Do AI engines cite the same sources? 79% of sources were cited by only one of six

How we measured it

The questions are the 40 fixed English buyer questions of our AI Source Index, across eight categories from CRMs to running shoes, in four shapes: "best X", "alternatives to X", head-to-head comparisons and "how do I choose". Every week they go to six answer engines: ChatGPT (OpenAI's API with web search), Perplexity, Gemini, Claude, Google AI Overviews and Google AI Mode. We used all ten runs the index has published, 3 September to 2 November 2026: 400 question-weeks.

• Unit: one domain cited for one question in one week. A domain is the lowercase host with a leading www. removed, counted once per answer however many of its pages were cited. Google's own shopping links are dropped, as the public index drops them.
• An engine takes part in a question-week only if its check succeeded and cited at least one domain. 37 checks failed, and 295 succeeded without citing a single domain (124 of them Google AI Mode, 36 of which linked only to Google's own shopping results); those engines are left out of that question-week, not counted as disagreeing.
• "Cited by one engine only" is counted over the 385 question-weeks where at least four of the six engines cited something, so every domain had at least three other engines that could have cited it too.
• Pairwise overlap is the domains both engines cited for a question, out of all the domains either of them cited, summed over every question-week where both cited something.

Four in five cited sources belonged to one engine

The share held week after week: never below 75.2% and never above 82.8% in any of the ten runs. It is not an artefact of engines that sat a question out. Over the 177 question-weeks where all six engines cited something, 4,005 of 5,202 domains, 77.0%, were cited by one engine only, and 10 by all six.

ChatGPT agreed least with everyone

Pair by pair, ChatGPT shared between 3.9% and 5.4% of the domains with each of the other five: of every domain that ChatGPT or Perplexity cited for a question, 5.4% were cited by both (220 of 4,062). The closest pair was Google's own two surfaces, AI Overviews and AI Mode, and even they shared 14.9% (267 of 1,794). Next came Gemini with AI Overviews at 12.4%, which fits Gemini and the Google surfaces drawing on the same search index.

Ahrefs compared 730,000 response pairs page by page; we compared 238 question-weeks domain by domain, which is the looser match and should come out higher. It did, at 14.9%, and the two figures tell the same story from very different samples: even Google's two AI surfaces mostly cite different sources for the same question.

AI Mode's low solo share is partly a matter of size: it cited the fewest domains of the six, and a short list overlaps more easily. Gemini cited the most, and a long list has more room to be alone.

Engines agree with themselves far more than with each other

One engine asked the same question a week later is the fairest yardstick, because a single answer is a sample and engines reshuffle their sources. Measured the same way, an engine's sources for a question overlapped with its own sources the following week by 27.3% (ChatGPT) to 38.7% (Claude). ChatGPT's overlap with any other engine in the same week was 3.9% to 5.4%. Two engines answering the same question on the same day differed far more than one engine differed from itself a week later.

For how long a single engine keeps citing the same source, see how long an AI citation lasts.

The few sources most engines shared

Counting the question-weeks in which at least four of the six engines cited the same domain for the same question, the most frequent were youtube.com (16), vercel.com (10), runrepeat.com (8), nerdwallet.com (7), cnbc.com (7), and healthline.com, ncoa.org, slack.com, garagegymreviews.com and wise.com (6 each). Most are either a specialist review site for one category or the official site of a product named in the question. YouTube is the one general platform on the list.

What could make these numbers wrong

• Domain, not page. Two engines citing different articles on nerdwallet.com count as agreeing here, so every overlap in this guide is an upper bound for overlap at the level of the page.
• ChatGPT changed model inside the window: gpt-4o-mini until 20 October, gpt-5-mini from 21 October. Its solo share was 72.5% and 62.6% in the two runs since, against 75.3% to 82.5% before. Two runs are too few to call that a change.
• From 15 to 28 October our data provider returned Google AI Overviews' and AI Mode's source lists cut to about three domains. Leaving out the three runs it touched (W42 to W44), the one-engine share is 78.4% (5,847 of 7,456) and ChatGPT's solo share 76.7%.
• One sample per engine per question per week, through the engines' APIs and our data provider rather than the consumer apps, in English, for the US. Answers in the apps, other languages and other countries can differ.
• The same domains recur within a question across weeks, so the observations are not independent. That is why we give the spread across the ten weeks rather than a confidence interval on the pooled share.
• 40 questions in eight categories, none about AI visibility tools. Your own questions can behave differently, so measure them rather than borrow these figures.

What to do with it

• Measure each engine separately. A score averaged across engines hides that the engines are, for the most part, reading different sites. Find out where your buyers actually ask, and look at the sources that engine cites.
• Do not assume a citation carries over. Being cited by Perplexity or Google's AI tells you little about ChatGPT on the same question: in this panel they shared about one domain in twenty.
• Start with the sources several engines share. They are few, and for a given question they are usually the category's specialist reviewers and the official sites of the products named. Being on those pages reaches more than one engine at once.
• Check our arithmetic. The rows (one line per question, engine and week, with every cited domain), both scripts and their outputs are public. If you count it differently, we would like to know.

For the sources the engines cite most overall, see what AI engines actually cite, and for Google's AI against Google's own results, Google's AI cites pages Google doesn't rank.

Frequently asked questions