Passage-Level Citability: Write Sections AI Can Quote

A passage is what an AI answer actually uses: a section, a paragraph, sometimes a single sentence, lifted out of your page and judged with no memory of what surrounded it. Passage-level citability is the property of surviving that lift - being quotable, attributable and correct on your own, out of context. I’ve written about what earns citations from AI search engines at the strategy level: authority, freshness, the measurement problem. This post is the construction manual, one level down. It is about how I build the individual sections, because in my own citation checks the difference between cited and invisible is usually not the quality of the page. It is whether the winning passage could stand alone.

The passage became the unit before AI answers arrived

This shift predates ChatGPT, which is worth knowing because it means the work pays in classic search too. In October 2020, Google announced it would rank by “understanding passages in addition to the relevancy of the overall page” - finding, in its words, the needle-in-the-haystack information - and estimated the change would improve 7 percent of search queries globally. That was Google admitting the page was too coarse a unit for specific questions.

Answer engines finished the thought. A retrieval system chunks your page, embeds the chunks, pulls the few that match a query, and writes an answer from what survived. Your page is never in the room when the decision is made. Only the passages are, competing one at a time against passages from every other site. So the question I ask of every section I write has changed from “does this serve the page?” to “if this section were the only thing retrieved, would it work?”

The standalone test

My editing rule for citability fits in one sentence: every ## section must make sense, and make its claim, to a reader who has seen nothing else on the page. In practice that means checking four things, and they are mundane enough that most content fails them by accident rather than by intention:

The section answers in its first two sentences. Not winds up, answers. A retrieval system scoring a chunk against a query rewards the chunk whose opening is the answer. Everything after the second sentence is elaboration for the human who clicked - the machine has often stopped reading by then, and so, per the same logic, has the skimmer.

The subject is restated, not referred to. Pronouns and back-references are how passages die. “This makes it slower” is unquotable, because outside the page nobody knows what this or it is. “Pinning headlines makes a responsive search ad less flexible” quotes cleanly. I write headings the same way - “How data-driven attribution works,” not “How it works” - because the heading often travels with the chunk and has to carry the context alone.

The claim and its evidence share a paragraph. If the number lives three paragraphs from the statement it supports, a chunk boundary will separate them, and the retrieved passage becomes an unsupported assertion - precisely the kind of passage a generation model is trained to pass over in favor of one with the receipt attached. One claim, its number, and its source, inside one paragraph, is the most citable shape I know.

The stat sentence is self-contained. When I cite a figure, I build one sentence carrying the number, the source and the context together - “in Pew Research Center’s browsing study, users clicked a result on 8% of visits with an AI summary versus 15% without” - because that sentence is the unit an answer engine wants to lift, attribution included.

None of this is exotic. It is the discipline of writing every section as if it might be someone’s first and only contact with your thinking - which, in AI search, it now literally is.

Structure that marks the chunk boundaries

The second half of the construction job is making your page chunk well, since you don’t control the knife. What I rely on:

One question per section. When a section answers two questions, a chunk that captures one half dilutes the other. Splitting into two headed sections gives each answer its own retrievable unit.
Headings as claims or questions. The heading travels with the chunk and doubles as its label. “Costs” is a weak label. “What data-driven attribution costs you in auditability” tells both the retriever and the reader what the passage contains.
Lists for genuinely list-shaped answers. “What are the three attribution models?” retrieves best from an actual list of three items, each item restating enough context to stand alone. I resist the urge to convert arguments into lists, though - a list of sentence fragments is compact but unquotable.
Tables for lookups, with a prose summary. Models parse tables, but a table cell is the least standalone text on a page. When a table carries the answer, I put the one-sentence version in prose above it.
Definitions where terms first appear. A crisp one-sentence definition is the single most retrievable passage type there is, because definition queries are endless. If your page owns a term, give the term a sentence that could be quoted as the definition.

What Google says about markup, and what to do instead

Here is where I save you from a category of vendor pitch. Google’s documentation on AI features is unusually direct: you don’t need machine-readable AI files, special markup, or any special schema.org structured data to appear in AI Overviews or AI Mode - the same foundational SEO practices apply as for Search overall. There is no citability meta tag. Anyone selling one is selling decoration.

The markup that does matter is the ordinary kind, for unglamorous reasons. Clean heading hierarchy shapes how your page chunks. Standard schema types help systems resolve who is making the claims - the authorship and freshness signals I covered in the citation playbook, and generating that markup with ChatGPT is a safe afternoon job if you validate it. And the same doc points at the controls for the opposite problem: nosnippet, data-nosnippet, max-snippet and noindex limit what Search, including its AI features, shows from your pages. I mention them because a nosnippet left over from an old CMS template is a self-inflicted disqualification I have actually seen in the wild.

The other plumbing check is crawler access, and each engine documents its own: OpenAI’s OAI-SearchBot surfaces sites in ChatGPT search, and blocking it removes you from those answers - it is distinct from GPTBot, the training crawler, so you can allow one without the other. Perplexity documents the same split for PerplexityBot. The full setups live in my ChatGPT search and Perplexity guides. The point here is only that no passage gets cited from a page the crawler never fetched.

Why the passage work is worth it

Two numbers frame the payoff for me. The cost of not being the citation keeps rising: in Pew Research Center’s study of real browsing behavior, users who hit an AI summary clicked a traditional result on just 8% of visits, versus 15% without one. When the answer engine absorbs the click, being quoted inside the answer is most of the visibility left on the table.

And the passage-level levers demonstrably move citation odds. The Princeton and Georgia Tech GEO study tested content modifications against generative engines and found visibility gains of up to 40% - with the effective tactics being exactly the ones above: adding citations, quotations and statistics to the passages themselves. Structure and evidence, not tricks.

The passage audit I actually run

When I audit a page - mine or a client’s - for citability, the workflow is thirty minutes and needs no tools beyond the page itself:

Isolate each ## section and read it as if it arrived alone in an answer. No peeking at the intro.
Ask the question the section should win, phrased the way a person asks an assistant. If I can’t phrase that question, the section has no retrieval target, which usually means it exists for the page’s flow rather than for any query - fine, but then it is not citation inventory.
Check the first two sentences for the answer. Move it up if it is buried.
Hunt orphaned pronouns - this, it, they, that approach - and replace each with its noun.
Reunite claims with evidence that drifted into neighboring paragraphs.
Verify the section could be quoted without embarrassment - because the failure mode of citability work is making bad claims easier to retrieve. A confidently structured passage with a wrong number in it is now a widely distributed wrong number, which is why the verification discipline is part of the same job, not a separate one.

Then I check whether any of it worked, using the citation-tracking loop from the measurement side of this category - and I hold my expectations to what that post says about small numbers.

FAQ

What is passage-level citability? Passage-level citability is how well an individual section of a page can be retrieved, quoted and attributed by an AI answer engine on its own, without the rest of the page. It depends on the section answering a specific question early, restating its subject explicitly, and carrying its evidence with it - not on any special markup.

Do I need llms.txt or AI-specific schema for this? No. Google states outright that no AI text files, special markup or special structured data are needed for its AI features, and no major answer engine documents a citability tag either. The controllable inputs are structure, evidence and crawler access.

How long should a citable passage be? I aim for sections of roughly 75 to 250 words: long enough to state a claim with its support, short enough that a chunk likely captures the whole thing. The stronger constraint is one question per section - a 400-word section answering one question beats two questions squeezed into 150 words.

Does this conflict with writing for humans? Less than the framing suggests, and this is my favorite thing about the whole discipline. Answer-first sections, restated subjects, evidence next to claims - that is just clear technical writing, the kind skimming humans reward too. The one real tension is redundancy: standalone sections repeat context a linear reader already has. I accept that cost deliberately, because the linear reader is now the minority of first contacts.

This post is part of the AI SEO and GEO guides. The strategy layer is in how to get cited by AI search engines, the per-engine plumbing in the ChatGPT and Perplexity guides, and the wider practice starts at the complete guide to AI marketing.