AI engines cite pages that answer first. Does the answer have to come first?

KEY TAKEAWAYS

1. Cited pages open with a direct answer more often, but the gap shrinks from 20% vs 5% to 23% vs 14% once retrieval is held closer to equal.
2. Within one site, a paragraph that matches the question raises citations on every engine studied. A matching opening raised them significantly only on Gemini.
3. Retrieval explains far more than page format. Pages a search never finds cannot be cited however they open.
4. Test opening rewrites on half of a matched page set, by engine, before rolling them out.

On 28 October 2026, Search Engine Journal reported a study of 193,507 answers from ChatGPT, Google AI Mode and AI Overviews. Pages the engines cited opened with a direct answer four times as often as pages they passed over. The obvious reading is to put a short answer at the top of every page.

That reading merges two claims. One is that a page should contain a clear answer to the question. The other is that the answer should sit at the top. The evidence supports the first on every engine. It supports the second mainly on Gemini.

What did the October study actually compare?

TJ Robertson, whose agency sells GEO services, collected the answers weekly from 27 August to 21 October 2026 using Peec AI. The questions were 4,167 prompts that did not name a company, drawn from 78 client businesses.

In the first round, AI raters blind-scored 195 pairs of pages. 20% of cited pages opened with a direct answer, against 5% of passed-over pages. But the passed-over pages were defined as pages in Google's top 10 that never came up in the AI's own searches. The engines never read them. So that round mixes two things: whether a page was found and whether it opened well.

The second round fixed this. It compared cited pages with pages the same answer retrieved but did not cite. The gap fell to 23% against 14%. By our arithmetic, the ratio drops from four times to about 1.6 times. Search Engine Journal noted that the size of this round and the rater model were not stated.

Two other results deserve weight. Half of the passed-over pages had no answer to the question in the part raters read, against 23% of cited pages. And pages with an author byline had 28% lower odds of citation, which Robertson could not explain. These are associations from one agency's clients. Search Engine Journal itself said a test against a control group is needed.

Does the answer's position change citations?

The best evidence on position comes from Mahe Chen and Nick Koudas at the University of Toronto, posted on 10 October 2026. They tracked ChatGPT, Gemini (through Google's API with Search grounding) and Perplexity daily from 24 June to 21 September 2026, collecting 158,099 answers. They compared each page only with other pages on the same website, which removes most of the brand and authority differences between sites. Our FAQ schema analysis explains why that design matters.

They scored how closely three parts of each page matched the questions: the title, the first paragraph, and the best-matching paragraph anywhere on the page.

Split by engine, the best-paragraph effect was significant on all three: ChatGPT +9.5%, Gemini +7.8%, Perplexity +4.8%. The first-paragraph effect was significant only on Gemini, at +8.6%. ChatGPT showed +0.7% and Perplexity +0.2%.

Where did the best paragraph sit? At a median of 45% of the way down the page. It was in the top quarter for 32% of cited pages and in the bottom quarter for 28%. The authors put it plainly: the paragraph "need not be at the top: the engines find it wherever it is."

Why Gemini differs is not settled. The paper notes that Gemini's searches stay closer to the user's wording than ChatGPT's, which rewrites questions. Our reading, not the paper's, is that an opening which restates the question matches those searches better.

Why do other studies say the top third wins?

Because they measure something else. Kevin Indig's February 2026 analysis of 18,012 ChatGPT citations found 44.2% came from the first 30% of a page's text, 31.1% from the middle 40% and 24.7% from the last 30%. That is widely quoted as proof of answer-first.

Two limits apply. It looked only at pages already cited, so it cannot say whether a better opening gets a page cited. And it did not adjust for section length. By our arithmetic, per unit of text, the first 30% drew 1.47 times its share, the middle 0.78 and the end 0.82. Engines quote openings somewhat more. They do not ignore the rest.

Discovered Labs found the most-cited paragraph sat at a median depth of 0.36, again on cited pages only. Its TL;DR and summary block signal was positive but below its own threshold for a small effect.

A third figure is often misread. AirOps reported that 40.7% of citations fell in the first third. That is the position of the citation within ChatGPT's answer, not within the page. Research on position inside the model, such as the Lost in the Middle paper, moved whole documents within a prompt, not answers within one document.

What decides whether a page is cited at all?

Retrieval. Chen and Koudas found retrieval explained 32% of the within-site variation in ChatGPT's citations, against 2.8% for every page feature they measured. ChatGPT cited 43.8% of the pages it ranked first and 3.5% of those at ranks 11 to 20.

AirOps saw the same order. Pages at search position one were cited 58.4% of the time, against 14.2% at position ten. A weakly matched page at the top position (55.9%) beat a strongly matched page at position six or lower (about 26%). On Google's side, Ahrefs found 76.10% of AI Overview citations came from pages ranking in the top 10.

Format edits look small next to that. In a September 2026 replay test reported by Search Engine Journal, structured rewrites raised the chance a page was cited at all by 4.5 points, with a 95% range from minus 1.4 to plus 10.4. One rerun in seven flipped the cite decision with identical inputs.

Titles are the exception worth acting on. Ahrefs found cited titles were closer in meaning to the prompt than uncited ones, 0.602 against 0.484. Title match passed every robustness test in the Toronto data.

Where did the 40 to 60 word rule come from?

From featured snippets. Portent's 2021 study of 7,854 Google featured snippets found paragraph snippets usually ran 40 to 50 words, two or three sentences, and rarely over 320 characters. The AI-era advice inherited those numbers. A Search Engine Land contributor study defined an answer capsule as 20 to 25 words and found one on 72.4% of cited blog posts across 15 domains, but did not compare uncited posts.

We found no study that measures an ideal answer length for AI engines. This page opens with a 40 to 60 word answer because readers want one, not because an engine has been shown to reward the length.

Is answer-first the same as chunking?

No, and the difference matters. In January 2026, Google's Danny Sullivan said on Search Off the Record that Google does not want content broken into bite-sized pieces for LLMs, while allowing that some sites may see short-lived gains. Mike King of iPullRank replied that passage retrieval makes structure matter.

Both are right about different things. Google's 2026 guide for marketers gives no AI-specific formatting rules. It asks for headings and comprehensive answers. Google has said since passage ranking launched in February 2021 that it can find a relevant passage within a longer page. Answer-first writing states the conclusion, then the detail, on one complete page. Chunking breaks the page apart. The evidence supports the first and gives no reason for the second.

What should you change, by situation?

A worked example

This example is hypothetical. A B2B software firm has 40 guide pages, each tied to three buyer questions. It rewrites the opening paragraph on 20 pages and leaves 20 matched pages alone. It runs each question five times per engine before and after, so 300 runs per group per engine.

On Gemini, the rewritten group rises from 45 to 63 cited runs and the control from 42 to 45. The net gain is 15 runs, or 5 points. On ChatGPT, the rewritten group moves from 51 to 54 and the control from 49 to 53, a net change of minus one. The firm keeps the new openings, credits them to Gemini, and does not expect ChatGPT to move.

How do you test it on your own site?

1. List the questions. Assign each page the buyer questions it should answer.
2. Check retrieval first. Confirm each page is crawlable and ranks for the question. A page that is never found gains nothing from a better opening.
3. Score three parts. Rate how directly the title, the first paragraph and the best paragraph answer each question.
4. Fix missing answers first. Pages with no clear answer anywhere come before pages whose answer sits low.
5. Split into matched groups. Pair pages of the same type and traffic, and rewrite openings in one group only.
6. Sample by engine. Run each question several times on each engine, as set out in how to measure AI visibility.
7. Compare the change, not the level. Subtract the control group's change from the rewritten group's, per engine, and keep only gains larger than run-to-run noise.

What is still unknown?

Robertson's per-engine results and second-round size had not been published when Search Engine Journal reported them. Chen and Koudas used API versions of the engines and English questions about the US market, and could not replay Perplexity. Nobody has published a controlled test of moving an existing answer from the middle of a live page to the top.

In digital authority terms, answer-first writing touches AI visibility and search presence. It does not replace press coverage or a clear brand entity, which decide whether you are found at all. If you want to see which of your pages engines find and skip, join the waitlist.

Questions

Do AI engines only cite the first paragraph of a page?

No. In Chen and Koudas's October 2026 study, the paragraph that best matched the buyer's questions sat at a median of 45% of the way down cited pages. It was in the top quarter of the page for 32% of them and in the bottom quarter for 28%.

Does a TL;DR or summary box get a page cited more?

The evidence is weak. Comparing pages on the same site, a summary box went with 6.9% more citations in one market and 8.0% in another, but passed only two and one of five robustness tests. Discovered Labs found a positive effect below its own threshold for a small one.

Where did the 40 to 60 word answer rule come from?

Mostly from featured snippets. Portent's 2021 study of 7,854 Google featured snippets found paragraph snippets usually ran 40 to 50 words over two or three sentences. No study we found has measured an ideal answer length for AI engines.

Is answer-first writing the same as chunking content for AI?

No. Answer-first means stating the answer before the detail, for readers. Chunking means splitting content into fragments for machines. Google's Danny Sullivan said in January 2026 that Google does not want the second. Nothing in his remarks argues against the first.

Sources

1. Search Engine Journal, Pages AI cited got straight to the answer more often, data shows searchenginejournal.com
2. arXiv, Chen and Koudas, What drives AI citations? Content, retrieval and model knowledge in generative engine optimization arxiv.org
3. Search Engine Journal, Kevin Indig, The science of how AI pays attention searchenginejournal.com
4. AirOps, The fan-out effect. What happens between a query and a citation airops.com
5. Discovered Labs, What actually drives AI citations discoveredlabs.com
6. Search Engine Journal, AI citation test finds source order matters less than it looks searchenginejournal.com
7. Ahrefs, Why ChatGPT cites one page over another. Study of 1.4M prompts ahrefs.com
8. Ahrefs, Study of 1.9M AI Overview citations against organic rankings ahrefs.com
9. Portent, Study. Featured snippet length and display limits portent.com
10. Search Engine Land, How to get cited by ChatGPT. The content traits LLMs quote most searchengineland.com
11. PPC Land, Google warns against breaking content into chunks for AI ppc.land
12. iPullRank, A refutation of misinformation about chunking ipullrank.com
13. Google, Navigating the future of search. 2026 marketer's non-technical guide to SEO in AI search services.google.com
14. 9to5Google, Google Search passage ranking now live 9to5google.com
15. arXiv, Liu et al., Lost in the middle. How language models use long contexts arxiv.org

NEXT STEP

See where your brand stands.

Your Digital Authority Score, who AI recommends for your target questions, and the placements that close each gap. Your first report is free.

Get your free report →