**TL;DR** — Across 26 client sites in early and mid May 2026 we audited a question that scroll-tracking answers but citation reports never look at: when the AI Overview lifts a passage from your page, where on the page did that passage physically sit — near the top, or buried deep — and does vertical position change your odds of being the lifted passage? Across 8,140 cited-passage events we mapped each lifted passage back to its position in the page's main content, measured as a percentile of content-body depth in words rather than pixels. The cited passage sat in the top quartile of the content body 61% of the time and in the top half 79% of the time; the bottom quartile supplied only 8% of lifted passages despite holding roughly a quarter of every page's words. The strongest predictor of being lifted was section-lead position — a passage that was the first paragraph under its heading was lifted at 3.3× the rate of an otherwise-comparable passage buried mid-section. The second was words-before-answer — pages that placed the answer within the first 150 words of main content were cited 2.7× more often than pages that pushed the same answer below 600 words of narrative intro. The third was self-containment — a deep passage that named its entity and stood alone overcame most of the depth penalty, while a passage opening with an anaphor ("it", "this approach", "as noted above") referring back to earlier text was lifted at 0.4× the rate of a self-contained one at the same depth. One change — promoting the answer to the lead of its section and rewriting it to stand alone — lifted cited-passage rate by 41% on the affected sites over a 30-day follow-up.
Why we ran this audit
Classic on-page work treats the page as a flat bag of passages — get the answer right somewhere on the page and trust the composer to find it. Our cited-paragraph anatomy work told us what a good passage looks like; it never told us where on the page that passage had to be. But we kept seeing pages with a genuinely excellent answer paragraph that were not cited, and on inspection the answer was sitting 800 words down, under three sections of preamble, while a competitor with a weaker answer in its opening line took the chip. We had no audit data on whether vertical position on the page was actually a factor or whether we were pattern-matching on noise, and the difference mattered because one explanation is fixable with a five-minute edit and the other is not fixable at all.
The second motivation was a structural habit in modern content — the long narrative intro. House style across most of our clients, and most of the web, opens an article with scene-setting: why this topic matters, what you will learn, a personal hook. That preamble is good for human dwell time and, we suspected, bad for being the first answerable passage the composer reaches. If position mattered, the long intro was a self-inflicted wound on exactly the pages we most wanted cited, and the fix — promote the answer, demote the throat-clearing — would be among the cheapest and most universal edits we could make. We needed to know whether the position effect was real before we asked dozens of clients to reorder their house style.
How we ran the measurement
26 client sites — 10 SaaS, 7 publisher, 5 DTC, 4 B2B services — each with a fixed 200-query basket of its real in-market queries. Twice daily through early and mid May 2026 we captured every AI Overview card, and for cards citing a client page we identified the specific passage the card had lifted by matching rendered text to on-page text, allowing paraphrase normalisation. For each lifted passage we recorded its position in the page's main content as a percentile of content-body depth — measured in words from the start of the main content region (the article or main body, excluding global nav, header, sidebar and footer), not in pixels, so a page with a heavy hero image was not penalised for chrome the writer could not control. We also logged whether the passage was the first paragraph under its nearest heading (section-lead) or sat below other paragraphs in the same section, how many words of main content preceded it, and whether its opening sentence was self-contained or began with an anaphor pointing back to earlier text. The cited-passage cohort was 8,140 events.
Two normalisation moves matter. We measured depth in content words rather than pixels or scroll percentage because pixel depth is dominated by images, embeds and CSS the writer often does not control, and we wanted a signal an editor could act on directly. And we excluded single-section pages and pages under roughly 300 words of main content, where "position" is meaningless because there is only one place the answer can sit; those were 14% of cited events and cannot inform a position argument. We report the position effect only on multi-section pages long enough for depth to be a real editorial choice rather than an accident of length.
The shape of the position pattern
The flat headline first. Lifted passages are heavily front-loaded. The cited passage sat in the top quartile of the content body 61% of the time and in the top half 79% of the time. The bottom quartile of the page supplied only 8% of lifted passages, even though it holds roughly a quarter of every page's words — so the back of the page is cited at about a third of the rate its word count would predict if position were neutral. The composer is not reading every page to the bottom with equal attention; it disproportionately lifts from the part of the page a hurried human would also read first, and the falloff with depth is steep rather than gentle.
The most decision-relevant cut was that this is not simply "the page's single best passage happens to live near the top." We controlled for passage quality using our existing cited-paragraph rubric — self-contained claim, 40–80 words, a concrete number or date — and the front-loading held within quality bands: among passages the rubric scored as equally citable, the ones near the top were still lifted far more than the ones near the bottom. Position is doing independent work on top of passage quality, not merely correlating with it. A strong answer buried deep is genuinely disadvantaged against the same answer near the top, which means position is a lever, not a symptom.
Driver one: section-lead position decides most lifts
The single strongest predictor was whether the passage was the first paragraph under its heading. A section-lead passage was lifted at 3.3× the rate of an otherwise-comparable passage that sat second, third or later under the same heading, and the effect was sharp rather than gradual — the drop from lead to second paragraph was larger than the drop from second to fifth. The composer appears to treat the heading as a question and the first paragraph beneath it as the candidate answer, weighting later paragraphs in the section as elaboration it is less likely to quote. A great answer written as the third paragraph of a section is, from the composer's point of view, sitting in the elaboration slot, not the answer slot.
We ran a structural test on 19 pages across 9 clients, each with a strong answer that sat mid-section under a relevant heading. We moved the answer to be the first paragraph under its heading, pushing the prior throat-clearing ("Before we get into X, it helps to understand Y…") down or out, and changed nothing about the answer's wording. Over the 45 days that followed, 13 of the 19 pages began being lifted on at least one target query where they had previously been skipped, and on 5 of those the exact promoted paragraph became the rendered passage. The lever was not better writing; it was moving the existing good writing into the slot the composer reads as the answer.
Driver two: words-before-answer is a measurable tax
Holding section-lead status constant, the second driver was how far into the page the answer sat at all — the count of main-content words before the answer-bearing passage. Pages that placed the answer within the first 150 words of main content were cited 2.7× more often than pages that pushed the same answer below 600 words of narrative intro, and the relationship was monotonic across the bands we cut (under 150, 150–400, 400–800, 800+). This is the long-intro tax made measurable: every hundred words of preamble before the first real answer lowered the odds that the answer was the passage lifted, because the composer either found an earlier, weaker passage first or treated the late answer as buried.
We ran a structural test on 16 pages across 8 clients, all carrying the house-style long narrative intro ahead of the answer. We cut or compressed the intro so the answer-bearing passage arrived within the first 120 words, moving the scene-setting below the answer where readers who want it can still reach it. Over the 60 days after the change, 11 of the 16 pages improved their cited-passage rate, and the two clearest gains were on pages where the old intro had contained a hedging sentence the composer had been lifting instead of the real answer — removing the early hedge stopped it from being mistaken for the answer. The fix costs nothing in total word count; it is purely about order.
Driver three: self-containment buys back depth
The third driver was the one piece of good news for pages that genuinely need depth. A deep passage that was self-contained — it named its entity, did not depend on the sentence above it, and could be quoted with no surrounding context — overcame most of the depth penalty: a self-contained passage in the bottom half of the page was lifted at 2.3× the rate of a non-self-contained passage at the same depth. The killer for deep passages was the anaphoric opener — a passage beginning with "it", "this approach", "that method", or "as noted above" referring back to earlier text was lifted at 0.4× the rate of a self-contained one at the same depth, because the composer cannot quote it without dragging in the antecedent it points at. Depth hurts, but un-quotability hurts more, and the two compound: a deep passage that also opens with an anaphor is almost never lifted.
We ran a structural test on 14 pages across 7 clients with strong but anaphor-dependent passages deep in long reference pages we did not want to shorten. We rewrote each deep passage's opening sentence to restate its subject explicitly ("INP" instead of "it", "the deferral technique" instead of "this approach") without moving the passage up the page. Over the following 45 days, 9 of the 14 deep passages began being lifted on target queries despite their depth, confirming that for genuinely long pages the cheaper fix is making deep passages independently quotable rather than restructuring the whole document. Self-containment buys back most of what depth costs.
What changed in our content checklist
Three changes. We added a "promote the answer" pass as the first editorial step on any page targeting an answerable query: the answer-bearing passage must be the first paragraph under its section heading and must arrive within the first ~120 words of main content, with scene-setting moved below it. We added a "self-containment" check to every passage we want cited — it must name its own subject and be quotable with zero preceding context, which we test literally by reading the passage cold and asking whether a stranger would know what it is about. And we changed our page templates so the long narrative intro, where house style demands one, sits after the lead answer rather than before it, preserving the human hook without making it the first thing the composer reaches.
We dropped one habit. Through 2025 we had defended the long narrative intro as good for engagement and assumed the composer, unlike a skimming human, would read the whole page evenly and find the answer wherever it sat. The audit shows the composer front-loads its attention much like a human skimmer does, and the long intro was costing citation on the exact pages it was meant to make compelling. So the intro stopped being the opening of the page and became a section the reader reaches after the answer — the engagement argument survives, the position cost does not.
- 01Put the answer first under its heading. Section-lead passages were lifted at 3.3× mid-section ones; 13 of 19 audited pages were cited after the existing answer moved into the lead slot — same words, better position.
- 02Get the answer in early. Answers within the first 150 words of main content were cited 2.7× more than the same answer below 600 words of intro; cutting the preamble lifted 11 of 16 audited pages.
- 03Make deep passages self-contained. A self-contained deep passage was lifted at 2.3× an anaphor-opened one; rewriting openers (no "it" / "this approach") cited 9 of 14 deep passages without moving them up.
- 04Measure depth in content words, not pixels. 61% of lifted passages sat in the top content-body quartile and only 8% in the bottom — front-loading is real and it is an editor-controllable variable.
Where this argument breaks
For very short pages and single-answer pages there is no position to optimise — the answer is the page, and these levers are overhead. For well-structured reference pages with an anchored table of contents and stable in-page jump links, the composer is better at navigating to the right deep section, so the depth penalty is softer there — a good ToC partly buys back depth, which is the structural version of self-containment. For pillar pages whose value is comprehensiveness, you cannot and should not promote every answer to the top; the right move is one lead answer plus self-contained section leads throughout, not a top-heavy page. The self-containment lever can be over-applied: restating the subject in every sentence reads robotically to humans, so the rule is self-contained passage openers, not self-contained sentences. For Chinese-language AI search (文心一言、元宝、通义) the front-loading was even stronger in our parallel audit — those engines lifted from the top content quartile at roughly 70% — but they were also more sensitive to a question-shaped heading immediately above the answer, so the section-lead lever matters more there and the words-before-answer lever slightly less. Our window was 60 days and the cohort was 26 sites; the per-vertical numbers are point estimates that will differ by template and content type. Outside those carve-outs the lesson holds: in 2026 the AI Overview lifts disproportionately from the front of the page and from the lead of each section, position does independent work on top of passage quality, and the cheapest citation win on a long page is to promote the answer and make it stand alone — not to write a better answer further down where the composer is unlikely to reach it.