Overwhelmingly from the top of the page. In MaxAEO's 2026 controlled study, an identical claim placed in a page's intro survived into AI answers about 72% of the time, versus about 24% for the same claim mid-page — and on pages over 4,500 words, mid-page claims survived just 9% of the time while top-of-page claims held near 68%. This reference lays out the position-and-length data and why the fix is leading each section, not writing shorter, for extractability.
Last reviewed: 29 August 2026. All figures below come from MaxAEO's "Page Length and AI Citations" study (2026): three topics in five depth variants (fifteen guides of 2,900-5,200 words), each probed with about 80 query phrasings across six AI surfaces (ChatGPT, Perplexity, Google AI Overviews, Gemini, Copilot), about 1,200 answer observations over four weeks in June-July 2026. A single distinctive claim was rotated through five page positions using a Latin-square design; "survival" meant the claim appeared verbatim or paraphrased in the answer. It is a first-party, controlled, directional study — cite "MaxAEO, 2026" with the date, and note engines shift weekly and the topics lean B2B.
Where on a page are AI engines most likely to quote from?
The start, by a wide margin, with a small lift at the end. Here is how the same claim fared at five depths:
| Position in page | Claim survived in answer | Cited with a link |
|---|---|---|
| Intro (top ~5%) | ~72% | ~61% |
| ~25% depth | ~51% | ~43% |
| ~50% depth (middle) | ~24% | ~18% |
| ~75% depth | ~38% | ~29% |
| Conclusion | ~46% | ~37% |
Source: MaxAEO, 2026. The shape is a U that heavily favours the opening: an intro claim survived roughly twice as often as a mid-page one. This matches the sibling finding that 64% of quoted passages are the first or second sentence of their section — engines reach for what sits early, both in the page and in each section.
Do long pages get quoted less by AI?
Not at the top — only in the middle. Length barely touched top-of-page claims but steadily crushed mid-page ones:
| Page length | Middle-claim survival | Top-claim survival |
|---|---|---|
| 1,000-1,800 words | ~41% | ~78% |
| 1,800-3,000 words | ~27% | ~73% |
| 3,000-4,500 words | ~16% | ~70% |
| 4,500+ words | ~9% | ~68% |
Source: MaxAEO, 2026. Read the two columns together: a top claim held near 68-78% no matter how long the page ran, while a middle claim fell from about 41% to about 9% as the page grew. The lesson is not "write less" — it is that depth on the page is where citations go to die, so a long page needs every section to earn its own place at the top.
Does this contradict "coverage, not length"?
No — it sharpens it. The finding that content length is not the lever, coverage is says a comprehensive page wins by answering more of the sub-questions a query fans out into. This study adds the mechanism: a long page only captures those branches if each branch's answer is front-loaded in its own section. Because engines retrieve at the passage level, every section is a separate candidate, so "put it at the top" applies to each chunk, not just the document. That is exactly the answer-first, one-question-per-section structure the extractable-chunks how-to prescribes — the position data is why it works.
On the page as a whole, the middle is where citations go to die. The fix is not a shorter page — it is a page where every section leads with its answer, so each chunk is "at the top" of itself.
What should you do about page position?
Structure long pages so no fact depends on where it sits in the scroll.
- Lead the page with your single most important claim. The intro is where survival peaks (~72%); do not build up to the point.
- Front-load every section, not just the intro. Each section is its own retrieval candidate, so put its answer in the first sentence and let the detail follow.
- Split any section where the key fact has drifted past the halfway mark. If it reads like two questions, it is two chunks — give each its own answer-first opening.
- Keep the page as long as coverage requires, no longer. Length is safe when structure is right; it is unstructured depth, not word count, that loses citations.
- Measure per engine after restructuring. Track your citation coverage engine by engine, since none reports which passage it pulled.
The honest read for late 2026: a fact's odds of being quoted depend as much on where you put it as on what it says — and on a long page, only the front of each section is reliably in play. Watching whether a restructure actually moved your citations, engine by engine, is exactly what Buffy Intel measures. Questions: [email protected].