When an AI search engine answers a question from the web, it does not read your page the way a person does, top to bottom with everything in mind. It reads pieces. The page is cut into passages, each passage is scored against the question, and only the best few are handed to the model that writes the answer.
That cutting step is chunking. It explains a lot of behaviour that otherwise looks random: why an engine quotes the wrong row of your table, why a page that clearly answers a question is passed over, why a single paragraph from deep in a long article ends up in an answer. For the short definition, see the glossary entry on chunking.
What is chunking, exactly?
Chunking is splitting a document into smaller passages before indexing it for retrieval. The usual pipeline, often called retrieval-augmented generation, has four steps:
- Split. The page is cut into chunks, usually by length, sometimes along headings or paragraphs, often with some overlap so a sentence at a boundary appears in two chunks.
- Index. Each chunk is turned into a form that can be searched, such as an embedding vector, a keyword index, or both.
- Retrieve. When a question arrives, the chunks most similar to it are pulled, from your page and many others.
- Generate. The model writes the answer from those chunks, and links the pages they came from.
The consequence is simple and important: your page competes chunk by chunk. A brilliant page with one weak section is, to the retriever, a set of strong chunks and one weak one. A mediocre page with one perfect paragraph can win on that paragraph alone.
What do the engines actually document about chunking?
Very little, and it is worth being honest about that. ChatGPT search, Perplexity, Claude and Google's AI features do not publish their chunk sizes, overlap or scoring.
The clearest documented example comes from a developer product. OpenAI's file search, part of its API, splits files into 800-token chunks with 400 tokens of overlap by default, and lets developers change both. That is not how ChatGPT search reads the web, and nobody outside OpenAI knows whether its web pipeline is similar. It is useful only as a picture of what chunking looks like when someone writes it down: fixed-length windows that ignore your headings, overlapping so nothing falls in a gap.
Google, for its part, says there is no special markup needed for its AI features: a page has to be indexed and eligible to show with a snippet. Chunking is not a setting you can change. It is something you write for.
What does chunking do to a real page?
Here is what it looked like on one of ours. On 2026-10-03, two of three ChatGPT answers to "Attensira vs Profound" reproduced a single row from a table on our old comparison page, the row that marked crawler analytics as missing for us. The page around that table had other, newer statements. The row travelled on its own, without them.
That is chunking behaviour. A table row is a compact, question-shaped piece of text: "Crawler analytics: Attensira No, Profound Yes". It scores well against a comparison question. The paragraph elsewhere on the site that said the feature had shipped was in a different chunk, on a different page, and lost.
Three things to take from it:
- Each chunk is judged without its neighbours. A caveat in the next paragraph does not travel with the claim.
- Compact, structured text scores well. That is good when it is right and bad when it is stale.
- Your own pages compete with each other. The chunk that best matches the question wins, even if another of your chunks is more accurate.
We wrote up the stale-page side of this in freshness conflicts.
How do you write sections that survive chunking?
One question per section
Give each section a heading phrased as the question it answers, and keep the section on that question. A section that answers two questions becomes a chunk that half-matches both and fully matches neither.
Answer in the first sentence
The opening sentence of a section is the one most likely to land in the same chunk as the heading. Put the answer there. Background and nuance come after.
Repeat the subject; drop the pronouns
"It supports this on Growth" means nothing once cut out. "Attensira's AI crawler traffic is available on Growth and above" survives. Name the subject in each section's first sentence, even if it feels repetitive to a human reader.
Keep facts out of positional context
"As mentioned above", "see the table below", "the second option": these break when the section is read alone. State the fact where it is used.
Make tables self-explaining
A table split across chunks can lose its header, leaving cells that mean nothing. Keep tables short, use full words in cells rather than ticks, and state the table's main finding in a sentence before it.
Do not hide text behind scripts
Text that only appears after JavaScript runs, inside tabs or click-to-expand widgets, may never be fetched at all by a crawler that does not run scripts. If it is not in the HTML, it cannot be chunked, so it cannot be quoted.