Plain LLM vs. retrieval-augmented LLM (RAG)
A plain LLM answers from what it learned during training. A RAG-based LLM fetches up-to-date information before answering. The difference changes everything for your visibility.
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 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, 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.