ChatGPT cites apart: under 4% of sources shared with any other AI

Ask six AI engines the same question on the same day, and compare the sources they cite. In August, two engines shared 6.4% of them on average. Five weeks later, on a corpus of similar size, the average is 7.3%, and one engine stands clearly apart from the other five: ChatGPT shares no more than 3.5% of its sources with any of them.

The overall finding holds: these are six distinct competitions, not a single one displayed on six scoreboards. What moved since August also shows how those competitions draw closer or drift apart, and the way each engine searches is what decides it.

The measurement

Between August 15 and September 21, 2026, our trackers put fixed panels of questions to six engines through their official APIs: 2,847 answers archived with every source they gave, 23,350 citations in all. The six engines read are ChatGPT, Claude, Gemini, Perplexity, Mistral and Grok.

The unit of comparison is still the cell: one question asked in one survey run, at one repetition. Within a cell, the engines were asked the same thing at the same moment. For each cell, we take the set of registrable domains each engine cited, and measure the share of their combined sources that both named. Cells where neither engine cited anything are counted apart and left out of the average.

Finding 1: ChatGPT stands apart, with each of the other five

Each pair carries its own denominator, from 323 to 532 comparable cells. The August column comes from the previous wave, measured with the same instrument:

The five pairs that include ChatGPT fill the last five rows, all between 3.1% and 3.5%. On the same question, ChatGPT and any other engine name roughly one source in thirty in common. At the other end, the closest pair, Claude and Mistral, shares 19.5%: even there, four sources in five belong to one or the other alone.

Finding 2: Claude has moved closer to the others, and it searches more

The four pairs that include Claude but not ChatGPT all rose, by two to seven points. In the same month, Claude's search habit changed more than any other engine's:

Claude searched the web in 81.8% of its answers, against 50.7% in August, and it cites more when it searches: 6.5 sources per sourced answer, against 4.6. An engine answering from memory draws on a training corpus frozen months ago; an engine that searches draws on today's web, which the other engines read too. The more an engine searches, the more likely its set of sources is to cross the others'.

ChatGPT shows that searching is not enough to bring engines together: it searched in 90.9% of its answers, and still sits furthest from all the others. What sets it apart is what it chooses to cite once the search is done.

Finding 3: most of what an engine cites, it cites alone

Of the sources an engine cites in a cell, what share does none of the other five cite there? This measure escapes the mechanical ceiling on sharing, which drops as soon as two engines cite different quantities:

More than seven sources in ten at Perplexity and at ChatGPT are cited by no other engine on the same question. Perplexity gets there through volume, with 16.9 sources per sourced answer; ChatGPT through selection, with 3.9. Mistral has become the most consensual of the six: fewer than three sources in ten are its own, which its closeness to Claude confirms.

Finding 4: concentration stays low, on all six engines

The share of citations taken by an engine's ten most cited domains tells whether its answer plays out on a few sites or on a wide web:

No engine concentrates more than a third of its citations on ten sites, and four of the six stay under 17%. The profiles behind these figures:

• ChatGPT is the most concentrated: Google (6.0% of its citations) and OpenAI (4.2%) lead, followed by marketing software publishers, arXiv, and travel guides such as Lonely Planet and National Geographic.
• Perplexity is the widest: 2,065 domains, with Reddit first at 3.9%.
• Gemini leads with YouTube (2.6%) and Reddit (2.1%). What it reads on YouTube is the transcript.
• Claude spreads its citations, no domain exceeding 1.8%, and reads far more than it shows: 2,834 pages read without being cited, led by Wikipedia with 4.2% of those reads, while Wikipedia does not appear among its thirty most cited domains.
• Grok relies on marketing software publishers, with Semrush and HubSpot first.
• Mistral cites the smallest web of the six, 375 domains, where agencies and specialist sites, many of them French, hold the top of the ranking.

What this changes for a brand

1. Read your visibility engine by engine, and act engine by engine. An average over six engines describes none of them, since their sets of sources barely cross. That is what our method is built on, and why a search ranking no longer predicts a presence in answers. The tools that produce this engine-by-engine reading are in our guide to AI citation analysis.
2. Treat ChatGPT as a project of its own. Its sources overlap with no other engine's beyond 3.5%, and it concentrates its citations the most: there are fewer places, and they are won on its own set. When a competitor holds the answer you wanted there, they are in the sources that engine reads.
3. Follow each engine's search habit. Claude now searches in more than eight answers in ten: a page published today can enter what it cites, where it used to answer from memory one time in two. Measuring which engines cite you, and which read you without citing you, is a job of its own, and Claude can do the reading.

Method

2,847 answers collected between 2026-08-15 and 2026-09-21 by Epovest trackers through the engines' official APIs, on fixed panels worded the way your customers ask. Overlap is measured cell by cell, a cell being one question in one survey run at one repetition, so the paired engines were asked the same thing at the same moment. Sources are counted as sets of registrable domains, citations only: pages read without being cited are left out of the overlap, since only one engine exposes them. Each pair is measured on the cells where both engines answered, from 323 to 532 cells, and cells where neither cited anything are counted apart. Web search is read from the signal each engine returns, answer by answer. The window picks up the day after the August wave, which remains available, so the two waves leave neither gap nor overlap. The figures may be reused with attribution.