Chunking: how the LLM splits up your content
Before an AI reads your article, it gets split into small, independent blocks. Understanding this splitting process, called chunking, changes how you should write each section.
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) 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, 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 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.