How to structure a page so an AI engine can quote it

An AI engine does not cite your page. It cites a paragraph from your page, stripped of everything around it, dropped into an answer next to a paragraph from someone else's page.

That is the whole model, and once you internalize it, most content advice reorganizes itself. Your intro does not matter, because it will never be retrieved. Your beautiful three-section build to a conclusion does not matter, because section two will be evaluated alone by a system that has never seen section one. What matters is whether each individual chunk of your page can stand up by itself, answer something, and be safely quoted by a model that will be blamed if it is wrong.

The stakes are not theoretical. When Google shows an AI summary, only 8% of users click any search result, versus 15% without one, and just 1% click a link inside the summary. The traffic is going to the answer, not the page. Being in the answer is the game.

Why AI engines cite passages, not pages

Because that is how retrieval works. Before a model writes an answer, a retrieval layer splits candidate documents into chunks of roughly 200 to 500 tokens, embeds each chunk as a vector, and matches the user's question against those vectors. The model then sees a handful of winning chunks, not your page.

Three consequences follow directly, and every technique in this post is a corollary of one of them:

• Your chunk competes alone. The quality of the rest of your page is invisible at match time. A great page with one bad chunk loses that chunk's query to a mediocre page with one great chunk.
• Chunk boundaries are not your boundaries. The splitter does not know where your argument begins. It will happily cut mid-section. Content that depends on setup gets cut away from its setup.
• The model has to be willing to quote it. A generative engine is a nervous quoter. It prefers passages that are unambiguous, attributable and hard to misread, because a hedged or context-dependent passage is a hallucination risk. Ambiguity is a filter, and you fail it silently.

What makes a passage liftable?

A liftable passage answers a specific question in its first sentence, names its own subject, carries its own numbers, and does not refer to anything outside itself. Run any paragraph through those four and you will usually find it fails at least one.

The single best diagnostic is what we call the lift-out test: delete everything on the page except this one paragraph. Would a reader who sees only this paragraph get a correct, complete, attributable answer? If they would be confused, or would need to know what "this approach" or "the tool" or "as mentioned above" refers to, the paragraph is unliftable and an engine will pass over it.

Here is the checklist we actually run against pages:

That last row is underrated. Engines cite sources that themselves cite sources. A paragraph containing a linked, dated statistic is dramatically safer to quote than the same paragraph with the number floating free, because the engine can pass the attribution through to its own answer.

A before/after rewrite

Abstract advice is worthless here, so here is a real paragraph from an early draft of our own pricing page, and the rewrite.

Before:

When considering the various options available in this space, it's important to think about what you're actually paying for. Many of the tools out there will advertise an attractive entry price, but once you dig in, you find that the coverage you need is locked behind a much higher tier. This is something we felt strongly about when designing our own approach, and we made a different choice. It means you get everything from day one, which we think is fairer.

Every sentence in that fails the lift-out test. What space? Which tools? What higher tier? What different choice? What is "everything"? A retrieval system embedding this chunk gets a vector that means roughly "vague opinions about pricing fairness," which matches nothing. And even if it did match, no model would quote it, because there is not a single checkable fact in it.

After:

Why do AI visibility tools gate engines behind higher tiers?

Most AI visibility tools advertise a low entry price that covers only one or two engines, and unlock the rest at three to ten times the price. Profound starts at $99/mo but tracks ChatGPT only; its three-engine Growth plan is $399/mo. Otterly.AI's $29/mo Lite plan includes four engines, but Gemini and Google AI Mode are paid add-ons. Writesonic advertises ten-engine coverage at $79/mo, but self-serve plans track three; ten is Enterprise-only.

Spottlo includes all four of its engines — ChatGPT, Perplexity, Gemini and Google AI Overviews — on every plan, at $39/mo base plus $19/mo per additional brand, with 25 tracked prompts per brand.

Same argument. Now it is quotable. Look at what changed:

1. The heading is a question a buyer actually types.
2. The first sentence is the answer. Someone who reads only that sentence has the point.
3. Every claim carries a name and a number. "Profound starts at $99/mo but tracks ChatGPT only" survives being lifted out of the page, out of the section, out of the site.
4. The brand mention comes last and is a factual statement, not a feeling. "Which we think is fairer" is unquotable; "$39/mo base, all four engines on every plan" is quotable.
5. It is two paragraphs, each one idea, each under 90 words. Whichever way a chunker slices this, both halves survive.

The counterintuitive part: the "after" version reads as less promotional, and gets recommended more. Models quote facts, not enthusiasm. We see this in scan data constantly — the pages that get cited are the ones that would work as a Wikipedia paragraph.

How should you structure the page as a whole?

Structure the page as a stack of independently useful answers, not as an argument that builds. The page-level shape that works:

Above the first heading, put a 40-to-70-word direct answer to the title's question. No throat-clearing. This block does double duty: it is the chunk most likely to be retrieved for the head query, and it is what you feed into your FAQPage or Article schema. If your title asks a question, the first paragraph answers it, completely, before anything else.

Make every H2 a question or a claim someone would search. "How much does AI visibility tracking cost?" beats "Pricing." "AI Overviews now appear on 48% of queries" beats "Market Context." The heading is part of the chunk's embedding, and a heading that matches the query is free relevance.

Keep sections to 150–300 words. That is roughly one to two chunks. A 900-word section under one H2 will be split into three or four chunks, and the ones that are not adjacent to the heading lose the heading's context.

Use tables for anything comparative. A table row is the most self-contained content shape that exists, because each cell carries its column header with it. Engines quote tables at a rate wildly out of proportion to how much of the web is tables. If you have three or more things being compared on two or more dimensions, that is a table, not prose.

Put dates and numbers in the text, not only in charts. An image of a chart is invisible to a text retriever. Write the number in a sentence.

Answer the adjacent questions on the same page. If someone asks "how do I structure content for AI citation," they will also ask "how long should a paragraph be" and "do headings need to be questions." Answering those on the same page gives you three retrievable chunks instead of one, which is why the FAQ block at the bottom of a post is not decoration.

The chunk-boundary trick

Write each section so it survives being cut anywhere. Concretely: repeat the subject noun at the top of every paragraph, even when it feels redundant to a human reading top to bottom.

A human reading in order does not need "Google AI Overviews" repeated five times. A retriever that grabbed only paragraph four absolutely does. Mild redundancy in the subject line is the cheapest insurance you can buy, and readers barely notice it because they are skimming anyway.

Does this actually change whether you get cited?

It changes whether you can be cited, which is necessary but not sufficient. Structure gets you into the candidate set. Authority decides whether you win it.

A perfectly structured page from a domain nobody links to and nobody discusses on Reddit will lose to a sloppier page from a source the engine already trusts. Structure is the part of the problem fully under your control, so do it first, but do not expect it to substitute for being talked about. That side of the problem is how AI engines choose sources and why competitors show up in ChatGPT.

The timeline in what we see: rewrites tend to show up in ChatGPT and Perplexity within two to six weeks, because both re-fetch pages aggressively and Perplexity in particular is retrieval-heavy. Google AI Overviews take longer, because they run over the regular index. If you rewrote a page and nothing moved in ChatGPT after eight weeks, structure was not your bottleneck.

Measuring that requires asking the engines the same questions repeatedly and watching what they say, which is what Spottlo automates: your 25 buyer questions, four engines, every week, with the passage that got quoted. You can do it by hand in a spreadsheet for one brand. It stops being fun at question fifteen.

What to do next

1. Run the lift-out test on your five highest-intent pages. Take each paragraph in isolation. If it needs the paragraph above it, rewrite it to stand alone. Pronoun-opening paragraphs are the fastest thing to grep for.
2. Rewrite every H2 as a question or a searchable claim, and make the first sentence under it the answer. This alone is usually a day's work and the highest-leverage change on the list.
3. Convert your comparison prose into tables. Anything with three-plus items and two-plus dimensions.
4. Add a 40-to-70-word answer block above the first heading on each page, and reuse that exact text in your Article or FAQPage schema so the markup and the HTML cannot drift. Schema markup for AI search has the JSON-LD.
5. Baseline before you rewrite. Pull a free AI visibility report now, so that in six weeks you can tell whether the rewrite moved mentions or whether you just enjoyed rewriting. Structure work is easy to feel good about and hard to evaluate without a before.