How to Write Content AI Actually Quotes (The Passage-Level Guide)
By the alphaa team — we run AI-visibility scans across thousands of small businesses and read the raw passages engines quote back. Last updated 30 July 2026.
Short answer: AI engines almost never quote a whole page — they quote a passage, a few hundred words of text that was retrieved on its own and had to make sense without the rest of the article around it. So the unit of AEO writing is not the page, it is the self-contained paragraph. Write each section so that if you cut it out and handed it to a stranger with no title, no headline, and no surrounding context, it would still answer one specific question completely and name the business it is about.
Why passages, not pages, get cited
A passage (or "chunk") is the slice of your page that a retrieval system stores and scores as one unit. Retrieval-augmented generation — the process most assistants use to answer with live web content — works in three steps, and each one operates below page level:
- Chunking. A page is split into segments, usually a few hundred words each, often along heading and paragraph boundaries.
- Embedding and matching. Each chunk is converted into a vector — a numeric representation of its meaning — and matched against a vector of the user's question. Chunks that are semantically close to the question are pulled.
- Synthesis. The model reads the handful of retrieved chunks and writes an answer, citing the ones it leaned on.
The consequence is blunt: your best paragraph competes on its own. It does not inherit credibility from your headline, your hero section, or the paragraph three screens above it. If your explanation of emergency call-out fees only makes sense after reading the intro, the chunk containing it reads as vague and gets passed over for a competitor's chunk that stands alone.
What makes a passage liftable?
A liftable passage is one a model can quote verbatim without editing, hedging, or guessing what it refers to. In practice, five properties do most of the work:
- It answers one question. One question per section. Two questions in a paragraph halves its match score against both.
- It leads with the answer. The first sentence is the claim; the rest is support. A model that quotes only your first sentence should still have said something true and useful.
- It re-states its subject. Names the business, the service, and the place, rather than "we" and "our team." Pronouns break when the chunk travels alone.
- It contains checkable specifics. Numbers, ranges, timeframes, conditions, qualifications — the things a model can attribute.
- It is bounded. Roughly 40–120 words. Long enough to be complete, short enough to be quoted whole.
A worked example: the same fact, written two ways
Here is a paragraph from a typical service page:
We pride ourselves on fast response times and transparent pricing. Our experienced team is available around the clock, and we always aim to get to you as quickly as possible — because we know how stressful these situations can be.
Extract that chunk and it says nothing checkable. Who is "we"? Where? How fast is fast? What price? A model asked "who does emergency plumbing in Tucson at night, and what does it cost?" has nothing to quote. Now the same facts, written to be lifted:
Cardinal Plumbing is an emergency plumber serving Tucson and Oro Valley, Arizona, available 24 hours a day including weekends and public holidays. Overnight call-outs between 8pm and 6am carry a flat $145 call-out fee, quoted before a technician is dispatched, with labour billed separately at an hourly rate agreed on the call. Typical arrival time within the Tucson city limits is 60 to 90 minutes.
Same business, same truth, but the second version survives extraction. It names the entity, the service, the geography, the hours, the fee, the conditions, and the caveat — and every one of those is a term a question might match against. This is also why entity clarity and passage writing reinforce each other: repeating your name and location inside the chunk is not clumsy, it is what keeps the chunk attributable.
One caution: only write specifics you will keep current. A stale price quoted confidently by an AI engine is worse than no price, because a customer arrives expecting it. Put a review date on any page carrying numbers.
How to structure a page for chunking
Chunk boundaries usually fall on headings, so headings are how you control what gets stored together. The practical rules:
- Make every H2 and H3 a real question or a precise noun phrase. "How much does an emergency call-out cost at night?" retrieves. "Our Promise" does not.
- Put a definitional sentence directly under each heading. No throat-clearing, no "in this section we'll look at." The first sentence after a heading is the sentence most likely to be quoted.
- Keep sections short and parallel. Roughly one screen each. A 900-word section under one heading gets split arbitrarily, and you lose control of where.
- Do not bury facts in tables or images alone. A table is fine, but state the key number in a sentence too — sentences survive chunking more reliably than layout does.
- Add a short FAQ block of genuine questions. A question heading with a two-to-four sentence answer is the single most extractable shape on the web, because it mirrors the query-answer pair the retriever is matching.
A 30-minute rewrite workflow you can run today
- Pick one page that matters — the service or location page you would most want named in an AI answer.
- List the five questions a customer actually asks before buying. Use the wording they use, not your marketing wording.
- Turn each into an H2. If the page has no section answering one of the five, that gap is your highest-value edit.
- Rewrite the first sentence under each H2 so it answers the heading outright and names the business and place.
- Run the isolation test. Copy each section into a blank document with the heading removed. Read it cold. If you cannot tell who it is about or what it commits to, it fails — fix it.
- Check it is in the HTML. None of this counts if the text is client-rendered; see why AI can't read your JavaScript site for the one-command check.
The isolation test is the whole method in one step. It is the closest thing to seeing your page the way a retriever sees it, and it costs nothing.
Questions people ask about writing for AI extraction
Does this mean I should write shorter articles?
No — write shorter sections. Length at the page level is fine and often helps, because more sections means more chances to match a question. What hurts is one undifferentiated wall of prose, where no single chunk resolves cleanly.
Should I stuff keywords into every paragraph?
No. Matching is semantic, not literal — the retriever compares meaning vectors, so a naturally worded sentence about overnight plumbing call-outs matches a query about night-time emergency plumbers without containing those exact words. Repeat your entity (name, service, place) for attribution; do not repeat keywords for their own sake.
Is this different from writing for Google?
It overlaps heavily but is not identical. Google has ranked passages for years, so the work compounds. The difference is the failure mode: a weak page can still rank on domain strength and get a click, whereas a weak passage simply is not quoted, and there is no second chance further down the answer. For the broader contrast see AEO vs SEO and why old-school agencies fall short.
Will good passages guarantee I get cited?
No, and be wary of anyone who says otherwise. Extractable writing makes you eligible and raises the odds; whether you are chosen also depends on third-party evidence, reviews, competing sources, and the model's own variation between runs — see why AI answers change every time. Passage writing is the part of the system you fully control, which is exactly why it is worth doing first.
The bottom line
Stop optimising pages and start optimising paragraphs. Every section should answer one question, lead with the answer, name the business and place, carry a specific you will keep accurate, and stand up when read in isolation. That is not a trick for machines — it is also how a hurried human reads. The rewrite costs an afternoon per page and it is the highest-leverage AEO work most businesses have not done.