What Actually Earns an AI Citation
We compared the pages AI chose to quote against comparable pages it passed over. The pages it ignored were longer, carried more headings and had more markup. The advantage lies somewhere else.
Key Takeaways
- The typical page AI quotes shares only a quarter of the question's words in its headline. One in twelve shares none.
- Its body covers 80% of the question. The answer is inside the page, just not announced in the title.
- Pages quoted first are almost identical to pages quoted twelfth in length, headings and structure.
- In our closest like for like comparison, the pages AI ignored were longer and carried more markup.
- Answering the question is the entry ticket. Fit is the advantage.
Why we looked into this
Two rules dominate almost every conversation about getting quoted by AI. Put the question in your headline. And write the most thorough page on the subject.
Both arrived from traditional search, where they earned their reputation honestly. For twenty years, matching a searcher's phrasing helped an engine understand what a page was about, and length was a reasonable proxy for effort. Teams built entire content operations on those two ideas. Headline templates were standardised around question phrasing. Briefs specified word counts. Budget was approved on the promise that a longer, more complete page would outrank a shorter one.
Both rules are also expensive. The first shapes how every page on a site is titled and structured. The second is why a single article can take six weeks and cost several thousand dollars before anyone checks whether the extra depth changes anything. When a rule that costly moves into a new channel unexamined, it is worth testing rather than assuming.
How we tested it
We started with every question and every page AI quoted to answer it, giving 39,284 pairings of a question with a page an engine chose as a source.
For each pairing we measured two things. How much of the question's wording appeared in the page's headline, which is what the first rule tells you to optimise. And how much appeared anywhere in the page's content, which tells you whether the page genuinely addresses the question regardless of what it calls itself.
Then we needed the harder half: pages AI did not quote. A dataset of winners tells you what winners look like, never what separates them. Every page in the main set had already won.
The comparison came from a second round of collection. We had visited the home page of every website appearing anywhere in the study. Some were quoted. A further 2,703 never were, despite sitting on websites AI clearly trusted enough to quote other pages from. That gives two groups of the same kind of page, on websites of comparable standing, where the meaningful difference is that AI chose one set and passed over the other.
Read: box spans the 25th to 75th percentile, the vertical line is the median, whiskers reach the 10th and 90th.
Finding one: AI is not matching your headline
The typical page AI quotes shares just a quarter of the question's words in its title. Only 19% share half or more. And 8% share none of the question's wording at all, yet were still chosen as the source.
Look inside the page and the picture reverses. The body of a quoted page covers 80% of the question's wording. More than one in five quoted pages barely gestures at the question in the headline while answering it thoroughly in the text.
This is a meaningful shift in how the machine reads a page. A traditional search engine leaned on the title because it was a compact, reliable signal of subject matter. An AI engine reads the whole document and decides whether the answer is in there. The headline has stopped being the summary the machine trusts and become simply the first sentence it encounters.
There is a practical consequence that catches teams out. A page can be quoted for a question its author never had in mind. Roughly one in five pairings work exactly that way: the page covers the ground thoroughly, the engine finds the answer in the middle of it, and the citation goes to a page whose title suggests it is about something adjacent.
Finding two: more is not better
If depth decided the outcome, pages quoted first would be visibly richer than those quoted last. Every engine returns its sources in an order, and that order is the closest thing to a ranking AI gives us.
They are not richer. A page quoted first and a page quoted twelfth are nearly indistinguishable on every measure of substance we captured.
| Position in the answer | Words | Headings | List items |
|---|---|---|---|
| Quoted first to third | 2,659 | 31 | 70 |
| Quoted fourth to tenth | 2,659 | 31 | 72 |
| Quoted eleventh or later | 2,559 | 32 | 74 |
The medians are effectively flat, and where they move they move the wrong way: pages quoted last carry marginally more headings and list items than pages quoted first.
Notice how substantial all of these pages are. The median quoted page runs to roughly 2,600 words with 31 headings. Nothing thin is winning. But once a page is in that territory, another thousand words does not move it up the answer.
Finding three: the pages AI ignored were the deeper ones
This is the result we did not expect, and the one that changes the advice.
Comparing quoted pages against ignored pages of the same type on comparable websites, the ignored pages were longer, carried more headings and list items, and were more likely to include structured markup. On every measure the standard playbook tells you to increase, the pages AI passed over scored higher.
Why this happens
Put the three findings together and a coherent explanation appears.
An AI engine is not choosing the best page on a subject. It is assembling an answer to one specific question, and it needs a passage that resolves that question cleanly. Depth helps up to the point where the answer is present and clear. Past that, more material is more to sort through, and it can work against you: a page covering fifteen subjects holds a weaker claim to any one of them than a page covering three.
That is why the ignored pages were the longer ones. Length and markup are the natural signature of a page trying to serve many purposes at once. The quoted pages were tighter, because they were built to do one job.
It also explains why the headline matters so little. The engine has already read the document. It does not need the title to tell it what is inside.
What this means for your business
Answering the question inside your content is the entry ticket, not the advantage. Every page AI quotes already does it, so doing it competently no longer distinguishes you from anyone else in the running.
That reframes the standard improvement plan. Make it longer, add more sections, add more markup: all investment in the part that no longer separates anyone. Our comparison suggests you can do every one of those things well and still be the page that gets passed over.
What separates a quoted page is fit. The right kind of page for the kind of question being asked, in the format that question calls for, scoped to one topic rather than spread across a category. Three further findings in this study measure each of those directly, and all three point the same way.
For a marketing team, the practical shift is away from asking "is this page thorough enough" and towards "is this the right shape of page for the question a buyer is actually asking". The second is cheaper to answer, faster to act on and considerably more productive. It also changes what you build: fewer sprawling guides, more pages that each resolve a real question completely and stop.
There is a budget consequence worth naming. If depth is table stakes rather than a differentiator, the marginal thousand words on an already substantial page is close to the least valuable thing your content budget can buy. The same money spent producing a page that fits a question you currently have no page for is likely to earn considerably more. Our work with growth teams starts from exactly that gap.
Methodology
89 questions, each agreed with a chief executive or business owner as one their company needs to monitor, put to ChatGPT, Google AI Overview and Perplexity every day through August 2026. That produced 6,601 answers and 61,536 citations across 9,702 unique pages, every one of which we visited and stored exactly as its server delivered it.
Headline and body matching count shared wording between the question and the page, after removing common words. A page answering a question in entirely different language scores lower than it deserves, so these figures are a conservative floor. We chose this over an AI judgment so any reader can reproduce the numbers exactly.
The comparison group is 135 quoted home pages against 2,703 never quoted, drawn from websites AI uses elsewhere. Home pages are occasionally quoted because an engine wants the company rather than the content, which is a separate finding in this study. All brands are anonymised.
Frequently Asked Questions
Does putting the question in my page title help AI quote it?
Not on this evidence. The typical page AI quotes shares only a quarter of the question's words in its headline, and 8% share none at all. AI engines read the whole document rather than relying on the title as a summary.
Do longer pages get quoted more often by AI?
No. Pages quoted first in an answer are almost identical in length to pages quoted eleventh or later, at roughly 2,600 words. In our comparison against pages AI passed over, the ignored pages were actually longer.
How long should a page be to get cited by AI?
Long enough to answer the question completely. The median quoted page runs to about 2,600 words with 31 headings, so nothing thin is winning, but adding length beyond that point did not correspond to better placement in our data.
If depth does not decide it, what does?
Fit. The right kind of page for the kind of question, in the format that question calls for, scoped to a single topic. Page format, question type and topic scope are each measured separately in this study.
How was this measured?
89 questions agreed with chief executives and business owners were put to ChatGPT, Google AI Overview and Perplexity daily through August 2026, producing 6,601 answers and 61,536 citations across 9,702 pages, each of which was visited and stored.
Are You the Right Shape of Page?
We map the questions your buyers actually ask, then show you which answers you appear in, which your competitors own, and what shape of page is winning them.