How to Write a Passage an AI Will Quote

Answer passage

A short, self-contained block of text that answers one specific question and keeps its meaning when lifted out of the page it sits on.

Also known as: Answer block, Quotable passage, Extractable passage

The test is mechanical rather than stylistic. Delete the heading, delete the paragraph before, and read the sentence cold. If a reader cannot tell what it is about, an extraction system cannot either, and a competing page with a clearer sentence gets used instead.

This is a different skill from writing well. Good editorial prose builds meaning across paragraphs, carries subjects forward with pronouns, and saves the conclusion for the end. Every one of those habits produces text that breaks when a single span is pulled out of it.

That number is unusually solid because it comes from the API documentation rather than from inference. Anthropic's web search tool returns each result with a url, a title, a page age and encrypted content, and every citation carries a cited_text field capped at 150 characters. There is no ambiguity about the size of the window on that engine.

A hundred and fifty characters is roughly 20 to 25 plain English words. So the practical unit is a single declarative sentence, and the useful discipline is writing that sentence first, then letting the supporting detail follow in a separate sentence where it cannot lengthen the quotable one.

Two widely repeated figures contradict this and neither has a source. The claim that AI Overviews extract passages of 134 to 167 words, and the related claim that 62% of featured content sits between 100 and 300 words, appear on many agency pages with no method, sample or date attached. Semrush, across 11,882 prompts and 304,805 cited URLs, found passage length had a negligible association with citation, which is the opposite of a length target.

A fifth failure is subtler and very common in company blogs: the sentence is fine but the page has no claim worth quoting. No amount of rewriting fixes an article whose every statement is available on a hundred other pages, because a model with a hundred equivalent sources has no reason to attribute any of them.

Put one passage at the top of the page answering the page title, then one under every H2 answering that H2. That gives an engine a clean candidate for the head question and for each narrower question it fans out to, which is how retrieval actually works: Google documents query fan-out as issuing multiple related searches across subtopics and data sources.

A figure circulating widely says 44.2% of AI citations come from the first 30% of the page. There is no primary publication behind it, and it should not be quoted. The underlying instinct is right, front-loaded answers get extracted more often, and that instinct is supported without needing an invented number to prop it up.

The one thing to avoid is a summary block that promises rather than answers. "In this guide we will cover how the Ad Library works" is the single most common opening sentence on question-shaped pages, and it is unquotable by construction, because it describes the page instead of answering the question.

This is where most AEO projects go wrong, and the evidence is unusually direct. Vishwakarma, Kumar and Jamidar, "What Gets Cited: Competitive GEO in AI Answer Engines", ACM SIGIR 2026, ran 252,000 trials across six models and 18 content factors. Formatting changes, structural changes and paragraph organisation were all negligible. Topic relevance, explicit concrete specifics, recency and retrieved position were unanimous gatekeepers.

Google says something compatible from the other direction: there is no requirement to break content into tiny pieces for AI to better understand it. So a quarter spent chunking a content library buys close to nothing, while the same quarter spent adding measured numbers and dates to those pages targets the factors that were significant.

The right reading is that a quotable passage is the container and a concrete specific is the payload. Rewriting the container matters when there is something in it worth quoting, and does nothing when there is not. Oppira pairs each rewritten passage with a figure we measured ourselves, because a sentence nobody else can copy is what turns a paraphrase into a citation.

So sequence the work. Find the ten pages where you know something specific, write real answer passages on those, and leave the generic pages alone until they have something to say.