Six AI engines, one question, and almost no sources in common

Ask six AI assistants the same question, on the same day, and they will each name sources to back their answer. It is tempting to picture them drawing from one pool of authoritative pages, with each engine ranking that pool a little differently. We measured what they actually have in common, question by question: two engines share 6.4% of their sources on average, and no pair of the six reaches 13%.

That number changes what a visibility strategy is. If the engines drew from a shared pool, being cited would be one contest with six scoreboards. They do not, so it is six contests, and a page can win several of them without ever meeting the same competitors twice.

The measurement

Between July 27 and August 14, 2026, our trackers put fixed question panels to six engines through their official APIs, in six languages: 2,883 answers archived with every source they cite, 22,229 citations in total. Mistral and Grok entered the measured catalogue in late July and are read here for the first time alongside ChatGPT, Claude, Gemini and Perplexity.

Comparing engines requires an appariable unit, and a total per engine is not one: two engines citing completely different sites and two engines citing the same sites produce the same totals. So the comparison goes down to the cell, which is one question put in one survey at one repetition. Within a cell, the engines were asked the same thing at the same moment, which is what makes the comparison legitimate. 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 both engines cited nothing are counted separately and left out of the average, since there is nothing there to share.

Finding 1: no pair of engines shares more than 13% of its sources

Each pair carries its own denominator, from 237 to 519 comparable cells:

Eight of the fifteen pairs sit below 5%. The closest two engines in the set, Claude and Mistral, still keep seven sources out of eight to themselves. And the two engines a marketing team is most likely to treat as interchangeable, ChatGPT and Claude, are the furthest apart of all: on the same questions, 2.3% of the sources they name are the same.

Finding 2: the result held a month later, on a corpus three times larger

A single measurement can be an accident of its window. We ran the same instrument on our July measurement, which covered four engines and a third of the answers, and the six pairs measurable in both waves moved very little:

The corpus tripled, two languages became six, and the ceiling stayed where it was. Whatever separates these engines is a property of how each one retrieves and selects, rather than a quirk of one fortnight of questions.

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

Sharing has a mechanical ceiling when two engines cite different amounts: Perplexity names 16.5 sources per sourced answer and Claude 4.6, so even if every source Claude named were also in Perplexity's list, their overlap would read 28%. A second measurement avoids that ceiling entirely by asking a different question: of the sources one engine cites in a cell, what share does no other engine cite there?

Three quarters of what ChatGPT cites is cited by none of the other five on the same question. Even Mistral, the most consensual engine of the six, brings four sources out of ten that nobody else brings. Being read by one engine tells you very little about the other five, and that is measurable on both instruments at once.

Finding 4: they cite differently because they read differently

The profiles behind these gaps are stable, and each one explains a share of the distance:

• ChatGPT leans institutional: Lonely Planet, National Geographic, UNESCO, Tripadvisor and Wikipedia lead its sources, and YouTube is absent from its top 100.
• Perplexity reads platforms first: YouTube and Reddit carry 10.9% of its citations between them, ahead of LinkedIn and Facebook. What it reads on YouTube is the transcript.
• Gemini spreads the widest: 5,124 citations across 2,318 domains, where specialist operators and personal blogs share the top with Reddit.
• Claude works the long tail: its most-cited domain carries 1.1% of its citations, and signed expert pages follow. It also reads far more than it shows, archiving 1,497 pages read without citing them.
• Grok anchors on forums and professional networks: Reddit first, then Semrush, LinkedIn and Malt.
• Mistral cites a small web of niche and review pages: 542 domains, led by specialist sites and Trustpilot.

The retrieval habit sets the rest: Perplexity searched on 100% of its answers, Grok 98.2% and Gemini 95.6%, against ChatGPT 70.9%, Mistral 60.6% and Claude 50.7%. Where an engine does not search, the training corpus answers, and that corpus froze months ago.

What this changes for a brand

1. Read your visibility per engine, and act per engine. An average across six engines describes none of them, since their source sets barely intersect. This is what our method is built on, and why a search ranking no longer predicts a presence in answers.
2. A gap on one engine is a separate job. Winning ChatGPT through institutional records and press does nothing for Grok, where forums and professional networks carry the answer. When a competitor holds the answer you wanted, they are in the sources that engine reads, and the fix is that engine's own source set.
3. Six near-disjoint sets means six openings. A modest page can enter one engine's cited set without competing against everything that already sits in the other five. Measuring which engines cite you, and which read you without citing, is a job of its own, and Claude can do the reading.

Method

2,883 answers collected between 2026-07-27 and 2026-08-14 by Epovest trackers through the engines' official APIs, on fixed panels phrased as customers ask them, in six languages. Overlap is measured cell by cell, a cell being one question in one survey at one repetition, so 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 excluded, since a single engine exposes them. Each pair is measured on the cells where both engines answered, from 237 to 519 cells, so adding an engine to the catalogue leaves the existing pairs unchanged. Cells where both engines cited nothing are counted apart. This wave supersedes the July measurement, which remains available. Figures may be reused with attribution.