Chunking: definition and who actually does it

Chunking (also called document chunking) is the process of cutting a document into segments, called chunks, so a retrieval system can store and search them separately. It is a core step in retrieval-augmented generation, where an AI engine pulls passages from an index and hands them to a language model to write an answer.

The key fact is who does it. The engine's operator chunks documents it already holds. Your page is one input. The segment boundaries, sizes and overlaps are settings inside someone else's pipeline.

This entry is a short definition. For the full question of whether to restructure your pages, read our guide on whether you should chunk content for LLMs.

A typical pipeline runs in this order:

1. A crawler fetches your page and extracts the text.
2. The operator splits that text into chunks, by fixed length, by sentence, or by detected topic boundaries.
3. Each chunk is converted into an embedding and stored in the index.
4. At answer time, the engine matches the query (often a rewritten version of what the user typed) against stored chunks.
5. The best-matching chunks enter the model's context, and the model writes the answer.

The method matters to the operator. A March 2026 study (arXiv:2603.06976) compared 36 chunking methods across 5 embedding models and 6 domains. Every variable it tested is one the retrieval operator sets. None is a variable a publisher controls.

Chunking sits in the retrieval stage. It happens after access (can the crawler fetch and store your page) and before generation (which retrieved passages the model uses and cites).

Chunking explains how engines read pages, but it is not a lever you can pull. You cannot tell an engine where to split your page, and no engine documents a way to request it.

What you can control comes earlier and is more concrete:

• Whether your text exists in the fetched HTML. Many AI crawlers don't execute JavaScript, so content that only appears after scripts run may never reach the index to be chunked.
• Whether the crawler is allowed in and can reach the server. robots.txt rules are per bot, and a firewall can block a bot that robots.txt allows.
• Whether key facts are stated plainly. A specific price, number or one-line definition on a fetchable page gives the engine a sentence it can quote, whichever chunk it lands in. Our playbook on publishing specific, fetchable facts covers this.

For a fuller split between what you can and can't change, see which AI search weaknesses you can actually fix.

"Chunk your content for LLMs"

This advice is widespread and has no first-party support. Google's guide to optimizing for generative AI features (July 2026) rejects chunking as a requirement and states there is no ideal page length. It also says you don't need to write in a specific way just for generative AI search. No other major engine documents a preferred page segmentation either.

"Shorter, self-contained blocks get cited more"

The experimental evidence points the other way. A 252,000-trial factorial study (arXiv:2605.25517) found that formatting-only edits have little impact on citation. A 2026 critical survey of 45 GEO studies (arXiv:2607.14035) warns that citation-oriented rewrites can impair retrieval. Our comparison of on-page rewrites versus access fixes walks through that evidence.

"So structure is useless"

That overcorrects. Clear headings and short paragraphs help human readers, and that is reason enough to use them. They just aren't a documented route into AI answers.

• Retrieval: the answer-time step where chunks are matched to a query. See how large language models find content.
• Embedding: the numeric representation of a chunk used for matching.
• Context window: the set of retrieved passages the model reads before it answers.
• AI-friendly writing: a related claim that a special style earns citations. Our page on whether an AI-friendly style exists covers what the evidence says.

Can I see how an AI engine chunked my page? No. As of October 2026, no major AI engine exposes its chunk boundaries, chunk sizes or the passages it stored from your page. Some APIs show which URLs were retrieved and which were cited, but the segmentation itself stays inside the operator's index. Any tool claiming to show your real chunks is showing its own simulation.

Should I split long pages into many short pages? Not for AI search reasons. Google says there is no ideal page length and that chunking is not a requirement for its generative AI features. Split a page when it helps readers find what they need. Splitting purely to influence an engine's chunking has no published evidence behind it, and thin, duplicated pages create their own risks.

Does chunking affect Google AI Overviews differently from ChatGPT? Each operator runs its own pipeline, so the method likely differs, but none publishes it. For Google, the documented requirement is simpler: a page must be indexed and eligible to be shown with a snippet to appear as a supporting link in AI Overviews or AI Mode. Google states there are no additional technical requirements.

Is a chunking study a reason to rewrite my content? No. The chunking literature, including the March 2026 study of 36 methods, measures choices the retrieval operator makes on documents it already holds. It tells engine builders which settings retrieve better. It does not test whether a publisher rewriting a page changes what gets retrieved or cited in a live engine.