Glossaire · GEO

Chunking

Chunking refers to splitting a piece of content into self-contained fragments, called chunks, that AI engines and RAG systems index, vectorize, and cite independently of the rest of the page. Each chunk groups a few sentences or a paragraph carrying one complete idea: a definition, an answer, a statistic. When a query comes in, the engine does not retrieve the whole page but the most relevant chunks, which it turns into embeddings to measure their semantic closeness to the question. A well-built chunk is short, factual, free of dependence on neighboring context, and understandable in isolation. It is the real unit of citation in generative search: ChatGPT, Perplexity, or AI Overviews cite a passage, never a full URL. Mastering chunking therefore means structuring your content so it segments cleanly and maximizes citability within the answers generated by LLMs.

Chunking has become a core skill in GEO (Generative Engine Optimization) because AI engines no longer reason at the page level, but at the fragment level. Understanding how your content will be split means understanding what will actually get cited.

How it works

When content enters a generative search system, it is first segmented into chunks. Each fragment is then converted into an embedding, a vector representation of its meaning. At the user's query, the engine computes the closeness between the question and every stored chunk, then selects the most relevant ones to feed the generated answer.

The split can follow a fixed logic (token count) or a semantic one (by paragraph or section). Modern approaches favor semantic splitting: one chunk = one idea. This is where your editorial structure makes the difference.

Why it matters

If your ideas are diluted across several paragraphs or depend on the preceding context, they segment poorly and lose their meaning once isolated. As a result, the AI does not cite them. Conversely, content designed around self-contained chunks mechanically increases your AI citability.

Key takeaway

Write each paragraph as if it could be cited alone, outside of any page. That is exactly what a generative engine does.

How to optimize chunking

A few concrete principles: start each section with a clear statement, avoid pronouns that point back to a distant paragraph, embed the statistic or definition inside the sentence that carries it, and use descriptive subheadings. These signals help the AI delimit clean chunks and tie them to the right search intent, which improves your chances of appearing in generated answers.

FAQ

Questions fréquentes

There is no universal figure, but an effective chunk usually fits in one paragraph carrying a single complete idea, around a few sentences. Too short, it lacks context; too long, it dilutes the semantic signal and loses citability.

No. Chunking complements SEO by optimizing how AI segments your content. You still need to be indexed and well structured for Google, but chunking adds a layer dedicated to generative search.

Pour aller plus loin