Chunking

What is actually known about how pages get chunked?

Less than the advice implies. The chunk sizes inside ChatGPT, Claude and Perplexity are not published, so any specific token count you read for them is somebody’s guess, and Google publishes none for its own surfaces either. Anyone quoting you an exact chunk size is quoting a reconstruction.

What is measured is the outcome rather than the mechanism, and it points the same way. Ahrefs’ analysis of 560,346 AI Overviews found the average cited page at 1,282 words, 53.4% of cited pages under 1,000, and a correlation between word count and citation position of 0.04. Pages are cited for a passage rather than for their size, which is the practical content of the word “chunking”: the thing being selected is smaller than the thing you published.

Does this mean you should write for the chunker?

Google’s own position is no, and the practitioners who instrumented retrieval say yes, and the disagreement is smaller than it looks. Nobody credible recommends shredding a page into fragments; Google states multi-topic pages are understood fine. What the measurements support is ordinary discipline: one subject per section, the answer at the top of it, and the entity named rather than left as “it”.

Where the two positions genuinely differ is on intent, and intent is not observable in a page. A section written to answer a reader’s question and a section written to be a chunk are identical in the HTML, so the argument settles nothing you can act on, and the advice collapses to the same thing whichever side you find more convincing.

The recommendation, which is a position and could be wrong: write the section so it would survive being quoted with the rest of the page deleted. If it would not, the problem is usually that the argument is split across a heading boundary, and moving one sentence fixes it.

Where does chunking make a page fail silently?

In pronouns and in tables. A section opening “It tracks five assistants” loses its subject the moment it is lifted out, because the sentence that named the subject stayed behind on the page. A table without a caption fails the same way and more completely, arriving as a grid of words with nothing saying what they are of, which is why every table in this glossary carries one. Neither failure shows up in any report: the page looks fine, and the passage is simply never the one selected.

Related terms

• GroundingGrounding is the step that ties a model’s answer to sources it retrieved, so the reply can cite them rather than assert them.
• EmbeddingAn embedding is a list of numbers representing a piece of text, positioned so that texts with similar meaning sit close together.
• RAG (Retrieval-Augmented Generation)RAG, or retrieval-augmented generation, is the technique of fetching documents at question time and giving them to a language model to answer from, rather than relying on what the model learned in training.
• Query Fan-OutQuery fan-out is Google’s own term for how AI Overviews and AI Mode answer a question: the engine breaks the question into subtopics and issues many searches at once, then writes one answer from the results.