No AI engine gives more than 27% of its citations to ten sites

Take the ten sites an AI engine cites most often, and count the share of its citations that goes to them. Across the six engines we track, that share runs from 15.5% to 27.1%. None of them builds its answers on a handful of sites: more than seven citations in ten go to domains outside its top ten, which leaves real room on every engine for a source that is not there yet.

What sets the engines apart is the web each one draws on. On the same question, two engines rarely name the same sources, and ChatGPT is still the one that shares the fewest with the other five, as it was in September.

The measurement

Between September 23 and October 19, 2026, our trackers put fixed panels of questions to six engines through their official APIs, and archived every answer with its sources, 3,834 citations in all:

• ChatGPT: 126 answers
• Perplexity: 110 answers
• Gemini: 110 answers
• Mistral: 98 answers
• Claude: 68 answers
• Grok: 44 answers

Concentration is read engine by engine: the share of an engine's citations that goes to its ten most cited registrable domains. Overlap is read pair by pair, on the cell: one question asked in one survey run, at one repetition, where both engines were asked the same thing at the same moment.

Finding 1: concentration stays low, on all six engines

The September column comes from the previous wave, measured with the same instrument:

Over two consecutive waves, no engine gave more than 30.1% of its citations to ten sites, and four out of six stay under 18%. Behind these figures, six profiles:

• Grok cites the smallest web of the six, 137 domains, where software comparison sites and directories (SaaSHub, SoftwareSuggest) sit alongside specialised French sites, with La Fabrique du Net first at 4.6% of its citations.
• ChatGPT draws on 219 domains, led by software publishers: OpenAI (3.5%), HubSpot (3.1%), Salesforce (2.4%), Google (2.1%). On travel, it relies on official sources such as Visit Costa Rica, UNESCO and the US State Department.
• Mistral cites 187 domains, led by French agencies and directories, Digitiz and Sortlist first at 2.4% each.
• Claude spreads its citations over 252 domains, none of them above 2.2%.
• Gemini cites 284 domains, with Reddit (1.8%) and YouTube (1.4%) among its first. What it reads on YouTube is the transcript.
• Perplexity is the broadest: 681 domains and 14.4 sources per answer, led by the press and travel guides, Lonely Planet (2.7%), Le Figaro (1.9%) and The Times (1.4%).

Finding 2: ChatGPT stands apart

For each pair, we measure in every cell the share of their combined sources that both engines named. Each pair carries its own denominator, from 42 to 110 comparable cells:

Four of the five most distant pairs include ChatGPT, and no pair that includes it goes above 5.7%: on the same question, ChatGPT and any other engine name fewer than one source in seventeen in common. At the other end, Claude and Mistral have drawn closer still and share nearly one source in four, which still leaves three quarters of their sources to one or the other alone.

Finding 3: Claude reads far more than it cites

Claude is the one engine of the six that also exposes the pages it read without citing them. For 451 citations, it read 605 pages it did not cite, spread over 290 domains. Wikipedia, Trustpilot and Zapier account for twelve reads each, without appearing among its twenty-five most cited domains. A page can therefore feed a Claude answer without showing in its sources, and the chances of being cited vary from one engine to another down to this gap between reading and citing.

What this changes for a brand

1. Aim for a place on every engine: it exists. Ten sites never take more than 27.1% of an engine's citations in this wave: most citations spread over hundreds of other domains, and a source that is not there yet can get in. The tools that show where you stand are in our guide to AI citation analysis.
2. Read your visibility engine by engine, and act engine by engine. Their source sets barely cross, and each draws on its own web: comparison sites for Grok, software publishers for ChatGPT, the press and guides for Perplexity. This is what our method is built on, and why a search ranking no longer predicts a presence in the answers.
3. Treat ChatGPT as a project of its own. Four of the five most distant pairs include it: when a competitor holds the answer you wanted there, it is in the sources that engine reads.
4. Measure what is read about you, too. A page Claude reads without citing already weighs on its answer. Knowing which engines cite you, and which read you without citing you, is a job in itself, and Claude can do the reading.

Method

556 answers collected between 2026-09-23 and 2026-10-19 by Epovest trackers through the engines' official APIs, on fixed panels worded the way your customers ask, 3,834 citations in all. Sources are counted as registrable domains. The top 10 share is computed on all of an engine's citations, citations only. Overlap is measured cell by cell, on the cells where both engines answered, from 42 to 110 per pair. Pages read without being cited, which Claude exposes, are counted apart and left out of both concentration and overlap. The window resumes the day after the September wave, which remains available, so the two waves leave neither gap nor overlap. The figures may be reused with attribution.