**TL;DR** — Across 24 client sites through May 2026 we audited a formatting choice writers make on almost every page without thinking about it as a citation lever: the wording of the heading directly above an answer — whether the H2 or H3 restates the user's query as a natural-language question ("How long does an AI Overview keep citing a page?") or compresses it into a keyword label ("AI Overview citation duration") — and whether that choice changes how often the AI Overview lifts the passage underneath it. Across 7,140 cited-passage events we joined each cited passage to the heading immediately above it, classifying the heading by form and by how well it matched the query, and recording whether the passage actually answered the question the heading posed. The headline is that heading-question match is a real, cheap and badly underused citation lever, but only when three things line up. Passages sitting under a heading phrased as the exact question being asked were cited 2.4× more often than otherwise-comparable passages under a keyword-label heading. The strongest predictor was question-form match — a heading that reads as the natural-language question the user typed, not a noun phrase stuffed with the keyword, carried most of the gain. The second was answer adjacency — the heading helped only when the passage directly beneath it answered the question in its first sentence; a question heading followed by three sentences of preamble before the answer lost almost all the benefit. The third, and the warning, was heading-answer agreement — a question heading whose passage did not actually answer it (a teaser heading, a heading promising more than the text delivered) was cited no more often than a keyword heading and, on 7% of audited pages, was cited less. One change — rewriting the heading above each key answer as the literal question and moving the answer into the first sentence under it — lifted cited-passage rate by 31% on the affected sites over a 30-day follow-up.
Why we ran this audit
Heading wording is one of the oldest arguments in on-page SEO, and for a decade the keyword-first school won it: put the target keyword in the H2, front-load it, keep it short. That advice was tuned for a ranking algorithm matching strings, and it produced a generation of pages whose headings read like index entries — "AI Overview citation duration," "page freshness signal," "schema markup extraction" — noun phrases optimised for a matcher, not questions a person would ask. The AI Overview is a different reader. It is answering a question, often the literal question a user typed, and it has to find the passage on your page that answers that question. We had a strong suspicion the composer was using headings as a map — reading the heading to decide what the passage below it is about — and that a heading phrased as the question would map more cleanly onto the query than a keyword label. But suspicion is not measurement, and the keyword-heading habit is deeply ingrained, so we needed to know whether rephrasing headings as questions actually moves citation or just feels more natural.
The second motivation was a failure mode we kept walking into. Several clients had pages that clearly contained the answer to a query — the right fact, well written, in a self-contained paragraph — but the paragraph sat under a heading that described the section's topic rather than posing its question, and the passage was not being cited while a thinner competitor page with a question-shaped heading took the chip. The obvious hypothesis was that the competitor's heading was telling the composer "the answer to this exact question is right here" in a way our topic-label heading was not. We needed to know whether the heading was doing that work, because if it is, rewriting a heading is the single cheapest citation intervention available — it changes no facts, adds no content, and touches one line of text.
How we ran the measurement
24 client sites — 9 SaaS, 6 publisher, 5 B2B services, 4 DTC — each with a fixed 200-query basket of its real in-market queries, weighted toward the question-shaped queries ("how", "what", "why", "when", "does") where a heading match could plausibly matter. Twice daily through May 2026 we captured every AI Overview card, and for cards citing a client page we identified the specific lifted passage by matching rendered text to on-page text with paraphrase normalisation, then located the heading immediately above that passage in the document structure. We classified each heading on two axes: form (a natural-language question, a keyword-label noun phrase, or a hybrid) and query match (how closely the heading's wording matched the user's query, scored independently of form so that a keyword heading containing the query terms was not penalised for missing words it actually had). We also recorded answer adjacency — how many sentences sat between the heading and the sentence that actually answered it — and heading-answer agreement — whether the passage delivered the answer the heading promised. We built a matched control for every cited passage: a comparable passage on a similar query with a different heading form, so the comparison was question-heading-vs-keyword-heading rather than good-page-vs-bad-page. The cited-passage cohort was 7,140 events.
Two normalisation moves matter. We scored query match separately from heading form, because the two are easy to confuse — a keyword-label heading can contain every query term and still not read as the question, and we wanted to know whether the question form itself carried signal beyond mere term overlap, which it did. And we matched on passage quality before comparing headings — we paired each cited passage with a control passage our existing cited-paragraph rubric scored as equally citable (self-contained, concrete, right length), so the effect we attribute to the heading is not just the question-headed pages being better-written overall. The 2.4× and 3.0× figures are from those matched comparisons, not raw averages across unmatched pages.
The shape of the heading-match pattern
The flat headline first. Question-headed passages are cited more. A passage sitting under a heading phrased as the natural-language question being asked was cited 2.4× more often than a matched passage under a keyword-label heading making the same point. The effect survived the quality match and survived the term-overlap control: among passages our rubric scored as equally citable, and among headings that contained the same query terms, the ones whose heading actually read as the question were lifted far more than the ones whose heading read as a topic label. The composer behaves as though it uses the heading to decide what question the passage below it answers, and a heading already shaped as the question maps onto the query with no inference required.
The most decision-relevant cut was that the gain is about the heading-passage pair, not the page. We had half-expected a page-level effect — "pages with question headings are well-structured FAQ-style pages and those get cited." That existed but it was weak. The strong effect was local: on the same page, a passage under a question heading was lifted while an equally good passage three sections down under a keyword heading was skipped, and rewriting only the second heading often brought the second passage into citation without touching the first. Heading-question match is not a page-wide structural credential; it is a property of the individual heading-and-answer pair, which is good news because it means the fix is targeted — rewrite the headings above the answers you want cited, not every heading on the site.
Driver one: question form beats keyword overlap
The single strongest predictor was whether the heading was actually phrased as a question rather than merely containing the query's keywords. Holding term overlap constant — comparing a keyword-label heading that contained all the query terms against a question heading containing the same terms — the question form was lifted at 2.4× the rate of the label. The composer appears to read a natural-language question heading as a direct claim about what the passage answers, while a noun-phrase label that happens to contain the words is read as a weaker topical hint. The words being present is not enough; the heading reading as the question is what carries the signal. A heading like "How long does an AI Overview keep citing a page?" outperformed "AI Overview citation duration" even though the second contains the core terms, because the first announces an answer to a question and the second announces a topic.
We ran a structural test on 19 headings across 10 clients, each sitting above a strong, self-contained answer that was going uncited. We rewrote each heading from a keyword label into the literal natural-language question the target query posed, changing nothing about the passage beneath it. Over the 45 days that followed, 13 of the 19 passages began being lifted on at least one target query where they had previously been skipped, and on several the newly-headed passage became the rendered answer. The lever was not new content or a better paragraph; it was relabelling the existing answer as the answer to a question, in the words a person would actually ask.
Driver two: the answer has to sit right under the heading
Holding question form constant, the second driver was answer adjacency — how quickly the passage under the heading actually answered the question the heading posed. A question heading followed immediately by a sentence answering it carried the full effect; a question heading followed by two or three sentences of preamble, context-setting or throat-clearing before the answer lost most of the benefit, citing at little better than a keyword heading. The composer appears to read the heading and then expect the answer to follow directly; when the answer is buried below preamble, the heading's promise and the passage's opening do not line up, and the passage reads as less directly responsive. A question heading is a promise that the answer is right here, and the composer appears to check whether the next sentence keeps it.
We ran a structural test on 16 passages across 8 clients that already had question-shaped headings but answered the question only after a paragraph of setup. We moved the answer into the first sentence under the heading — stating the answer plainly, then giving the context that had previously preceded it — and changed neither the heading nor the facts. Over the 60 days after the change, 11 of the 16 passages improved their cited-passage rate, confirming that the question heading only pays off when the answer is adjacent to it. The pairing matters as a unit: a question heading with the answer two sentences away is a half-built signal, and closing that gap was as productive as adding the question heading in the first place.
Driver three: heading-answer agreement, and the teaser heading that backfires
The third driver was the warning. A question heading whose passage did not actually answer the question it posed — a teaser heading designed to pull a reader down the page, a heading promising more specificity than the text delivered — was cited no more often than a plain keyword heading, and on 7% of audited pages was cited measurably less after a teaser heading was added than before. The reading consistent with the data is that the composer cross-checks the heading's question against the passage's content and discounts, or distrusts, a heading whose answer the body does not deliver; a heading that asks "What is the exact threshold?" over a passage that never states a number is making a promise the content visibly breaks, and the broken promise appears to cost more than an honest topic label would. Heading-answer agreement, not heading form alone, is what the signal rewards.
We confirmed this deliberately on 12 pages across 6 clients where a previous team had added engaging question headings as a copywriting flourish over passages that did not directly answer them. We did not strip the headings; we either rewrote the passage to actually answer the question the heading posed, or where the passage could not answer it we rewrote the heading to match what the passage genuinely covered. Over the following 45 days, the pages we made honest-by-answering recovered citation, and the pages we made honest-by-rephrasing held steady rather than declining further — both better than the teaser state they started in. The actionable rule is blunt: never pose a question in a heading the passage below does not answer in its first sentence, because the composer appears to check.
What changed in our content checklist
Three changes. We rewrote heading guidance from keyword-first to question-first on every page targeting question-shaped queries: the heading above a key answer should read as the literal natural-language question a user would type, not a noun phrase stuffed with the keyword, even though the keyword version still looks tidier in an outline. We added an answer-adjacency check to the same pass: the sentence immediately under a question heading must answer the question, with context and caveats following rather than preceding the answer, so the heading's promise and the passage's opening line up. And we added a heading-answer agreement gate: every question heading must be backed by a passage that genuinely answers it in its first sentence, and a heading that promises more than the text delivers must be either backed with the answer or rephrased down to what the text actually covers — never left as a teaser.
We dropped one habit. For years our house style had favoured short, keyword-dense headings as the safe default, on the theory that a matcher rewards the keyword and a question heading just wastes characters. The audit removes that default for AI citation: a keyword label is legible to a string matcher but illegible as a question, and the composer is reading for the question. So the reflexive keyword-label heading left our playbook for answer sections — we now write the heading above an answer as the question that answer resolves, and keep keyword-label headings only where the section genuinely is a topic rather than an answer.
- 01Phrase the heading as the question. Passages under a natural-language question heading were cited 2.4× more than matched passages under a keyword-label heading making the same point — the question form, not the keywords, carries the signal.
- 02Form beats overlap. Holding query terms constant, a heading that reads as the question outperformed a label containing the same terms; 13 of 19 passages were cited after their heading was rewritten from label to question — same passage, reworded heading.
- 03Put the answer right under the heading. A question heading only paid off when the next sentence answered it; burying the answer below preamble lost most of the gain — moving the answer adjacent lifted 11 of 16 passages.
- 04Do not tease. A question heading the passage did not answer was cited no more than a keyword label and hurt on 7% of pages — back every question heading with an answer in its first sentence, or rephrase the heading down to what the text delivers.
Where this argument breaks
For navigational, brand or transactional queries there is often no question to phrase — a user searching your brand name or a product SKU is not asking a question, and forcing an interrogative heading onto a category or product page is artificial and helps nothing; the question-heading lever is for informational, question-shaped queries, not for every heading on the site. For long-form narrative, opinion or story-driven content, question headings can fragment a piece that is meant to read as continuous prose, and the citation gain on a single buried answer rarely justifies chopping an essay into a Q&A; the lever is for reference and answer content, not for everything. For pages whose structure is already a genuine FAQ the gain is smaller because the headings are mostly questions already — the win there is in answer adjacency and heading-answer agreement, not in the form, which is done. For YMYL topics (health, finance, legal) the heading-answer agreement bar is higher than our averages suggest — the composer appears to penalise a question heading over an incomplete or hedged answer more heavily there, so a question you cannot answer cleanly is better left unasked in a heading than asked and half-answered. The 7% teaser-heading penalty is small and noisy; we are confident teaser headings do not help and mildly confident they hurt, but that is the weakest finding in the audit and we would not restructure a page on it alone. For Chinese-language AI search (文心一言, 元宝, 通义) the heading-question effect was present but weaker in our parallel audit, and the matching was looser — the engines tolerated a topic-label heading more readily and leaned more on the body text to locate the answer, so the heading rewrite bought less there than in Google's AI Overview, though answer adjacency still mattered. Our window was 60 days and the cohort was 24 sites; the multipliers are point estimates that will move by vertical and query type. Outside those carve-outs the lesson holds: in 2026 the AI Overview reads your heading to decide what question the passage answers, the unit is the heading-and-answer pair rather than the page, the question form beats raw keyword overlap, the answer has to sit directly under the heading, and the cheapest citation win on an answer page is to write the heading as the question a person would ask and put the answer in the very next sentence.