We asked 25 questions a buyer of an AI visibility tool might ask, once each on five assistants, on 2026-10-21, and recorded every citation: 617 in all.
| Assistant | Questions answered (of 25) | Sources cited | Per answered question | Questions with any citation | Own or search-engine documentation | Redirect links |
|---|---|---|---|---|---|---|
| ChatGPT | 25 | 47 | 1.9 | 18 | 43 | 0 |
| Perplexity | 25 | 250 | 10.0 | 25 | 22 | 0 |
| Gemini | 25 | 202 | 8.1 | 25 | 2 | 179 |
| Claude | 25 | 118 | 4.7 | 13 | 2 | 0 |
| Google AI Overviews | 22 | 61 | 2.8 | 22 | 3 | 11 |
Perplexity cited ten sources on every question, 187 different domains over the run, and most of them are small: consultancies, tool vendors, personal blogs. Its most cited domain, spawned.com, is a company's learn hub with pages titled the way people ask.
Claude cited 118 sources from 108 domains, but only on 13 of the 25 questions. On the other 12 it answered from memory and cited nothing. When it does search, it reads the same kind of small page Perplexity does.
ChatGPT cited 47 sources, fewer than two per answer, and 43 of them were OpenAI, Google or Perplexity documentation. A third-party page was cited only where no official document answered the question.
Gemini cited 202 sources, but 179 of them were redirect links that hide the destination, so we can say little about who it favours. Google AI Overviews cited 61 sources over the 22 questions where an overview appeared, and leaned on YouTube, LinkedIn and Reddit in a way no other assistant did.
Agreement was rare. Of 617 citations, 5 went to a source that three or more assistants chose. The typical top source for a question was named by two of five.
What this means for a small site
Your first citation will almost certainly come from Perplexity or Claude, and it will come for a page whose title matches the question. ChatGPT follows later, if at all, and for how-to questions it would rather cite the platform than you. Measure the assistants separately: a brand can be named in every Perplexity answer and no ChatGPT answer, and the fix is different for each.
One run, one day, one topic. The six measured pages on the blog carry the full tables and the limits.
