot...)
The bots used by generative AI companies to crawl the web, either to train their models or to power their real-time search features. GPTBot (OpenAI), ClaudeBot (Anthropic), and PerplexityBot (Perplexity) are examples; their activity can be tracked in a site's server logs, and allowed or blocked via the robots.txt file.

## AI Overviews
AI-generated summaries that appear at the top of certain Google search results, ahead of standard organic results. They represent a form of AEO built directly into traditional search, and can reduce traffic to cited sites if the user finds their answer without clicking through.

## Hub & Spoke architecture
A site organization model in which a pillar page ("hub") centralizes an overview of a topic, linked to several more specific pages ("spokes") that each dig into a sub-topic. This structure helps engines understand the hierarchy and relationships between content on the same subject.

## Pillar page / spoke page
The pillar page is the reference content on a broad topic, designed to cover the essentials and link out to more detailed content. A spoke page covers a specific sub-topic in depth and links back to its pillar page for broader context — together, they form the hub & spoke architecture.

## Entity (SEO)
A person, place, organization, product, or concept identified as a unique thing by search engines, independent of the exact words used to describe it. Working on entity recognition means helping engines clearly associate a brand with what it represents, beyond just the text on its pages.

## Search intent
The actual motivation behind a query: looking for information, comparing options, making a purchase, or finding a specific site. Identifying the intent behind a question is a prerequisite for creating relevant content, whether for SEO or to be picked up in a generative answer.

## Link building
The set of practices aimed at earning external links pointing to your site from other domains. Historically central to SEO as an authority signal, it retains a role in GEO insofar as it contributes to a brand's overall reputation and the mentions that stem from it.

## Internal linking
The set of links connecting a site's pages to one another. Consistent internal linking makes crawling easier, distributes authority across pages, and helps engines understand how a site's content relates thematically.

## Local SEO
The set of practices aimed at improving a business's visibility for geolocated searches, relying notably on the Google Business Profile listing, consistent business information (NAP), and customer reviews.

## Multimodal content
Content that isn't plain text — images, video, audio, infographics. For it to be usable by engines and generative AIs, it needs to come with descriptive text (alt text, captions, transcripts).

## MCP (Model Context Protocol)
A standardized protocol that lets an AI assistant connect to external tools and data sources in a structured way. At Vurto, it's offered on the Business and Agency plans, letting third-party systems query visibility data directly.

To see these concepts in practice, head to [prompts & visibility tracking](/en/academy/avec-vurto/prompts-personas-suivi-visibilite-ia) and [dashboards & scoring](/en/academy/avec-vurto/lire-dashboards-vurto-scoring-semantique).

### Scaling up with Vurto

### [Getting started with Vurto in 5 steps](https://vurto.ai/en/academy/avec-vurto/bien-demarrer-avec-vurto) (2026-07-13)

Vurto measures how ChatGPT, Claude, Gemini and Perplexity talk about your brand, your products and your categories — and helps you improve that presence. Here's how to get started.

## 1. Tell us who you are

From the welcome screen, choose your profile: agency, e-commerce or brand. This choice shapes the questions asked next and the vocabulary used across the app (single client vs. multi-client, product catalog vs. services, etc.).

## 2. Tell us about your business

Enter your company name, your website URL and your country. That's the only manual input needed to get started — everything else is inferred automatically.

## 3. Let the AI analyze your site

Vurto analyzes your URL and automatically detects your industry, your brands, your product or service categories, and relevant keywords. You can adjust this selection: up to 3 keywords, 3 categories and 1 to 3 brands depending on your profile.

## 4. Pick your test prompts

Based on these elements, Vurto generates around ten prompts — the kind of questions real users would ask a generative AI ("what's the best brand for...", "what do you recommend for...", etc.). Select the ones that best reflect the questions your customers might ask (at least one prompt is required).

## 5. Run your first analysis

Once your account is created, your first project and active prompts are set up automatically. You're redirected to your dashboard, where the first visibility results appear: by brand, by category and by keyword, engine by engine (ChatGPT, Claude, Gemini, Perplexity).

## What's next?

Once this foundation is in place, explore [dashboards and semantic scoring](/en/academy/avec-vurto/lire-dashboards-vurto-scoring-semantique) to interpret your results, and the [prompts & tracking](/en/academy/avec-vurto/prompts-personas-suivi-visibilite-ia) section to refine your testing strategy over time.

### [Reading your AI visibility dashboard](https://vurto.ai/en/academy/avec-vurto/lire-dashboards-vurto-scoring-semantique) (2026-07-14)

Whether you're tracking a brand, a category or a keyword, Vurto displays the same types of indicators on every dashboard: an average visibility score, its breakdown by AI engine, its evolution over time, and the competitors identified in the answers. On top of that, your product pages get an independent semantic score evaluating how well they can be understood and reused by generative AIs.

## The indicators common to all three dashboards

The brand, category and keyword dashboards share the same structure. At the top, the number of tracked elements (active brands, categories or keywords) and the average visibility score, which sums up in one number how often and how well you appear in AI answers across all your prompts. That score then breaks down by engine — ChatGPT, Claude, Gemini, Perplexity, up to eight models depending on your plan — revealing gaps the overall average hides: it's common to be well positioned on one engine and nearly absent on another.

## Reading the trend curve and the engine breakdown

The trend curve shows how your score moves over time, wave after wave. It's this curve, more than the instant score value, that should guide your decisions: a score of 40 trending steadily upward is better news than a score of 60 that's been dropping for three waves. When a score drops on one engine in particular, don't look first for a technical explanation on your side — start by checking whether a competitor gained visibility at the same time on that same engine. For example, if your Perplexity score drops from 55 to 38 over two waves while staying stable on ChatGPT and Claude, chances are a competitor got better cited on the sources Perplexity favors for your sector — the competitor table on the same dashboard lets you verify this directly rather than assume it.

## The competitor table

Below the score indicators, Vurto shows the competitors identified in the AI answers linked to your prompts, with, for each one, the persona active when they were mentioned and the sources the AI cited to recommend them. This table gives directly readable context to any score change: a drop can almost always be explained by one or more specific competitors gaining ground, identifiable row by row rather than as a diffuse phenomenon.

## The detailed analysis

On results that warrant it, Vurto offers a detailed analysis that goes beyond raw numbers: an executive summary of the situation, the dominant and absent players on the tracked topic, a structured competitive comparison, an identification of risks and opportunities, and recommendations organized by time horizon — immediate actions, medium-term actions, foundational actions. This is the level of reading to favor before deciding where to invest time, rather than reacting to a single score change.

## Semantic scoring of your product pages

Independently of the visibility dashboards, Vurto evaluates each product page on ten criteria: title, description, structured data, FAQ, technical specifications, customer reviews, multimodality (photos, videos), alt text, out-of-context readability, and hierarchical structure. Each page gets an overall score out of 100, along with detailed recommendations per criterion. Out-of-context readability deserves particular attention: a generative AI usually only shows an excerpt of your page, without the rest of the site around it — a page that assumes context it doesn't provide (a vague page title completed elsewhere, for instance) loses clarity once isolated, even if it reads well in its original setting. A page scoring 45/100 with FAQ and structured data both at 0 doesn't call for the same action plan as a page also scoring 45/100 because of missing customer reviews and uncaptioned visuals: the overall score guides priority across pages, the per-criterion detail guides the work within a single page.

## Cross-referencing the visibility dashboard with competitor analysis

These two readings complement each other and shouldn't be done separately. A visibility score that stagnates despite a good semantic score generally points to an external issue: the [sources cited by AIs](/en/academy/avec-vurto/analyser-ses-sources-de-citations) in your sector don't yet mention your brand, even though your own content is solid. Conversely, a low score paired with a low semantic score points first to a content issue to fix before even looking for external explanations. The [detailed competitor analysis](/en/academy/avec-vurto/analyser-ses-concurrents-dans-les-reponses-ia) lets you settle between these two hypotheses instead of guessing.

## How often to check these dashboards

Visibility scores vary from one analysis wave to the next, sometimes for one-off reasons (a slightly different phrasing in an AI answer, momentary news). Checking the dashboard after every single wave without stepping back exposes you to reacting to noise rather than to a real trend. A weekly review, focused on the trend curve rather than the latest isolated point, gives a more reliable basis for decisions than a daily check.

## Reading the three dashboards together rather than in isolation

The brand dashboard, the category dashboard and the keyword dashboard each answer a different question, and reading them in isolation can hide important situations. The brand dashboard shows whether your company itself is recognized and recommended as such. The category dashboard shows your position on a type of product or service, regardless of whether your brand gets named. The keyword dashboard narrows this further, down to specific queries. A brand that's rarely named directly can still be well represented at the category level if its products keep coming up without the AI systematically naming the brand — which points toward strengthening brand awareness rather than product content. Conversely, a well-recognized brand that's absent on specific keywords points to a more targeted issue, often tied to specific pages that aren't sufficiently structured rather than an overall awareness gap.

## A 3-point prioritization method

Facing several tracked brands, categories or keywords, prioritize in this order. First, scores dropping continuously over several waves, which signal active deterioration rather than a simple one-off variation. Second, high-commercial-stakes product pages with the lowest semantic scores, since these are the fastest fixes to implement and the most direct payoff. Third, prompts where the detailed analysis identifies a clear opportunity — a topic no player yet dominates in AI answers — because that's where effort produces the best ratio between investment and visibility gain.

### [Prompts and personas: building reliable visibility tracking](https://vurto.ai/en/academy/avec-vurto/prompts-personas-suivi-visibilite-ia) (2026-07-15)

A tracked prompt is a template question, phrased the way a real user would ask a generative AI, that Vurto regularly sends to ChatGPT, Claude, Gemini and Perplexity to measure how your brand, your products and your categories show up in the answers. The quality of your tracking depends directly on the quality and diversity of these prompts.

## The three sources of your prompts

Your prompts come from three sources. System prompts are suggested by Vurto based on common question patterns in your sector — comparisons, recommendations, best value for money. Onboarding prompts are the ones generated automatically during the initial analysis of your site, based on the detected industry, brands, categories and keywords. Custom prompts are the ones you write yourself, to cover a question specific to your business that no system prompt addresses — a particular warranty, a specific use case, a comparison naming a competitor. Solid tracking combines all three: system and onboarding prompts give a broad comparison base consistent with your sector, custom prompts cover your specific blind spots.

## Dynamic variables

To avoid manually duplicating the same prompt for every tracked brand, product or keyword, Vurto uses variables: {{marque}}, {{mot_cle}}, {{categorie}}, {{nom_du_produit}}. A prompt written once with these variables automatically declines across all matching tracked elements, with no rewriting needed. This is what keeps tracking consistent even with a large catalog, instead of managing dozens of manual variants of the same prompt.

## Organizing by intent

Vurto classifies prompts by intent: brand, category, keywords, products, price, competition, recommendations. This isn't just filing — it shapes what you can later read in your dashboards. An imbalance between intents skews the reading: if nearly all your prompts are about the brand and none about price or competition, your overall visibility score will reflect your brand awareness well but say nothing about where you stand in comparisons, even though that's often where purchase decisions are made. Regularly check that all seven intents are represented in reasonable proportion relative to your customers' actual questions.

## The role of personas

A persona is a fictional user profile — for example a professional buyer focused on technical specifications, or an individual primarily price-sensitive — used to simulate different audiences facing the same topic. This diversity of viewpoints ties back to the goal of covering the widest possible [AI Share of Voice](/en/academy/comprendre-le-geo/objectif-etre-partout-share-of-voice): being visible to a single type of audience isn't enough if your real customers span several different profiles. The same prompt asked from two different personas' viewpoints can produce very different AI answers, with the competitors and sources shifting depending on the angle taken. Without personas, tracking only captures a single average viewpoint, which can hide very unequal positions depending on the actual target audience.

## Example: a prompt template and its persona variations

A prompt template like "What's the best {{categorie}} brand for {{mot_cle}}?" automatically declines across every tracked category and keyword. Paired with a "professional buyer" persona, this same prompt can lean toward reliability and technical compliance; paired with a "first-time individual buyer" persona, it leans more toward ease of use and price. Across both personas, {{categorie}} and {{mot_cle}} stay identical, but the AI's answers — and therefore the competitors and sources that surface — can diverge sharply. That divergence is what makes the prompt × persona cross-analysis useful: it reveals competitive positions invisible in an average reading.

## How many prompts and personas to track

There's no universal ideal number, but a few practical benchmarks help. Start by covering each intent with at least one system or onboarding prompt, add custom prompts only for topics where your customers actually ask specific questions, and limit active personas to those matching real, distinct customer segments for your business — multiplying personas without a real difference in buying behavior dilutes the reading without adding information. The number of tracked prompts is also bound by your plan (500 tracked prompts on Standard, up to unlimited prompts on Business), which should factor into your decisions too.

## Managing your prompts over time

As your prompt library grows, managing prompts one by one becomes limiting. Vurto lets you export your full set of prompts as CSV to review offline, have another team validate them, or prepare a new wave in bulk, then re-import them. This is especially useful for periodic review: rather than editing each prompt one at a time in the interface, export them, flag the ones that have become outdated or redundant, adjust under-represented intents, and re-import the updated list.

## Avoiding common writing pitfalls

A poorly worded system or custom prompt skews the reading before the first analysis even runs. A prompt too close to an SEO keyword ("best cheap {{categorie}} brand") doesn't match how a user actually talks to a conversational AI, and produces less representative answers than a real question would. A prompt that's too long, or that stacks several questions at once (brand, price and delivery in the same sentence), dilutes the answer and makes the result harder to interpret. Best practice is to keep each prompt centered on a single clear intent, phrased as a real question — creating several prompts if needed, rather than one overloaded one.

## Evolving your tracking instead of freezing it

A prompt set is never final. A prompt that no longer produces useful information (because the AI answer has become stable and predictable) can be replaced with a prompt on an intent that's still under-covered. Conversely, a topic that becomes strategic — a new competitor, a new product line — deserves dedicated custom prompts. This review work naturally connects with reading your [visibility dashboards](/en/academy/avec-vurto/lire-dashboards-vurto-scoring-semantique) and extends into [analyzing your competitors in AI answers](/en/academy/avec-vurto/analyser-ses-concurrents-dans-les-reponses-ia) to understand why a given prompt tilts in their favor: the results you get are what tell you where your tracking needs more depth.

### [Analyzing competitors in AI answers](https://vurto.ai/en/academy/avec-vurto/analyser-ses-concurrents-dans-les-reponses-ia) (2026-07-16)

Vurto automatically identifies the competitors that show up in generative AI answers to your tracked prompts, and shows you which prompts they appear on and which sources the AIs cite to recommend them. Here's how to read that table and turn it into concrete actions.

## What the competitor table contains

On each dashboard (brands, categories, keywords), Vurto displays a table of competitors identified in the AI answers linked to your prompts. For each competitor, you see the persona that was active when they were mentioned, along with the sources the AI cited to justify recommending them. This table isn't static: it evolves with every new analysis wave, at the same pace as the engines update their answers.

## Spotting the competitors that keep coming back

The first useful read is to sort by frequency of appearance. A competitor cited on a single isolated prompt is a weak signal; a competitor who keeps showing up across several prompts, several engines and several personas is structurally better positioned than you on that segment. This reading connects to the idea of [AI Share of Voice](/en/academy/comprendre-le-geo/objectif-etre-partout-share-of-voice): being visible across as many prompts, engines and personas as possible, rather than on a single isolated point. Also distinguish cross-cutting competitors (visible across your whole category) from niche competitors (visible only on one keyword or one persona): the two call for different treatment — the first for a structural response, the second for a targeted one.

## Understanding why they get cited

The table doesn't just name a competitor: it shows the sources the AI relied on to recommend them. That's the most actionable piece of information. If a competitor keeps coming back because a third-party comparison, a buying guide or a trade press article cites them, you know the issue isn't only on their own site — it's their presence in the external sources the AI consults. If, on the other hand, a competitor shows up through their own content (product page, category page, FAQ), that's a sign their content structure is better exploited by generative engines than yours.

## Cross-referencing competitors and personas

The same competitor can dominate on one persona and be absent on another. A "demanding professional buyer focused on technical specs" persona won't necessarily cite the same players as an "individual shopper looking for the best value" persona. Reading the competitor table persona by persona avoids drawing overly general conclusions: you can perfectly well dominate one segment while being invisible on another, without the overall average revealing it.

## Building an action plan

Once competitors and their sources are identified, the work breaks down into three parts. First, on prompts where a competitor dominates because of third-party content, identify those sources and assess how to get a presence there — the natural continuation of [analyzing your citation sources](/en/academy/avec-vurto/analyser-ses-sources-de-citations). Second, on prompts where a competitor dominates through their own content, compare the structure of their product pages to yours: title, description, structured data, FAQ, customer reviews — the same criteria evaluated by Vurto's semantic scoring on your own product pages. Third, track how the table changes over time: a competitor gaining ground quickly across several prompts within a few weeks deserves closer monitoring than average.

## A reading example

Say that on the "running shoes" category, three competitors keep coming back in AI answers to your prompts, but for very different reasons. The first appears almost exclusively through an independent comparison heavily cited by all four engines: acting on that third-party source is the priority. The second appears through their own product pages, richly structured with customer reviews and detailed technical data: that's a signal to compare directly against your own semantic scoring. The third only appears on the "beginner runner" persona and never on the "experienced runner" persona: the issue isn't dislodging them everywhere, but understanding why they specifically capture that audience — for instance because their content emphasizes guidance and coaching rather than technical performance.

## Don't confuse visibility with relevance

A visible competitor isn't necessarily a direct one. Before spending time on a given player, check that they're actually a relevant competitor for your business, not a generalist brand mentioned by default for lack of anything better. This analysis should be read alongside the [visibility dashboard](/en/academy/avec-vurto/lire-dashboards-vurto-scoring-semantique): a score dropping on a given engine is often explained by a specific competitor gaining ground on that same engine — something the table lets you verify directly rather than assume.

### [Analyzing your citation sources](https://vurto.ai/en/academy/avec-vurto/analyser-ses-sources-de-citations) (2026-07-17)

Vurto analyzes the domains and pages that generative AIs cite most often when answering your tracked prompts, so you know where to focus your efforts to get a presence in those answers — on your own site or through third-party sources.

## What the sources dashboard shows

For each tracked brand, category or keyword, Vurto records the sources cited by AI engines in their answers: this can be a page on your own site, a competitor's page, an independent comparison, a buying guide, a specialized forum, or a press article. This list builds up prompt by prompt, wave after wave, and connects with the competitor table: the same source can justify the presence of several players at once.

## Your own sources versus third-party sources

The first distinction is simple but structural. A first-party source is a page on your own site cited directly by the AI: a sign that your content is judged clear and structured enough to be reused as-is. A third-party source is an external site cited instead of you, or instead of your competitors: a comparison, a review, an article. If most of the citations in your sector come from third-party sources rather than the brands' own sites, that means the AIs prefer to rely on intermediaries they consider more reliable or more synthetic than the original product pages or category pages.

## Prioritizing: frequency, consistency, accessibility

Not all sources carry equal weight. A source cited once on an isolated prompt matters less than a source that keeps coming back across several prompts and several engines — that's the one to address first. Also look at consistency over time: a source that appears at every analysis wave is structurally embedded in AI answers, unlike a one-off source. Finally, assess accessibility: a third-party source open to contributions (guest article, partner listing, participative comparison) is a faster opportunity to activate than a closed outlet.

## Two distinct levers

Once priority sources are identified, two levers emerge. On your own pages, the goal is to strengthen their structure so they become cited sources themselves: clear titles, direct answers to common questions, structured data, FAQ — the same principles behind [GEO-ready content](/en/academy/comprendre-le-geo/contenu-geo-ready). On third-party sources identified as influential, the goal is to get a presence there: press contact, partnership, contributing to a comparison, updating an existing listing. This is the same territory as the [off-site signals](/en/academy/strategie-seo-geo/signaux-off-site-qui-comptent) that affect your generative visibility, except here you know exactly which sources to target since Vurto names them for you.

## Cross-referencing sources and competitors

The sources table makes full sense when read alongside your [competitor analysis](/en/academy/avec-vurto/analyser-ses-concurrents-dans-les-reponses-ia). If a competitor dominates a prompt through a specific third-party source, that same source is probably accessible to you too — often the fastest way to close a visibility gap, faster than waiting for a redesign of your own pages to take effect.

## A concrete example

Say you sell office furniture, and the prompt "which ergonomic chair brand do you recommend for remote work" consistently cites a comparison published on a site specialized in office equipment, more than any brand's own product page. That comparison becomes priority number one: check whether it's updated regularly, whether it accepts additional brands, and whether its owner can be reached for an update or a partnership. Reaching out to that site has a good chance of moving your visibility on that prompt faster than a full rewrite of your product page — even though both actions remain complementary.

## Treating a media outlet and a forum differently

Not all third-party sources get activated the same way. A media outlet or a professional comparison site usually works on a classic editorial basis: press contact, content pitch, or in some cases a commercial partnership. A specialized forum or community (user reviews, Q&A) works differently: presence there is built through genuine participation — answering questions, providing useful and verifiable information — rather than through a classic commercial approach. Mixing up the two wastes time: pitching a partnership to a community forum, or waiting for a spontaneous reply from a media outlet that runs on press relations, rarely gets the hoped-for result.

## Tracking changes over time

The sources cited by AIs aren't fixed: a comparison can be updated, a new guide can emerge, an old article can lose relevance. Come back to this dashboard regularly to spot new sources gaining frequency and ones dropping out of answers, rather than basing your action plan on a single snapshot.

### [Generating GEO-ready content with Vurto](https://vurto.ai/en/academy/avec-vurto/generer-du-contenu-geo-ready-avec-vurto) (2026-07-18)

Vurto turns the insights from your analyses — uncovered topics, frequent persona questions, angles used by better-positioned competitors — into concrete content ideas: guides, comparisons, FAQs, or rewrites of existing product pages.

## Where the insights come from

Three sources feed this work. First, the dashboards' detailed analysis, which identifies dominant and absent players on a given prompt, along with risks and opportunities by time horizon — missing content on a topic dominated by a competitor is a directly actionable opportunity. Second, [analyzing your citation sources](/en/academy/avec-vurto/analyser-ses-sources-de-citations), which reveals the questions and formats AIs favor in your sector: if a third-party comparison keeps coming back, content of that type is probably missing from your own site. Third, the semantic scoring of your existing product pages, which points to specific criteria to fix — missing FAQ, missing structured data, underused customer reviews.

## From insight to brief

A raw insight isn't enough to produce publishable content: it needs to become a brief. A good GEO-ready brief specifies the exact question the content should answer first — phrased the way a user would ask an AI, rather than as a classic SEO keyword — the persona(s) concerned, the competitors or sources currently dominating the topic, and the elements expected by the target format (a comparison with clear criteria, a structured FAQ, technical specifications). The closer the starting question is to what a real prompt would contain, the more likely the resulting content is to get picked up by generative engines.

## The formats that best answer each type of insight

Depending on the type of insight, one format is generally more relevant than another. A topic where several players get compared in AI answers calls for a structured comparison, with explicit criteria rather than a narrative text. A topic where the same question keeps recurring across several prompts (warranty, compatibility, delivery, price) calls for a dedicated FAQ, with short, self-contained answers. A broader topic, where your brand is absent from answers despite genuine legitimacy, calls for an in-depth guide that builds your expertise over time rather than a one-off answer. For example, if the "B2B buyer" persona consistently asks about technical compatibility and those questions have no clear answer on your site, a dedicated FAQ is more likely to get picked up than a generic guide.

## Rewriting a product page from the scoring

For product pages, the logic is different: rather than creating new content, the goal is to fix the criteria the semantic scoring flagged as weak. If a page scores low on structured data and FAQ but well on the description, the rewrite priority is on the first two criteria, not the whole page. This targeted approach is faster than a full overhaul and lets you address first the pages with the lowest overall score or the most searched-for product.

## Avoiding generic content

A brief built from a Vurto insight is always more precise than a topic picked at random, but that precision can get lost if the final content stays generic in its writing — a comparison with no real numbers, an FAQ with vague answers that don't actually settle anything. The simplest test before publishing is to re-read the content and ask whether it genuinely, fully and independently answers the starting question the way a user would have asked an AI. If answering it fully requires looking elsewhere on the site, the content isn't ready yet.

## Publish, then re-check

Content generated from an insight is only useful once it actually shows up in AI answers. Once published, tracking doesn't stop: add a custom prompt covering the exact question addressed if needed, and watch, on the following analysis waves, whether the new content starts getting cited as a source or whether your visibility score improves on that prompt. If nothing moves after several waves, go back to the [semantic scoring](/en/academy/avec-vurto/lire-dashboards-vurto-scoring-semantique) to check that the published content actually meets the expected criteria — clear hierarchical structure, readability out of context, direct answers near the top of the page.

## Who writes, and based on what

Vurto turns insights into content ideas and briefs, but the actual writing stays classic editorial work: an in-house writer, a freelancer, or a marketing team. The brief produced by this analysis work exists precisely to make that writing faster and more targeted than blind topic research: instead of guessing what might interest your customers, the writer starts from a question actually asked to generative AIs, with the competitors and sources already identified as a reference for what works or is missing on the topic.

## Prioritizing when insights pile up

With a large catalog or several tracked categories, insights accumulate fast. Prioritize high-commercial-stakes prompts where one or more [competitors clearly dominate](/en/academy/avec-vurto/analyser-ses-concurrents-dans-les-reponses-ia) your own results, rather than trying to cover every identified topic at once. One piece of content well-targeted on a high-stakes prompt beats ten shallow pieces on secondary topics.

### [The API and MCP server: connecting Vurto to your tools](https://vurto.ai/en/academy/avec-vurto/api-et-mcp-connecter-vurto) (2026-07-19)

Vurto's REST API and MCP (Model Context Protocol) server, available from the Business plan up, let you connect your AI visibility data to your own tools and automate report retrieval, without going through Vurto's web interface.

## What the REST API exposes

The REST API gives programmatic access to the data you'd normally check in the dashboards: visibility scores by brand, category and keyword, results by AI engine, the table of competitors and cited sources, and the results of your product pages' semantic scoring. The main benefit is getting this data out of the interface and into somewhere else — an internal dashboard, an existing reporting tool, a data warehouse — without manual re-entry.

## Two complementary tools, not competing ones

The REST API and the MCP server don't replace each other: they serve two different ways of working with data. The API fits when you know precisely which data you want to retrieve, how often, and in what format — a structured, repeatable use, usually set up once by someone technical and then run automatically. The MCP server fits when the question isn't formulated in advance: you realize mid-conversation with your AI assistant that you need a number or a piece of context about your visibility, without having planned that question beforehand. A team can perfectly well use both in parallel: the API for recurring reports, the MCP server for the ad hoc needs that come up in the flow of work.

## What the MCP server adds

The MCP server addresses a different need: it lets a compatible AI assistant (Claude, Cursor, ChatGPT) query your Vurto data directly in natural language, within the context of a conversation, rather than writing code against the API. In practice, this means you can ask a question about your visibility or your competitors directly inside the tool you already use daily, without opening Vurto or writing a query.

## Connecting Vurto to your tools

The connection is set up by configuring Vurto's MCP server as a data source in the compatible assistant you use — Claude, Cursor or ChatGPT, depending on supported integrations. Once connected, the assistant can pull the relevant information from your Vurto projects (a brand's visibility score, a category's latest analysis, competitors identified on a prompt) right when you need it, in the flow of your work, rather than switching tools.

## Concrete use cases

Automating reports is the most direct use case: a periodic API extraction of visibility scores to feed an internal report or a recurring client review, without manually exporting data every week. For an agency tracking several client projects, this makes it possible to build a custom consolidated report, with its own formatting, rather than sticking to the agency dashboard's format. The MCP server, on the other hand, is more suited to ad hoc, exploratory use: checking the state of your visibility while working, without breaking out of a session in another tool.

## A concrete automation example

Picture a marketing team producing an internal review every Monday morning on the AI visibility of three tracked brands. Without the API, that means opening Vurto, browsing each dashboard, and exporting or copying the numbers by hand into a presentation. With the REST API, a script scheduled for Monday at 8am can automatically pull each brand's average visibility score, its engine breakdown and its top competitors, and feed them straight into the reporting document used in the meeting. The gain isn't just time saved: it's also the certainty that the number presented matches the latest analysis wave, with no gap caused by a forgotten export.

## Before you start

The API and MCP server require a Business plan or above — they aren't available on the Standard and Pro plans. If your main use remains occasional dashboard browsing, web access is more than enough; the value of the API and MCP grows with how often you need automated reports or with the number of projects to consolidate, which makes it a particularly relevant tool in agency mode. See the [plans and credits](/en/academy/avec-vurto/comprendre-plans-credits-vurto) page for eligibility details. Once automation is in place, it fits naturally into a [weekly GEO tracking workflow](/en/academy/systeme-pour-durer/workflow-hebdomadaire-geo), replacing manual exports with automatic retrieval of the metrics tracked in your [dashboards](/en/academy/avec-vurto/lire-dashboards-vurto-scoring-semantique).

## Managing access

As with any programmatic access to tracking data, limit the distribution of API credentials to people who actually need them, and revoke access as soon as an integration project ends. This matters especially in agency mode, where several clients and collaborators may be involved with the same data exposed through the API.

### [Tracking AI crawlers and auditing robots.txt with Vurto](https://vurto.ai/en/academy/avec-vurto/suivre-les-crawlers-ia-avec-vurto) (2026-07-20)

Vurto tracks the passage of AI crawlers (GPTBot, ClaudeBot, PerplexityBot, and others) on your site and audits your robots.txt file, to verify that these bots can actually access the pages you want cited in generative answers.

## Why this tracking matters

Content can be perfectly structured and relevant: if generative AI crawlers can't reach it, it will never be cited. [The robots.txt file and controlling AI crawlers](/en/academy/setup-technique/robots-txt-et-crawlers-ia) are the first technical gateway to your generative visibility: before working on a page's content or semantic scoring, you need to make sure it's accessible to the bots feeding ChatGPT, Claude, Gemini or Perplexity.

## What Vurto shows about crawler activity

Vurto tracks which identified AI crawlers visit your site and how often, giving a concrete indication of the attention generative engines already pay to your pages. A site regularly visited by GPTBot, ClaudeBot and PerplexityBot is being actively explored as a potential source; a site where these bots rarely or never visit certain sections points to a blind spot to fix before anything else.

## A signal to read over time, not as a snapshot

A single AI crawler visit on a given day doesn't tell you much on its own: what matters is how regularly it visits over several weeks. A crawler that used to visit a section of the site regularly and suddenly stops is a signal to prioritize, even with no apparent change on the site — it can point to an infrastructure issue (server response time, intermittent technical errors) that isn't necessarily tied to robots.txt. Conversely, a crawler that has never visited a section doesn't necessarily mean it's blocked: that section may simply never have been judged relevant enough yet by that engine to warrant deeper exploration.

## The robots.txt audit

Vurto analyzes your robots.txt file to identify whether rules block, intentionally or by mistake, AI crawler access to all or part of your site. A blocking rule can be old, inherited from a generic configuration, and may never have specifically targeted AI bots — a frequent case, especially on sites whose robots.txt hasn't been reviewed since these new crawlers appeared. The audit pinpoints exactly which rules apply to which bots, rather than guessing from a raw text file.

## Interpreting a lack of crawling or a block

Two distinct situations call for different responses. If robots.txt explicitly blocks an AI crawler, the fix is direct: adjust the rule to allow access to the relevant sections. If, on the other hand, nothing blocks the crawler but it simply isn't visiting, or rarely does, the issue isn't technical but editorial or structural — the site may not yet be considered a sufficiently relevant or up-to-date source in your sector, which ties back to the questions covered by [analyzing your citation sources](/en/academy/avec-vurto/analyser-ses-sources-de-citations).

## What to fix, and in what order

Always start with explicit blocks in robots.txt: these are the fastest, highest-impact fixes, since they remove a total access barrier. Once these blocks are lifted, watch for crawlers returning to the affected pages in the following analysis waves — a delay of a few days to a few weeks before seeing an effect on your visibility scores is normal. Build this check into your [weekly tracking routine](/en/academy/systeme-pour-durer/workflow-hebdomadaire-geo): a robots.txt audit is only needed once, but checking that nothing gets re-blocked after a technical site update is a control worth repeating regularly.

## What a gap between site sections reveals

Crawler activity isn't necessarily uniform across a whole site. It's common for category pages and the homepage to be visited regularly, while deeper sections — individual product pages, FAQ pages, older blog content — get much less attention. Before concluding there's an access problem, check that these pages are actually linked from the rest of the site (internal linking, an up-to-date sitemap): a crawler, AI or not, only visits what it can reach by following links or reading the sitemap. A simple oversight in updating the sitemap after adding new product pages can explain a lack of crawling just as well as a robots.txt block.

## A common scenario

A technical update to the site (CMS change, hosting migration, new security configuration) can reintroduce a block without any deliberate action — an application firewall or a generic anti-bot rule can treat AI crawlers as unwanted traffic. That's exactly why the robots.txt audit should stay a recurring check rather than a one-time verification done only at the start.

### [Working in a team and in agency mode](https://vurto.ai/en/academy/avec-vurto/equipes-mode-agence-vurto) (2026-07-21)

Vurto offers two levels of collaboration: a team mode, for working together on a single account, and an agency mode, for managing several distinct client projects with access and credits kept separate or shared as needed.

## Team mode: the available roles

In team mode, each invited member is assigned one of four roles: Owner, Admin, Member or Viewer. The Owner has full control over the account, including billing. The Admin can manage projects, prompts and invitations, without necessarily having access to billing, which stays reserved for the Owner. The Member can view and generally contribute to tracking (prompts, personas) without administration rights. The Viewer has read-only access, useful for sharing results with someone who doesn't need to change the configuration. This hierarchy lets you open up Vurto to an entire marketing or product team without giving everyone full control over the account.

## Inviting collaborators

Invitations are sent by email: the invited person gets a link to join the account with the role set at invitation time. The role can be adjusted later if the person's responsibilities change — for example, moving a Member to Admin once they take on coordinating GEO tracking for the whole team.

## Agency mode: a distinct project per client

For an agency managing several clients, team mode alone isn't enough: each client needs its own tracking scope, without one client's prompts, brands or results mixing with another's. Agency mode addresses this by creating a distinct project per client, each with its own tracking of brands, categories, keywords and prompts, all grouped under a single agency account. Every new client project goes through the same 5-step onboarding process described in [Getting started with Vurto](/en/academy/avec-vurto/bien-demarrer-avec-vurto), which keeps the tracking foundation consistent from one client to the next.

## Roles at the project level

Inside each client project, Vurto offers three specific roles: Owner, Exporter and Viewer. The project Owner has full control over it. The Exporter can view results and export them (useful for an account manager who needs to produce a client report without needing to change the tracking configuration). The Viewer is limited to viewing, which suits giving the client themselves guest access so they can follow their results without touching the prompts or personas defined by the agency.

## Shared credits and balance visibility

Credits consumed by analyses (number of requests, number of engines queried) are shared across the projects of a single agency account, rather than siloed project by project. This gives real flexibility: an agency can put more credits toward a client in the middle of an AI visibility push and use fewer on a client in routine tracking mode, without manually reallocating quotas. In addition, if a client is invited as Viewer or Exporter on their own project, the agency can choose to hide the credit balance from them — useful when the agency's business model is a flat monthly fee rather than usage-based billing visible to the client.

## The multi-project agency dashboard

At the agency account level, a consolidated dashboard gives an overview of all active client projects, without having to open each project one by one to check on tracking status. Each collaborator invited on a specific project, meanwhile, only sees that project's scope — the multi-project agency dashboard stays reserved for roles with visibility across the whole account. The exact mechanics of credit consumption by analysis type are detailed in the [plans, credits and billing](/en/academy/avec-vurto/comprendre-plans-credits-vurto) page.

## Team mode or agency mode: how to choose

Team mode is enough as long as you're tracking a single brand or a single catalog, even with several people involved. Agency mode becomes relevant once you manage GEO tracking for several distinct organizations, with brands, categories and prompts that aren't meant to cross over, and where some access (like the end client's) needs to stay strictly confined to their own scope. An agency starting with a single client can perfectly well begin in simple team mode, then switch to agency mode once onboarding a second client.

## Assigning the right roles at the right time

A common trap is granting broader roles than needed for simplicity — for instance, making everyone Admin or Owner to avoid managing permissions. This works in the short term but becomes risky as the team or the number of projects grows: overly broad access multiplies the number of people who can change the tracking configuration (prompts, personas, tracked brands) even when that isn't required for their role. A simple principle: assign the most restricted role that still lets the person do their job, and only raise permissions when their actual responsibility grows.

## Example: a small agency with 3 clients

Picture a GEO agency with three clients: a cosmetics brand, a furniture e-commerce site and a restaurant chain. The agency creates one project per client, each with its own tracked brands, categories and prompts. The agency's founder is Owner of the agency account and of all three projects. A junior account manager is added as Exporter on all three projects: they can produce monthly reports without risking changes to the tracking configuration. Each client, in turn, is invited as Viewer on their own project only — they see their AI visibility results in real time, but not the other two clients' data, nor the credit balance consumed, which stays hidden since the agency charges a flat fee. Credits themselves stay managed at the agency account level: during a spike of activity on the furniture client (a new product line launch), the agency can temporarily put more requests toward that project without changing plans or buying dedicated extra credits.

### [Understanding plans, credits and billing](https://vurto.ai/en/academy/avec-vurto/comprendre-plans-credits-vurto) (2026-07-22)

Vurto offers four plans: Standard, Pro, Business and Agency. Every analysis consumes credits, and the number consumed varies depending on the type of analysis and the number of AI engines queried — there's no permanent free plan, but a 14-day trial with no credit card required, so you can test the product before committing.

## The 4 plans

The Standard plan, at €99/month, gives access to 4 AI models, 1 project, 500 requests per month and tracking for up to 500 prompts. It's the entry point suited to a brand or site starting its generative visibility tracking on a limited scope. The Pro plan, at €199/month, widens access to 8 AI models, 3 projects, 3,000 requests per month, and unlocks assisted product page rewriting. The Business plan, at €499/month, goes up to 10 projects, 20,000 requests per month, unlimited tracked prompts, and gives access to the [REST API and MCP server](/en/academy/avec-vurto/api-et-mcp-connecter-vurto) to connect Vurto to your own tools. The Agency plan, quote-based, is built for agencies managing many clients: up to 50 projects, 50,000 requests per month, a white-label offering and advanced multi-client management, suited to the [agency mode](/en/academy/avec-vurto/equipes-mode-agence-vurto) covered in its dedicated lesson.

## How credits get consumed

Every analysis — whether it's a prompt sent to an AI engine, a semantic scoring analysis, or a detailed analysis — consumes credits within your plan's monthly request quota. The number of credits consumed depends on two main factors: the type of analysis performed, and the number of AI engines queried at once. Querying a prompt across four engines costs more than querying just one, which is why the number of available AI models and the monthly request volume scale together from one plan to the next: the more engines you have active, the more requests each analysis wave consumes, for the same number of tracked prompts.

## What Pro and Business actually unlock

Beyond quota numbers, two features are worth a close look when choosing a plan. Assisted product page rewriting, unlocked from Pro up, makes full sense once you've reviewed the semantic scoring: it lets you act directly on the weak criteria identified rather than starting from a blank page. The API and MCP server, reserved for Business, are useless to a team with no need for automation or external integration — in that case, it's better to stay on Pro and reassess later, once that need becomes concrete, rather than paying for a feature that won't get used.

## The 14-day free trial

Vurto offers a 14-day free trial, with no credit card required at signup, to test onboarding, the automatic site analysis, prompt generation and the first dashboard results before choosing a paid plan. It's a chance to calibrate the number of prompts and engines you actually need for your tracking before committing to a monthly quota. To make the most of this trial period, follow the steps described in [Getting started with Vurto](/en/academy/avec-vurto/bien-demarrer-avec-vurto).

## Choosing your plan based on catalog size

The main criterion for choosing between Standard, Pro and Business isn't just available budget, but the number of elements to track in parallel — brands, categories, keywords and products — combined with how many AI engines you consider essential. A site with a small catalog and a single brand to track is often well-sized on Standard, even if 4 models rather than 8 limits per-engine granularity. An e-commerce site with several categories and an interest in assisted product page rewriting finds more value in Pro. An organization managing several brands or projects, or wanting to automate its reporting via the API and MCP, should aim for the Business plan, the only one unlocking these technical features along with unlimited prompts.

## A concrete benchmark for estimating your consumption

To estimate whether a plan holds up, start from the number of prompts you plan to actively track, multiply by the number of AI engines queried each wave, then by how often waves run in a month. An account tracking 100 prompts across 4 engines, with a weekly wave, consumes on the order of 1,600 requests a month just on base tracking — beyond the Standard quota, which would push toward a Pro plan even with a modest-sized catalog. This calculation stays indicative since the exact cost varies by analysis type, but it helps avoid mid-month surprises.

## What really separates Pro from Business beyond price

Between Pro and Business, the price gap reflects less a simple volume increase than a change in the nature of usage. Pro is still designed for a team steering its tracking mainly from the web interface, with a comfortable number of projects and requests for several categories or brands. Business is aimed at an organization whose GEO tracking becomes a data flow to integrate elsewhere — automated reporting, connecting to other tools via the API and MCP, unlimited prompts so you no longer have to trade off coverage against quota. The choice between the two shouldn't come down to comparing monthly request counts alone, but to asking whether the team actually needs to get data out of Vurto to use it elsewhere.

## Adjusting your testing pace

The monthly requests available need to be split between the number of tracked prompts and how often analysis waves run. Tracking 500 prompts across 4 engines on every wave quickly eats through a Standard quota if waves run too often; in that case, it's better to trim active prompts down to the most strategic ones, or space out analysis waves, rather than upgrading plans by reflex. Conversely, if your monthly quota is never reached even as your catalog grows, that's a signal that a higher plan could unlock more coverage (more engines, projects or prompts) without changing your tracking habits. Good practice is to review this sizing roughly every quarter, especially after a significant addition of products, categories, or opening a new project.

### A system built to last

### [The weekly GEO workflow](https://vurto.ai/en/academy/systeme-pour-durer/workflow-hebdomadaire-geo) (2026-07-23)

A GEO workflow that lasts is built on a fixed weekly ritual, not on occasional sprints. Every week you measure your scores, look at what competitors are doing in AI answers, pick one priority piece of content and publish it. Without that rhythm, GEO turns back into a project you restart every so often, with results that stall.

## Why weekly, not monthly

Generative AI answers change faster than Google results. A model updates its cited sources, a competitor publishes content that takes your spot on a prompt, a piece of news temporarily shifts attention to a related topic. If you only check your scores once a month, you discover problems too late to react effectively. A weekly check-in, even a short one, lets you spot a drop in the first week and adjust before it settles in.

The ritual doesn't need to be long. Budget 45 minutes to 1.5 hours a week once it's running smoothly — that's the price of staying visible in a channel that keeps moving.

This modest time commitment is also what makes the ritual sustainable. A heavier weekly process, one that requires half a day of preparation, tends to get skipped the first time the team is under pressure on something else, and a skipped week almost always turns into a skipped month. Keeping the ritual light is a deliberate choice to protect its regularity, not a shortcut that compromises its usefulness.

## Monday: measure without drowning in numbers

At the start of the week, check your visibility score by engine, by brand and by priority keyword. The goal isn't to analyze everything in detail but to answer three questions: is the overall score rising, flat, or falling? Is one engine lagging behind the others? Has a keyword or product lost its place in the answers?

If you use Vurto, [reading your AI visibility dashboard](/en/academy/avec-vurto/lire-dashboards-vurto-scoring-semantique) correctly at this stage avoids two classic mistakes: reacting to a one-day variation with no statistical meaning, or ignoring a trend that's confirmed over several consecutive weeks. Log the numbers in a simple table, week by week. That time series, more than a single snapshot, tells you whether your work is paying off.

## Tuesday-Wednesday: competitor and cited-source review

Once the measurement is done, look at who shows up in your place or alongside you in the answers. Two angles to cover: which competitors get cited more often than you on your priority prompts, and which sources (articles, comparison pages, forums, product pages) come up most often in AI citations for your sector.

[Analyzing your competitors in AI answers](/en/academy/avec-vurto/analyser-ses-concurrents-dans-les-reponses-ia) on this basis gives you concrete material: when a competitor gains ground, it's almost always because a source that mentions them gained authority, or because they published content that answers the user's question more precisely. This review feeds directly into Thursday's prioritization.

## Thursday: prioritize the week's content

This is the day you decide what to produce. Cross three signals: keywords or prompts where your score is weak but the business stake is real, topics where a competitor is gaining ground and you have a better answer to offer, and external sources worth prioritizing to strengthen the [off-site signals that matter](/en/academy/strategie-seo-geo/signaux-off-site-qui-comptent).

Don't try to cover everything. One well-built piece of content that precisely answers a frequent question and cites verifiable data outweighs a series of generic pages churned out one after another. Pick one or two priority projects for the week, not ten.

## Friday: publish and document

Close the week by publishing what's ready and logging what was done: content published, source worked on, technical fix made. This logbook serves two purposes. First, it lets you connect an action to a visible effect several weeks later, when you review your [KPIs](/en/academy/systeme-pour-durer/kpis-qui-comptent). Second, it prevents you from redoing the same work twice or forgetting a lead identified the month before.

This five-day breakdown is a starting point, not a rigid template. Working solo or in a small team, you can perfectly well merge the measurement and competitive review into a single 45-minute session on Monday, then spend the rest of the week producing content. In a larger team, it's often more effective to name a single owner for the ritual, responsible for collecting and sharing the numbers, even if the Thursday prioritization decision stays collective.

The risk to watch for, regardless of team size, is that the ritual turns into an empty formality: a table filled in without being read, a meeting held without any decision coming out of it. A weekly workflow only has value if it systematically leads to a concrete action, even a modest one, every week.

## The workflow on one page

To keep this ritual simple, here's the structure to repeat every week:

- **Monday (30 min)**: check visibility scores by engine and keyword, compare to the previous week.
- **Tuesday-Wednesday (30-45 min)**: review cited competitors, review sources rising in AI citations.
- **Thursday (30 min)**: choose 1-2 content or link-building priorities for the week.
- **Friday (15-20 min)**: publish, document, note hypotheses to check the following week.

This rhythm may look modest compared to the scope of the topic, but its regularity is exactly what makes the difference. GEO isn't a project you finish, it's a discipline you maintain. A weekly workflow kept up for three months will almost always produce better results than a one-off audit followed by silence. Start small if you have to, but start every week.

### [The KPIs that matter](https://vurto.ai/en/academy/systeme-pour-durer/kpis-qui-comptent) (2026-07-24)

Four indicators are enough to steer a GEO strategy over time: visibility score, share of voice, citation rate, and AI-driven traffic. The trap isn't missing one, it's tracking too many and losing sight of which one actually drives your decisions.

## Visibility score, your main barometer

The visibility score measures, across a representative set of prompts, how often your brand or product appears in generative AI answers. It's the indicator closest to a thermometer: it tells you where you stand, not why. Track it by engine (ChatGPT, Claude, Gemini, Perplexity), because a global score can hide strong performance on one engine and near-total absence on another. Track it by product category or keyword too: a brand can be well cited on its own name and invisible on the generic queries that actually bring in new customers.

This score really pays off once you know how to [read your AI visibility dashboard](/en/academy/avec-vurto/lire-dashboards-vurto-scoring-semantique): looking at the trend over several weeks rather than a single number, and crossing the score by engine with the score by keyword to spot where to focus effort.

## Share of voice, to position yourself against competitors

The visibility score alone says nothing about your relative position. Share of voice compares your presence to your competitors' on the same set of prompts: out of a hundred answers generated around your category, how many mention you, how many mention each competitor. It's the indicator that answers "am I gaining or losing ground" rather than "am I visible in absolute terms."

It's also the indicator that directly connects this module back to the first one in the training: [being everywhere is the goal of AI share of voice](/en/academy/comprendre-le-geo/objectif-etre-partout-share-of-voice), and this number tells you whether you're getting closer to it or further from it, month after month.

## Citation rate, to understand the quality of your presence

Appearing in an answer and being cited as a source aren't the same thing. Citation rate measures how often your pages are explicitly mentioned as a reference by the AI, with a link or direct reference to your content, rather than just a passing mention of your brand in the middle of a list. It's a more demanding indicator than raw visibility score, but more revealing of your actual authority on a topic.

A citation rate that rises while the visibility score stalls often signals that your content is gaining depth and credibility, even if how often it appears hasn't moved yet. It's a leading signal worth watching, since it usually precedes a rise in visibility score by a few weeks.

In practice, tracking citation rate also tells you which content format is working. A page that gets cited repeatedly usually has something in common with your other cited pages: original figures, a clear structure, a direct answer stated early rather than buried after several paragraphs of introduction. Comparing your cited pages against your non-cited ones is one of the more useful diagnostic exercises you can run once you have a few months of data.

## AI traffic, the indicator that ties it all to the business

The three previous indicators measure your presence in the answers. AI traffic measures what that presence actually produces: visits to your site coming from clicks in ChatGPT, Perplexity, or other AI interfaces. In Google Analytics 4, this traffic generally shows up under referral channels, under sources like chatgpt.com, perplexity.ai, or gemini.google.com. You sometimes need to build a dedicated channel grouping to see it clearly rather than have it buried in generic referral traffic.

This is the indicator that turns a visibility-measurement exercise into a business metric: how many visitors, how many leads, how many conversions come from this channel. Without that link, the visibility score stays an abstract number; with it, it becomes an argument for the time invested every week.

## A concrete example to visualize it

Take a fictional brand starting its GEO tracking. In month 1, its visibility score is 12% across its ten priority prompts, its share of voice is 8% against three identified competitors, its citation rate is close to zero, and its AI traffic in GA4 is anecdotal — a few dozen sessions for the month. After three months of steady publishing and stronger off-site signals, the visibility score reaches 22%, share of voice hits 15%, the citation rate starts showing up on two well-sourced pieces of content, and AI traffic passes a hundred monthly sessions.

What makes this example useful isn't the numbers themselves, which vary by sector, but the order in which the indicators move: citation rate often improves before the overall visibility score, which in turn improves before traffic becomes meaningful. Expecting an immediate effect on traffic within the first few weeks almost always leads to avoidable disappointment.

## Connecting the four indicators to concrete goals

These four KPIs shouldn't stay as four isolated charts. Connect them to each other and to an explicit business goal: for example, targeting a 20% share of voice on the quarter's five priority keywords, with a rising citation rate on the two engines your audience uses most, and a monthly AI traffic target that justifies the investment. Review these goals every quarter, not every week: the [weekly workflow](/en/academy/systeme-pour-durer/workflow-hebdomadaire-geo) is there to adjust execution, not to keep redefining the target.

One last word of caution: never judge a single month in isolation. GEO has natural variability, models evolve, tested prompts change wording. It's the trend over eight to twelve weeks that matters, not the week-to-week swing.

### [The mistakes that kill your progress](https://vurto.ai/en/academy/systeme-pour-durer/erreurs-qui-tuent-la-progression) (2026-07-25)

Most GEO strategies don't stall from a lack of knowledge but because of a handful of recurring mistakes: content that's too generic, giving up too early, relying on a single engine, neglecting the technical side. Fixing them one by one is often enough to unblock progress that looked stuck.

## Producing generic content instead of precise answers

The most common mistake is publishing content that looks like ten competitors' content: a general definition, a list of tips already seen elsewhere, no proprietary data. Generative AIs favor sources that give a specific, verifiable answer to a precise question, not ones that rephrase an already widely available consensus.

The simple test: if your content could have been written by any competitor in the sector, it won't set you apart in AI answers. A figure drawn from your own experience, a detailed real-world case, an answer that takes a clear position on a point where opinions diverge — that's what gives a model a reason to cite you rather than someone else.

## Giving up after two or three weeks

GEO almost never produces visible results before several weeks, sometimes several months for the most competitive keywords. Many teams run a first audit, publish two or three pieces of content, don't see their score move, and stop. This is the costliest mistake, because it happens right when the effort is starting to pay off behind the scenes: the AIs still need to crawl the new content, index it, and associate it with the right queries before it shows up in answers.

That's why a [weekly workflow](/en/academy/systeme-pour-durer/workflow-hebdomadaire-geo) kept up over time is more effective than an intense sprint followed by a stop. A modest but continuous pace over six months beats one month of intense effort followed by abandonment.

## Tracking only one AI engine

Focusing all your attention on ChatGPT because it's the most used is a classic mistake. Perplexity works differently (more search-oriented, with direct source citation), Claude and Gemini have their own logic for selecting sources, and the split between these engines varies by audience and sector. A brand can be well positioned on Perplexity and nearly absent on Gemini, and you won't see it if you only look at an aggregated score.

Tracking the four engines separately, even at a basic level, prevents you from building a strategy optimized for a single channel that could lose relevance or change behavior overnight.

## Ignoring the technical side in favor of content alone

Publishing good content isn't enough if it stays hard to crawl, poorly structured, or blocked for generative AI robots. Many teams pour all their effort into writing and neglect the technical foundations: whether AI crawlers can access the content, clear page structure, no unintentional blocks. Excellent content that's invisible to crawlers has the same effect as content that doesn't exist.

A visibility score can shift from one week to the next without any of your actions being the cause: providers regularly update their models, which sometimes changes how sources are selected, without a detailed advance announcement. Reading every variation as the direct consequence of your latest publication leads to the wrong conclusions either way: crediting an action that had nothing to do with it, or doubting a strategy that was actually working well.

Good practice means telling apart one-off swings, which may come from a model update, from trends confirmed over several consecutive weeks, which more likely reflect a real effect of your work. This is one of the reasons regular weekly measurement matters more than reacting to a single isolated number.

Another frequent mistake is running GEO in isolation, disconnected from SEO, existing content, and acquisition channels already in place. The [off-site signals that matter](/en/academy/strategie-seo-geo/signaux-off-site-qui-comptent) for GEO largely overlap with those that matter for classic SEO: domain authority, external mentions, backlink quality. Treating the two separately doubles the work and weakens each channel instead of reinforcing them together. This topic is developed further in this module, in the lesson on [GEO as an acquisition channel](/en/academy/systeme-pour-durer/geo-comme-canal-dacquisition).

## Never checking what competitors are doing

Finally, moving forward without ever checking who's capturing citations in your place deprives your strategy of an essential signal. Regularly [analyzing competitors in AI answers](/en/academy/avec-vurto/analyser-ses-concurrents-dans-les-reponses-ia) helps you understand why a score is stalling: often it's not that your content is bad, it's that a competitor answered the same question better, with a more authoritative or more recent source.

One last, quieter mistake is measuring your progress by comparing yourself right away to the biggest players in the sector, the ones who've been publishing content for years and already have well-established domain authority. Against that benchmark, any real progress looks negligible, which wrongly leads to the conclusion that the strategy isn't working.

It's more useful to track your own trajectory over time and compare yourself to competitors of a similar size, before aiming at the sector leaders. A visibility score that doubles in three months is a positive signal, even if it stays far behind that of a long-established player. It's this relative progress, week after week, that shows whether the pace of work you've put in place is working.

## What to remember

None of these mistakes is irreversible. They share one thing in common: they all come from a lack of consistency or a lack of overall perspective. Fixing just one of them, without changing the rest of the strategy, often already produces a visible unblock within a few weeks.

### [GEO as an acquisition channel](https://vurto.ai/en/academy/systeme-pour-durer/geo-comme-canal-dacquisition) (2026-07-26)

GEO works as a full acquisition channel, with its own metrics, budget, and place in the marketing mix — not as an isolated technical project handed to one person for two weeks. Integrating it properly changes how you steer it and how you justify the time spent on it.

## Getting out of the "one-off project" silo

Many companies approach GEO as an audit they run once, fix, then forget. This is the most common mistake listed in the lesson on [mistakes that kill your progress](/en/academy/systeme-pour-durer/erreurs-qui-tuent-la-progression): treating the topic as a silo rather than as a channel you feed continuously, the same way you would SEO or paid.

An acquisition channel is defined by three things: a recurring budget (time or money), quantified goals, and integration into overall marketing reports. GEO needs to meet these three criteria to survive past the first quarter of attention it usually gets.

## Building on what already exists in SEO

The good news is GEO doesn't start from zero. A large share of the [off-site signals that matter](/en/academy/strategie-seo-geo/signaux-off-site-qui-comptent) for AI visibility (domain authority, backlink quality, mentions in trade press) are the same ones that matter for classic SEO. If your SEO team is already working on link building and press relations, that work also feeds your visibility in generative AIs, with no extra dedicated effort.

In practice, this means GEO shouldn't be handed to a separate team that reinvents its own content and link-building priorities. It should build on the editorial calendar and link strategy already in place, with targeted adjustments: rewording certain content to answer prompt questions more directly, prioritizing sources that appear most often in your sector's AI citations.

## Coordinating with paid and social

Paid and social indirectly influence GEO. A campaign that generates search volume and brand mentions feeds the signals AIs use to assess your notoriety. Recurring thought leadership on LinkedIn or in specialized communities can become a cited source if it gives a clear answer to a frequent question in your sector.

The reverse is also true: a brand that's highly visible in AI answers builds a form of trust that improves click-through rates on paid campaigns, because the user has already encountered the brand elsewhere. These channels reinforce each other when coordinated, which means sharing the same thematic priorities quarter over quarter rather than handling them in separate meetings.

An acquisition channel that rests entirely on one person, with no documentation or shared process, disappears as soon as that person changes roles or priorities. This is a particularly common risk for GEO, often carried by a single motivated person inside a broader marketing team that doesn't yet fully engage with it. Documenting the workflow, the prompts being tracked, and the prioritization decisions lets someone else pick up the ritual without starting from scratch.

This documentation doesn't need to be elaborate: a shared table with the scores logged each week and the actions taken is largely enough. The goal is for the channel to survive organizational changes, the same way any other established acquisition channel in the company would.

The same logic applies to knowledge about competitors and cited sources: if that understanding only lives in one person's head, it has to be rebuilt from scratch every time roles shift. Writing it down, even briefly, turns a personal habit into an organizational asset.

## Measuring the return like any other channel

To justify its place in the mix, GEO needs to be measured with the same rigor as other channels: traffic generated, attributed conversions, cost per action when a dedicated budget is involved. The [KPIs that matter](/en/academy/systeme-pour-durer/kpis-qui-comptent) — visibility score, share of voice, citation rate, AI traffic — enable this comparison, provided you explicitly connect them to business goals and not just to abstract visibility scores.

One caution: AI traffic is often still lower in volume than classic SEO, at least for now, across most sectors. Comparing GEO to SEO in absolute traffic terms can be misleading. It's more accurate to compare growth dynamics and the quality of the visitors it brings: a user arriving via an AI answer has usually already received a detailed answer to their question, which can translate into a different conversion rate than classic SEO.

## Growing the budget with the results

As with any acquisition channel, the budget devoted to GEO doesn't need to be set once and for all. At the start, it's reasonable to limit the investment to internal time: a few hours a week for the tracking ritual, with no dedicated budget for link building or outsourced production. Once a positive trend is confirmed over several months, both in visibility score and in traffic generated, it becomes fair to gradually increase the investment, the same way you would for a paid campaign that's starting to convert.

This logic avoids two opposite pitfalls: over-investing in the first month in a channel whose return isn't yet proven, or under-investing indefinitely because the first results looked modest. The budget should follow the results curve, not precede it or ignore it.

## Giving GEO the place it deserves in reports

Finally, integrating GEO into your monthly or quarterly marketing reports, alongside SEO, paid, and social, anchors it durably in the company's priorities. A channel that never appears in a leadership report always ends up losing budget and attention, regardless of its results. Presenting a simple trend — visibility score, share of voice, AI traffic — in the same format as other channels avoids this pitfall and legitimizes the time invested each week in the [GEO workflow](/en/academy/systeme-pour-durer/workflow-hebdomadaire-geo).

### [The future of GEO](https://vurto.ai/en/academy/systeme-pour-durer/futur-du-geo) (2026-07-27)

No one can predict with certainty exactly how GEO will evolve, but several current signals point in a reasonable direction: the rise of AI agents acting on the user's behalf, and a likely gradual structuring of technical standards dedicated to AI crawlers. This module is deliberately cautious: the goal is to anticipate trends, not to predict a fixed future.

## From answers to actions: the rise of AI agents

So far, most interactions with generative AIs are limited to a question followed by a text answer. Several signals suggest this is shifting toward agents capable of acting directly: browsing a site, comparing offers, filling out a form, even completing a purchase on the user's behalf. Recent announcements around browsing and automated-purchase agents from several model providers point in this direction, though their large-scale adoption remains to be confirmed.

If this trend is confirmed, being visible in an answer won't be enough on its own: the site will also need to be usable by an agent, with clear pages, structured product information, and a purchase flow that doesn't depend on complex interactions designed only for a human. This is a logical extension of what [AI share of voice](/en/academy/comprendre-le-geo/objectif-etre-partout-share-of-voice) already measures: being present isn't enough, you also need to be usable.

## Toward a structuring of technical standards

The llms.txt file, inspired by robots.txt, offers a standardized way to signal to generative AIs which pages to prioritize when understanding a site. Its adoption remains partial today, and its actual impact on visibility results isn't yet clearly established by large-scale independent data. It's reasonable to set it up as a precaution, at low cost, without expecting a guaranteed short-term effect.

More broadly, it's likely that the coming months will see more technical conventions emerge for AI crawlers, similar to what happened for classic SEO with the XML sitemap or schema.org structured data. Nothing indicates, however, that a single standard will quickly dominate; several approaches will probably coexist for a while.

## A probable consolidation around sources deemed reliable

Generative AIs appear to be continuously refining their criteria for selecting cited sources, with a growing preference for content that's verifiable, recent, and clearly attributable to an origin. If this trend continues, it would strengthen the advantage of brands that publish original, documented data over those that simply rephrase content already available elsewhere. This is consistent with what's already recommended in the lesson on [mistakes that kill your progress](/en/academy/systeme-pour-durer/erreurs-qui-tuent-la-progression): favor precision and original data over volume.

## Staying critical of the announcements

GEO attracts a lot of announcements and promotional talk, often phrased in the certain future tense rather than the conditional. A new tool, a new model feature, an article declaring the end of classic SEO: this kind of communication has an incentive to dramatize, which doesn't mean the underlying analysis is wrong, but that it deserves checking before you act on it.

Good practice means distinguishing a change confirmed by observable behavior in your own data from an announcement that, for now, remains an intention or a feature in limited testing. A visibility score that genuinely moves over several weeks after an announced change is a more reliable signal than an article promising a coming revolution.

## What probably won't change

Despite these developments, some GEO fundamentals have little reason to disappear. A site's authority, measured through its [off-site signals](/en/academy/strategie-seo-geo/signaux-off-site-qui-comptent) (backlinks, mentions, reputation), will most likely keep playing a role, because it reflects a trust that goes beyond a single page's content. Similarly, the need to answer a question precisely rather than produce generic content will probably remain a central criterion, regardless of how the engines evolve technically.

Without rushing, a few adjustments have a low cost and a defensible upside even if the trends discussed above only partly play out: making product pages more explicit and easier for an automated system to interpret, clearly documenting essential information (price, availability, specifications) rather than scattering it across complex visuals or scripts, and following the evolution of technical standards without rushing to adopt each one the moment it's published.

These adjustments largely overlap with the good practices already recommended throughout this module: precise content, clean technical structure, regular measurement. There's no need for a separate action plan to prepare for the future of GEO — the discipline already in place is enough to absorb most of what's coming.

If a bigger shift does materialize, such as agents completing a significant share of purchases on a user's behalf, it will most likely show up gradually in your own data before it becomes obvious in the trade press: a slow rise in sessions with unusual navigation patterns, a growing share of traffic from sources you can't fully attribute yet. Keeping an eye on these edge cases in your analytics, without over-interpreting a handful of sessions, is a reasonable way to stay ahead without acting on speculation alone.

## How to position yourself without over-investing in uncertain bets

The right posture in the face of this uncertainty isn't inaction, but proportionality: put in place the low-cost, reversible adjustments now (technical structuring, llms.txt, factual and sourced content), and stay alert to signals that would confirm or contradict the bigger trends, like a real rise in transactional agents. Regular tracking of the [KPIs that matter](/en/academy/systeme-pour-durer/kpis-qui-comptent) is exactly the tool that lets you catch these shifts early, without having to bet today on one scenario over another.

GEO will likely remain, in the months ahead, a field that moves faster than classic SEO did in its early days. The discipline of continuous measurement and adjustment built throughout this module is what lets you keep pace without having to rebuild everything at every change.

### [The 90-day GEO action plan](https://vurto.ai/en/academy/systeme-pour-durer/plan-daction-90-jours) (2026-07-28)

A 90-day GEO plan breaks down into three one-month phases: laying the technical foundations, producing the first content and setting up tracking, then scaling with a sustainable weekly system. This plan revisits, in order, the essentials covered throughout this training.

## Before you start: what this plan assumes

This plan assumes you've already identified your priority keywords and prompts, and that you know where your brand stands against competitors on those queries. If that's not done yet, spend the first few days on this step before entering Month 1: without this baseline, the priorities in the following months will rest on assumptions rather than data.

## Month 1: technical foundations

The first month has a single goal: making sure your site is technically accessible and understandable to generative AI crawlers. Without this base, no content published afterward can fully pay off.

**Week 1 — Technical audit**
- Check that the site isn't unintentionally blocked for AI crawlers (robots.txt, server rules).
- Review the structure of key pages: headings, hierarchy, semantic markup.
- Identify the product or category pages most strategic for your priority prompts.

**Week 2 — Technical fixes**
- Fix the blockers identified in week 1.
- Set up or verify the XML sitemap and basic structured data.
- Assess whether an llms.txt file is worth setting up to flag priority pages to generative AIs.

**Week 3 — Technical SEO basics**
- Check the fundamentals that also drive classic SEO: load speed, indexability, no critical errors.
- Clean up or merge pages that cannibalize each other on the same queries.

**Week 4 — Set up tracking**
- Define the list of prompts and keywords to track over time.
- Configure AI traffic tracking in Google Analytics 4 (dedicated channel grouping).
- Take a first visibility-score reading to serve as a baseline going forward.

**End-of-month-1 milestone**: the site is technically clean, accessible to AI crawlers, and you have a baseline visibility score on your priority prompts.

## Month 2: understanding GEO and producing first content

The second month is dedicated to a deeper understanding of how GEO works and to producing the first content designed to be cited by generative AIs, along with setting up tracking through Vurto.

**Week 5 — Understand the goal**
- Clarify the [AI share of voice](/en/academy/comprendre-le-geo/objectif-etre-partout-share-of-voice) target on your priority prompts: appearing once isn't enough, the goal is consistency across all engines.
- Map the frequent questions from your audience that don't yet have a clear answer on your site.

**Week 6 — First GEO-ready content**
- Write one or two pieces of content that precisely answer the identified questions, with proprietary, verifiable data.
- Avoid generic content: every page needs to offer an answer competitors don't provide in the same way.

**Week 7 — Set up Vurto tracking**
- Configure the project and test prompts in Vurto.
- Learn to [read your AI visibility dashboard](/en/academy/avec-vurto/lire-dashboards-vurto-scoring-semantique): score by engine, by keyword, trend over time.
- Establish a first comparison with competitors using [competitor analysis in AI answers](/en/academy/avec-vurto/analyser-ses-concurrents-dans-les-reponses-ia).

**Week 8 — Adjust**
- Compare the visibility score to the month-1 baseline.
- Prioritize two to three additional pieces of content based on the gaps observed against competitors.

**End-of-month-2 milestone**: at least three GEO-ready pieces of content published, Vurto tracking operational, and a first quantified comparison against your main competitors.

## Month 3: scaling up and installing the system

The third month turns the first results into a lasting system: content production at scale, stronger off-site signals, and permanently establishing the weekly rhythm.

**Week 9 — Content strategy at scale**
- Build an editorial calendar covering the full set of priority prompts, not just the first two or three tackled.
- Prioritize topics where the gap with competitors is largest.

**Week 10 — Link building and off-site signals**
- Identify the external sources that matter most in your sector's AI citations.
- Launch the first link-building or press-relations actions targeted at those sources.

**Week 11 — Install the weekly workflow**
- Formally set up the [weekly GEO workflow](/en/academy/systeme-pour-durer/workflow-hebdomadaire-geo): measurement, competitive review, prioritization, publishing.
- Assign a clear owner for this ritual, even if the time devoted to it stays limited.

**Week 12 — Review and set next quarter's goals**
- Compare visibility score, share of voice, and AI traffic to the baseline set in week 4.
- Set the next quarter's quantified goals on this basis.

**End-of-month-3 milestone**: a weekly system kept up for at least two weeks, measurable progress on the baseline KPIs, and an editorial calendar that goes beyond the initial content base.

## Adjusting the pace if needed

This 90-day calendar assumes a minimum availability of a few hours a week. If your team can only devote a third of that time, it's better to spread the plan over five or six months rather than skip steps: Month 1's technical foundations condition the success of everything that follows, and rushing them in favor of content production rarely produces good results. Conversely, a team with more time available can run some weeks in parallel — for instance, launching Month 3's first link-building actions as early as Month 2, if the technical foundations are already solid.

## After 90 days: maintaining the system

Once this plan is complete, the work doesn't stop — it changes shape. The first 90 days build the foundations and demonstrate initial traction; the following months are about maintaining the weekly rhythm, watching for signals of a shifting landscape, and adjusting the content strategy as results come in. That discipline, more than the intensity of the first sprint, is what will determine your position in generative AI answers a year from now.

### Other lessons

### [GEO and AEO: understanding optimization for generative AI](https://vurto.ai/en/academy/geo-aeo-basics/geo-aeo-understanding-generative-ai-optimization) (2026-07-23)

SEO took twenty years to become an unavoidable discipline. GEO (Generative Engine Optimization) and AEO (Answer Engine Optimization) are going through the same shift, in a fraction of the time.

## From a results page to a single answer

With Google, you got a page of ten results and the user chose. With ChatGPT, Claude, Gemini or Perplexity, the user gets one synthesized answer — often a single brand, a single product, a single recommendation. Being #2 in an AI answer is worth almost nothing: either you're cited, or you're invisible.

## GEO vs AEO: two angles on the same problem

- **GEO (Generative Engine Optimization)**: optimizing your content and digital presence to be cited as a source by generative engines (ChatGPT, Gemini, Claude, Perplexity) when they build an answer.
- **AEO (Answer Engine Optimization)**: optimizing specifically to become THE direct answer to a precise question, including in Google's AI Overviews or voice assistants.

In practice, the two overlap heavily: a well-structured, well-documented brand that's cited by trustworthy sources has better odds of being chosen, whether it shows up as a "source" or as "the answer."

## What generative AIs actually look at

A generative AI doesn't "browse" your site like a human visitor. It relies on what has been indexed, summarized and judged trustworthy: clear content structure, structured data (schema.org), FAQ sections, customer reviews, consistency of information across the web (social profiles, directories, press), and how up to date that information is.

## A simple 4-step method

To turn this theory into an action plan:

1. **Decode** what AIs currently answer about your brand, your products and your industry.
2. **Audit** your digital presence (site, product pages, third-party content that cites you — or doesn't).
3. **Compare** your visibility against competitors, engine by engine.
4. **Prioritize** a concrete action plan, from quick wins (structuring an FAQ) to a long-term content strategy.

That's exactly what Vurto does: turning uncertainty ("does ChatGPT even mention us?") into actionable data. To go further, head to [dashboards & semantic scoring](/en/academy/dashboards-analytics).