What is chunking and why does it affect my page?

Retrieval systems split documents into passages of a few hundred tokens before indexing them, because a whole page is too coarse to match against a question. Your page is retrieved one passage at a time, so a passage that only makes sense in sequence is a passage that fails on its own.

How do I write a section that survives extraction?

Name the subject in the first sentence rather than referring back to it, keep the section under a heading that describes it, and avoid opening with a pronoun or a connective such as however or as a result. The test is to read the section with everything above it deleted.

Does this mean I should repeat myself in every section?

No, and the overcorrection reads badly. Restating the subject once per section is normal writing. Restating the whole context in every paragraph produces prose nobody wants to read, which costs more with human readers than it gains with machines.

Content segmentation: writing sections that can be quoted alone

Your page is unlikely to be read as a page. Before a model sees it, a retrieval system usually cuts it into pieces of a few hundred words and stores them separately, and the piece that gets retrieved arrives without the paragraph that came before it.

The short answer

Before a model reads your page it is cut into passages, and the cuts land at heading boundaries and paragraph breaks rather than where your argument divides. A section that survives extraction states its own subject, resolves its own pronouns and does not depend on the paragraph above it to make sense.

ImportantAudit check · Content segmentation

What happens to a page before a model reads it

The step nobody writing for the web thinks about is chunking. A document goes in, a set of fragments comes out, each stored with its own embedding, and a question retrieves the fragments closest to it rather than the document. The model then answers from those fragments.

Which produces a constraint that has no equivalent in classic SEO: every section of your page has to make sense in isolation, because isolation is the condition it will be read in. A paragraph opening with "this is why it matters" is fine on the page and useless as a fragment, since the thing it refers to is in a different fragment now.

Where the cuts are made

Chunkers differ, but almost all of them prefer structural boundaries and fall back to arbitrary ones. In rough order of preference: heading boundaries, then paragraph breaks, then sentence breaks, then a character count with an overlap window.

That order is the whole practical lesson. If your page has headings at sensible intervals, the cuts land where you would have put them. If it does not, the cuts land at whatever character offset the window reached, which routinely means a fragment that begins mid-idea and ends mid-sentence.

You do not control the chunker. You control whether it has anything better than arithmetic to work with.

This is why heading structure shows up as an AI concern rather than only an accessibility one. Headings are the boundary markers, and a page with one heading and four thousand words offers none.

What a passing section looks like

Sections of a size that survives chunking, marked by real headings, containing text that does not depend on its neighbours. In practice: how much prose sits between structural boundaries, and whether the sections open with references pointing outside themselves.

It is not a word-count check and it does not reward more headings. Six headings over eight hundred words is worse than three, because fragments too small to contain a complete thought are their own failure: the retrieval system gets a heading and two sentences that answer nothing.

Writing a section that survives being cut out

Three habits do almost all of it, and the first one is the one people resist because it reads as repetitive when you scan the whole page.

Name the subject in the section, not just in the heading. Headings are frequently dropped or stored separately, so a section whose only mention of the subject is in its heading has no mention of it at all.

segments.md

## Pressure rating

<!-- Depends entirely on the section above. As a
     fragment: what is rated to 12 bar? -->
It is rated to 12 bar, which is higher than the
previous generation. This matters because of the
issue described above.

<!-- Stands alone. Names the subject, states the
     figure, and explains the consequence without
     pointing anywhere. -->
The 40mm stainless widget is rated to 12 bar,
up from 9 bar in the previous generation. That
covers mains-pressure systems without a reducer,
which is the usual reason to specify it.

Resolve pronouns and demonstratives at section boundaries."It", "this", "the above" and "as mentioned" are all references across a boundary you cannot see. Inside a paragraph they are fine; as the first words of a section they are a broken link.

Keep sections between roughly two hundred and five hundred words. Not a rule, a range that tends to survive: long enough to contain a complete answer, short enough not to be cut in half.

Testing a section in isolation

The manual version is quick and unpleasant in a useful way: copy one section out of your page, paste it somewhere on its own, and read it. Anything you have to scroll back for is the dependency this check is about.

This one needs the full crawl

Section boundaries are judged per page across the whole site, because the useful finding is which pages have no usable boundaries at all rather than whether one page does.

The instant search on this site audits a single page, so rather than show you a verdict it cannot support, this guide sends you to the place the check actually runs.

Run the full audit, free

What this does not mean

It does not mean writing each section as a self-contained article.

The failure mode of taking this too far is a page that repeats its own context five times and reads like a form. A human is still the primary audience, and a page that annoys them to please a chunker has traded something real for something speculative. The target is a section that can be understood alone, not one that assumes the reader has never been to the page.

There is a second overcorrection worth naming: splitting one good page into six thin ones so each is "focused". That produces six pages competing with each other and none of them substantial, which fails an entirely different check. Segmentation is about structure within a page, not about splitting pages up.

Why imperfect extraction is not Critical

Because the cost is quality of extraction rather than presence. A badly segmented page is still fetched, still indexed, still readable, and will still get quoted. It just gets quoted less accurately, and a fragment that reads as incoherent is one a system is less likely to surface next time.

It sits above Refinement because the fix is genuinely editorial work with a real payoff, and below Critical because nothing about it stops your content being used. Compare text extractability, where the content is not there at all. The scale separates absent from imperfect deliberately.

Does this affect AI search?

It is close to being an AI-only check, which makes it unusual on this list. Chunking is a property of retrieval pipelines, and a classic search engine indexing your page does not cut it up this way.

There is a real search benefit, though it arrives by a different route: the same discipline that makes a section survive chunking is what makes it eligible to be lifted as a featured snippet. A passage that stands alone is a passage an engine can quote without editing, and that has been true since long before retrieval pipelines existed.

So the honest framing is that this is a bet on how machines read, which happens to align with a much older and better-established preference for self-contained passages. That alignment is the reason to do the work now rather than waiting to see how retrieval evolves.

Confirming the sections hold

Read the page as a list of its own sections, in isolation, ideally not on the day you wrote it. The specific thing to look for is the first sentence of each: a section beginning with a reference to something not named inside it is the defect, and it is almost always fixable by rewriting that one sentence.

Re-crawl for the structural half, the section sizes and the heading intervals. The dependency half is a judgement no crawler makes, which is worth saying plainly rather than implying a tool will catch it for you.

Questions this check raises

What is chunking and why does it affect my page?: Retrieval systems split documents into passages of a few hundred tokens before indexing them, because a whole page is too coarse to match against a question. Your page is retrieved one passage at a time, so a passage that only makes sense in sequence is a passage that fails on its own.
How do I write a section that survives extraction?: Name the subject in the first sentence rather than referring back to it, keep the section under a heading that describes it, and avoid opening with a pronoun or a connective such as however or as a result. The test is to read the section with everything above it deleted.
Does this mean I should repeat myself in every section?: No, and the overcorrection reads badly. Restating the subject once per section is normal writing. Restating the whole context in every paragraph produces prose nobody wants to read, which costs more with human readers than it gains with machines.