Does the title of a page matter to be cited by an AI?
Published on
Yes, provided that it contains the words of the question asked. To be cited as a source in an AI's response, this is one of the most consistent elements in our measurements: with ChatGPT, a page whose title contains the question is cited 1.3 to 1.9 times more often depending on the corpus. With Claude, the address of the page matters just as much, 1.2 to 1.7 times more when it contains the question. Ahrefs finds the same connection in ChatGPT. On the other hand, the format of the title, the position of the words, or the terms that the AI adds in its own searches have no stable effect.
For what type of question? We are talking here about being cited as a source in an AI's response, not citation in the sense of copyright. The clarification is not useless: on this exact question, ChatGPT understood 'cited' in the legal sense.
Even ChatGPT misunderstood the question
On September 15, 2026, we asked this exact question to the three AIs. Claude and Gemini searched for studies on titles and citations by AI engines. ChatGPT, however, understood it to mean whether a page title can be cited, i.e., reproduced, in the copyright sense: its searches focused on the protection of titles, and it cited Légifrance, the US Copyright Office, and European law. We asked the question again: same reading.
This is a direct lesson for any title: a word with multiple meanings can send the AI off on a different topic. We also saw this with 'GEO', which all three AIs mistook for the magazine or an organization. A title that resolves the ambiguity, by specifying 'cited as a source' or 'AI engine optimization', protects the page from this misinterpretation.
With ChatGPT: the words of the question and a separator
With ChatGPT, a page whose title contains the terms of the question asked is cited 1.8 and 1.9 times more often on our first corpora, HVAC in French and artificial intelligence in English, 1.3 times across about ten sectors in French and 1.7 times across these same sectors in English. The effect is positive everywhere, stronger or weaker depending on the field.
A separator in the title, such as 'Title | Site' or 'Title: subtitle', is also associated with more citations in ChatGPT: 1.5 to 1.8 times on our industry corpora. This is probably the sign of an editorial page rather than a lever in itself. In Claude, the separator has no effect on these same corpora. The year in the title, sometimes recommended to appear recent, lowered ChatGPT's citations across the ten sectors in French, and had no effect in English.
With Claude: the page address matters as much as the title
Claude reads the pages it retrieves in full. What helps it remember a page is an address that contains the question: 1.5 to 1.7 times more citations on our first corpora, 1.3 times across the ten sectors in French, 1.2 times in English. A readable address, made of words rather than identifiers, is cited 1.3 times more often. A title containing at least two terms of the question is cited 1.1 to 1.2 times more.
The words that Claude itself adds to its searches, a related topic or a type of document, do not help in the title: they lowered citations on our first corpora and had no effect on the subsequent ones. These words belong elsewhere, in the text: that is where the AI looks for them to choose the passage it cites.
What other studies measure
Ahrefs, in its study published on April 15, 2026 on ChatGPT 5.2, measures the proximity of meaning between the title of the pages and the question asked: 0.602 for cited pages versus 0.484 for discarded pages. The proximity is even stronger, 0.656, with sub-questions that ChatGPT asks itself. Pages with addresses written in natural language were cited in 89.78% of cases, compared to 81.11% for others.
Both measurements point in the same direction regarding the question and the address. They diverge on the words of the AI's internal searches: Ahrefs finds them close to the cited titles in ChatGPT, whereas we measure no benefit in placing them in the title, and even a loss in Claude on part of our corpora. The difference in methodology—semantic similarity for Ahrefs, presence of words for us—may explain this gap.
What does not matter in a title
Several formatting tips do not hold up from one corpus to another. The position of the first word of the question in the title, the interrogative form, the difference between the title tag and the visible h1 title, or the exact number of terms repeated have all yielded opposite results depending on the field or the AI.
What carries through are the simple variables: the presence of the words of the question, and with ChatGPT, a separator. Anything that seeks to refine the form beyond that is a matter of tuning that works on one example and fails on the next.
What this changes for a company
A good title for AIs uses the question exactly as your customer poses it, without ambiguity, followed by a separator and the name of the site. The page address carries the same question in readable words. This is the structure of every page in this section.
The rest of the vocabulary, the one with which AIs search, goes into the body of the page, in developed paragraphs. A title stuffed with keywords brings no measured benefit, and with Claude, it can cost citations.
What we do not know
Our reports compare pages that carry or do not carry a characteristic; they describe associations, not the effect of modifying the title on the same page. They focus on the models' API.
Sources
- Ahrefs, Louise Linehan — “Why ChatGPT cites pages”, .
- FaireDuBruit measurements, from September 4 to 16, 2026 — see "How we measured".