Passage-Level Content Structure for Search and AI Retrieval
Google runs a passage ranking system and retrieval systems index chunks. Here is how to structure content for that without writing artificially for machines.
The unit of selection has been drifting downwards for years. A search engine used to return documents; it now identifies relevant sections within them. An AI answer engine retrieves chunks of text and feeds them to a model. In both cases the thing being assessed is smaller than the page. That has real implications for how a page should be organised — and it is also the premise behind a large amount of advice that amounts to writing badly on purpose. This article covers the structural practices that hold up, what Google actually says, and where the line is.
What Google says about passages
Google's guide to Search ranking systems lists passage ranking among its systems and describes it plainly: "Passage ranking is an AI system we use to identify individual sections or 'passages' of a web page to better understand how relevant a page is to a search."
Two things are worth noting about that sentence. It says passages are used to understand the relevance of a page — it is a page-level ranking system that reads at passage granularity, not a system that indexes passages as independent documents. And it is documented as a Search ranking system, with no statement connecting it to AI Overviews or AI Mode.
That is the whole of the first-party evidence, and it is OFFICIAL PROVIDER GUIDANCE for exactly what it says and nothing more. Everything else in this article about how AI answer engines handle passages is inference from how retrieval systems are generally built.
What retrieval systems generally do
Systems that ground a generated answer on documents almost always split those documents into chunks, embed or index the chunks, and score chunks against the query. The retrieved chunk — not the page — is what reaches the model's context.
This is STRONG EVIDENCE as a general information-retrieval principle, drawn from the published literature on retrieval-augmented generation, and it is not provider documentation. No provider states how they chunk your content, at what size, on what boundaries, or whether they preserve heading context. Any specific number you have read about ideal passage length is someone's guess.
The consequences that follow from the general design, all at EMERGING PRACTICE as practical advice:
- A section that answers its own heading is usable as a retrieved chunk. A section that only makes sense after reading the previous three is not.
- Pronouns and back-references — "it", "this approach", "as described above", "the company" — lose their referent when the passage is separated from the page.
- A very long section either becomes one chunk that matches many queries weakly, or several chunks split at an arbitrary point mid-argument.
- A very short section carries too little to be worth retrieving.
The practices that hold up
None of these are novel. They are what a careful editor does, which is the point: the structural properties that make content retrievable are the ones that make it readable by a person who arrives mid-page from a search result or a table of contents.
Headings that name the section's subject
A heading is a label for the passage beneath it. "Overview", "Details", "More information" and "The bottom line" label nothing. "How fee status is assessed" and "Retry behaviour on rate-limit errors" name a subject a reader — or a retrieval system — can match against a question.
Google's starter guide is worth reading carefully here, because it is more relaxed about headings than most SEO advice: it says semantic heading order helps screen readers, but "from Google Search perspective, it doesn't matter if you're using them out of order", and that there is no ideal number of headings, though a page with an excessive number is worth reconsidering. So the case for descriptive headings is a readability and accessibility case, plus a plausible retrieval case — not a Google requirement.
OFFICIAL PROVIDER GUIDANCE that headings help users and that order is not a Search issue. EMERGING PRACTICE for the retrieval benefit of descriptive heading text.
The answer before the elaboration
If the section answers a question, put the answer in the first sentence or two and then explain, qualify and give examples. This is inverted-pyramid writing and it long predates any of this.
Two independent reasons it works. A reader scanning a long page decides in a few seconds whether this section is the one they want. And a chunk that begins with the answer contains the answer regardless of where the chunk boundary falls.
EMERGING PRACTICE as a retrieval lever. Settled editorial practice on its own terms.
Definitions near the top
If a page introduces a term, define it in plain language before using it in analysis. One sentence, using the term itself, without requiring the surrounding paragraphs:
A content orphan is an indexable page that no other page's body content links to, even though a navigation menu or an archive template may link to it.
That sentence works alone. Compare: "As the previous section showed, these are the pages that fall through the gap." Same information, useless in isolation.
Named entities instead of pronouns at section starts
Open each section by naming what it is about rather than referring back. Not "It also affects how the deadline is calculated" but "Fee status also affects how the application deadline is calculated." This is the practice entity clarity describes at page level, applied at passage level.
EMERGING PRACTICE. It also happens to be clearer prose.
Tables for comparisons and enumerable facts
When the content is "these three options differ along these four dimensions", a table states it once, unambiguously, in a form that survives extraction. Prose that walks through twelve cells in sequence does not.
A caution: a table is only good structure if it is a real HTML <table> with header cells. A grid of styled <div> elements built by a page builder looks like a table and is not one. Which brings us to the constraint that outranks everything above.
The content has to be in text
Google's AI features guidance asks that important content be available in textual form. A passage inside an image, a PDF nobody links to, a tab that loads its content on demand, or a client-side render is not reliably a passage at all.
OFFICIAL PROVIDER GUIDANCE, and the highest-value item in this article. On WordPress the usual culprits are accordion and tab blocks that inject content on interaction, and infographics carrying facts that appear nowhere in the page text.
What SiteRank measures, and what it deliberately does not
SiteRank's GEO audit includes a passage retrievability check. It defines a passage as the text between two headings, and flags indexable documents containing a section of 450 words or more without a subheading, reporting the count of long sections and the length of the longest.
Three deliberate limits on that rule are worth stating because they are the difference between a measurement and an opinion:
- 450 is a threshold, not an ideal. It is a point at which a section is long enough that a reader benefits from a subheading. There is no evidence for an optimal passage length and the plugin does not claim one.
- The finding is tagged as a general retrieval principle, not documented provider behaviour. The rule's own explanatory text says so: breaking a long section up does not change the content, only its addressability.
- It measures structure, not quality. A long section can be excellent. The check says a section is long, gives the number, and leaves the judgement to the person who wrote it.
There is no check for sentence length, reading level, "AI-friendliness" or question-shaped headings, because there is no evidence to support one and inventing thresholds would produce confident nonsense.
The line: structure versus performance
The failure mode of this whole subject is content written to look like it was structured for a machine. It is easy to spot and it makes pages worse:
| Anti-pattern | Why it fails |
|---|---|
| Every heading rewritten as a question, including ones nobody asks | Headings stop labelling the content and start advertising; a reader scanning the page learns less |
| A "quick answer" box at the top of every section, duplicating the section | Repetition without information gain; the page gets longer and says the same amount |
| Sentences chopped short because "AI prefers short sentences" | No provider documents any such preference. The prose degrades for nothing |
| An FAQ block bolted onto every page | Google's guidance is about helpful content, and questions nobody asks are not helpful. This also risks the scaled-content problem when applied across a site |
| A fixed word count per section | There is no published optimal length. A section is as long as its subject |
| Keyword-loaded headings | Headings should name the section's subject in the reader's words |
| Facts stated only in a summary table, never in the prose | A table with no supporting text is a claim without an argument |
The test that separates good structure from performance: would a competent editor, with no knowledge of retrieval systems, make this change? Descriptive headings, answers before elaboration, defined terms, self-contained sections, tables for comparisons — yes to all. Chopped sentences, question-shaped headings for questions nobody asks, and duplicated summary blocks — no.
Google states there are no additional requirements or special optimisations necessary to appear in its AI features, and specifically that no new machine-readable files, AI text files or markup are needed. Anything that fails the editor test is being done for a mechanism that Google says does not exist.
Evidence summary
| Practice | Tier | Basis |
|---|---|---|
| Put important content in server-rendered text | OFFICIAL PROVIDER GUIDANCE | Google AI features guidance |
| Headings help users; out-of-order headings are not a Search problem | OFFICIAL PROVIDER GUIDANCE | Google SEO starter guide |
| Google identifies individual passages to assess a page's relevance | OFFICIAL PROVIDER GUIDANCE | Google ranking systems guide, passage ranking |
| Retrieval systems generally operate on chunks, so self-contained passages are more usable | STRONG EVIDENCE | General IR/RAG literature; not provider-specific |
| Descriptive headings improve retrievability | EMERGING PRACTICE | Plausible mechanism, widely practised, unconfirmed |
| Answer-first sections improve extractability | EMERGING PRACTICE | Widely practised; some vendor-run studies |
| Naming entities instead of pronouns at section starts | EMERGING PRACTICE | Plausible for chunk independence, unconfirmed |
| Tables for enumerable facts aid extraction | EMERGING PRACTICE | Plausible, unconfirmed |
| A specific optimal passage or sentence length exists | HYPOTHESIS | No published basis. Do not act on it |
| Question-shaped headings cause citation | HYPOTHESIS | No provider documents this |
What not to assume
- That passage ranking is an AI-features system. Google documents it as a Search ranking system. Do not extend the claim.
- That you know how a provider chunks your page. Nobody outside the provider does. Design for the general principle, not for a guessed chunk size.
- That a retrievable passage will be retrieved. Structure affects usability of a passage as context. It does not control selection, and selection does not control citation.
- That restructuring an existing page is worth it everywhere. Do it where a page is genuinely hard to read in sections. A wholesale rewrite of a working site for an unconfirmed mechanism is a bad trade.
- That structural changes can be credited with a visibility change. Model updates, index updates and sampling noise move these numbers too. Reporting association rather than cause is a discipline, not a hedge — the methodology page sets out what a defensible comparison needs.
Key takeaways
- Google documents passage ranking as a Search system that identifies sections to judge a page's relevance. That is the only first-party statement in this area.
- Retrieval systems generally index chunks; self-contained passages are more usable as context. Strong as an IR principle, unconfirmed by any provider.
- The practices that hold up — descriptive headings, answers first, defined terms, named entities, tables — are ordinary good editing.
- Content that is not in server-rendered text is not a passage at all. Fix that before anything else.
- SiteRank flags sections of 450 words or more without a subheading, and states that 450 is a threshold rather than an ideal.
- If a competent editor with no knowledge of retrieval would not make the change, do not make it.
Official sources & further reading
- A guide to Google Search ranking systems — Google Search Central
- AI features and your website — Google Search Central
- SEO starter guide — Google Search Central
- Creating helpful, reliable, people-first content — Google Search Central