If AI assistants drew on the same sources, you would expect them to answer roughly the same. We did not want to assume that but measure it. So we took one Dutch market (video production), twenty questions that real customers ask, and asked each question three times to ChatGPT, Gemini and Claude: 180 measurements in one day.
Outcome one: the assistants fundamentally disagree
The most mentioned brand in the category appeared in 36% of Claude's answers (interval 26 – 47%), in 5% for ChatGPT (2 – 13%) and in 1% for Gemini (0 – 7%). The intervals of Claude and ChatGPT do not even touch: this is not measurement noise, this is a demonstrable difference.
The average of those three (14%) describes no single assistant. Whoever reports their visibility as one number reports a brand that does not exist.
Outcome two: they read different sources
Why do they differ so much? Because they cite from different sources. Across all answers, 398 unique domains were cited. Of those, 316 appeared with exactly one assistant. Only nine domains were cited by all three.
Gemini mainly cites individual company sites, Claude leans on comparison platforms and directories, ChatGPT cites sparingly but variedly. For anyone working on visibility that is worth gold: a mention on one of those nine shared sources counts everywhere, while a mention on an assistant-specific source only works in one place.
Outcome three: half the questions are no-man's-land
On ten of the twenty questions none of the eight tracked brands were mentioned, by any assistant. The assistants did give an answer, but with other names or without names. That is not bad news; it is the cheapest opportunity there is. On uncharted ground the first clear source is immediately the winner.
What this means if you want to work on it
Measure all three. Every assistant is its own field with its own sources. A report on ChatGPT alone is a third of a report.
Steer to the shared sources first. A mention on a domain that all three assistants read costs the same work and pays off three times.
Claim the no-man's-land. Look for the questions where no one is mentioned yet; there the competition is literally zero.
About this measurement
One market (nl-NL), twenty category questions, k=3 per assistant, measured on 29 July 2026 via the official APIs with web search enabled. Sources come from the assistants' own citation metadata, not from the running text. Every statement above carries a 95% confidence interval in the underlying data. This is the same measurement Geony runs weekly for every client.