The title AI assistants cite is the one shaped like the question, 58% against 24%

Someone finishes writing an article and has to give it a title. Two pieces of advice arrive, and they contradict each other. Search optimisation says put a number at the front, "The 10 best", because that is what search engines rewarded for twenty years. AI optimisation says the opposite, ask the question the page answers, because an assistant is looking for an answer and not a list.

We settled the argument on 4,791 answers from six assistants. Neither piece of advice holds.

What decides is not a shape. It is a match.

What you see looking only at the pages that were cited

The reflex is to collect the titles of cited pages and count. Across 15,024 article citations, that count returns this.

The conclusion writes itself: publish numbered lists with superlatives, never publish questions. It is false, and for a reason that has nothing to do with AI. That count has no denominator. If 36% of cited pages carry a number at the front, it may only be that 36% of the pages the engine had in front of it carried one. What is measured is the composition of the web, not a preference.

What is missing is the pages the engine read and did not cite.

The control

One of the six engines returns them. When Claude answers, the record keeps both the pages it cited and the ones it opened without citing. The two groups come from the same answer: same question, same engine, same moment, same language, same search pass. What separates them is the citation itself.

Six hundred and sixty-nine answers carry that control, over 201 distinct questions and six languages. They hold 3,561 cited pages and 4,061 pages read without citation. The title was retrieved for 96.3% of the 3,419 addresses; once anti-bot walls and pages that are not articles are set aside, 5,606 observations remain across 657 answers.

One question is then put to those data: among the pages an engine had in front of it for a given question, are the ones whose title carries a given mark cited more often than the others?

Both recipes fall together

An odds ratio of 1 means the mark changes nothing. The number at the front measures 1.02 and the question mark 1.10, and the interval of each one contains 1. This is not a shortage of data: with 5,606 observations, the interval on the number at the front rules out any effect above 16%.

The question flips the sign

The effects are not absent. They point in opposite directions depending on what the question asks, and they cancel out when you add them up. Here are the observed citation rates, once the answers are split by what was asked.

The same superlative title, of the "Best CRM for small business" kind, is cited 24% of the time when the question asks how to do something, and 57% when it asks which one is best. Same corpus, same engine, same week.

In odds ratios, across the pages seen in answer to a "how" question: a how-to title is worth 2.39, a numbered list 0.28, a superlative 0.15. On a ranking question the superlative crosses back above 1, at 1.32, and the how-to falls to 0.66. The question mark follows the same law: it is worth 2.14 in front of a "how" question and 0.60 in front of a question that is not one.

A title whose shape answers the shape of the question is cited 1.50 times more. And getting the shape wrong costs more than carrying no shape at all: on a "how" question, a superlative title is cited 24% of the time where a title with no particular mark is cited 50% of the time.

What wins everywhere

One measure rises across every family of questions, in the four languages populated enough to be tested, and on both models of the control engine: the share of the question's content words that the title repeats.

The objection is obvious: repeating the words of a question is simply being on topic. The corpus answers it in part. A page's address already says a great deal about its subject, and the alignment of the address alone is worth 1.37. Yet among pages whose address already repeats half the words of the question, those whose title repeats them too are cited 4.36 times more. The title weighs most where the subject is already settled.

By language, the effect runs from 1.38 in English to 2.32 in German, 6.78 in Spanish and 8.66 in French. English is the most worked market, so the one where titles already resemble each other most, and the one where repeating the question pays least.

Two marks of punctuation, and neither comes from folklore

Two traits survive correction for multiple testing. A parenthesis in the title is worth 1.46, and in two cases out of three it holds a year: "(2026)", "(2026 Guide)". It says nothing about the content; it dates the page inside the only fragment of text an engine sees before opening anything. A colon is worth 0.80: the two-part title, "Subject: promise", is cited a fifth less.

The engines do not agree with each other

The paired control exists for one engine only. A second comparison covers all six, and it is weaker: for the same question in the same week, the pages cited by at least one engine are pooled, and we look at which ones each engine kept.

ChatGPT and Perplexity move in opposite directions on the same pages: the numbered list is worth 0.29 on one and 1.86 on the other. The root of the disagreement is in what each one cites, and the site profiles behind the answers already said so: 82% of Perplexity's citations are articles, against 43% on ChatGPT, which cites far more home pages, product pages and documentation.

The sources that matter do not resemble one another either. Forty domains pass twenty article citations in the corpus. The share of their titles carrying a number at the front runs from 0% to 100%, median 39%. Thirty-one of those forty domains never use a question mark, and none passes 18%. Imitating "what the cited sites do" amounts to drawing lots.

What this changes for what you do

Write the title after the question, not before. Start from the wording your market actually types, and repeat its content words in the title: it is the one move that wins everywhere, in the four languages tested as on both models of the control engine.

Then match the shape to the demand. A "how" question calls for a how-to title, a "which is best" question calls for a superlative, and the wrong shape costs more than no shape.

Date the page in the title, in parentheses. The effect is small, but it is real and it does not depend on the question.

And stop believing in shape recipes. The number at the front and the question mark, taken for themselves, return nothing. What makes them work is what they have in front of them.

Method

The corpus covers 4,791 answers from six assistants to 309 distinct questions, from 6 July to 2 September 2026, in six languages and across some twenty sectors. Every answer is archived with its cited sources and its date, giving 46,712 citations to 24,978 addresses across 8,763 domains.

The main analysis rests on the 669 answers where the engine also returns the pages read without citation. The test is a Cochran-Mantel-Haenszel stratified by answer: each answer is its own comparison, which neutralises the question, the language, the sector and the moment. Intervals are Robins-Breslow-Greenland, confirmed by a bootstrap over answers, and Holm correction applies to the family of ten marks tested. Three counter-tests leave the results in place: one page per domain per answer, raw titles with the site banner left in, and the largest panel removed.

The title classification grid was written before any calculation, then audited by hand on seventy titles drawn at random. Eight defects were found, among them anti-bot walls taken for titles and a list missed because the word after the number was two letters long. All eight were corrected and the whole analysis re-run.

Three limits are worth stating. The control exists for one engine only, on two of its models, where the result reproduces; the comparison across the six is descriptive. The rank at which a page was found is not kept in the record, which prevents separating what the search layer ranked from what the model chose: what is measured is the outcome a publisher cares about, being cited rather than merely read. And the title is all we observe of the page, while a title that repeats the question is usually a page that answers it; the comparison with the address limits the objection without removing it.

The composition shares and the comparison across engines rest on titles retrieved for a random draw of 41% of the corpus addresses. The paired corpus is retrieved at 96.3%.

Measured with Epovest.