Contents
In brief
A page can rank well in Yandex or Google and still never show up under a Claude or ChatGPT answer. The assistant rewrites the person's question into short subqueries, sends those to its own search provider, and cites what comes back. Claude uses Brave, ChatGPT mostly Bing, and Yandex's Alice uses Yandex Search. Checking "AI visibility" in one familiar report is like looking up a book in the wrong library.
What happened
A Habr write-up traces where the link under an answer comes from. When the product searches at all, it does not copy the top results for the user's original sentence. It writes several short queries, reads pages from someone else's index or its own, and keeps the ones it used. If your site is missing from that provider, there is nothing to cite. Rank on the original question barely matters. Work by Meihui Chen and Nick Koudas on ChatGPT makes the same point: pages its search finds rarely match Google or Bing.
The model does not search every time. In a Profound sample, Claude searched on only 36.6% of prompts and answered the other 63.4% with no search. Until search is on, Brave rank changes nothing.
The provider map, as the author assembled it from docs and measurements:
| Product | Whose results it uses | What backs that |
|---|---|---|
| Claude | Brave | Anthropic trust center (via Peec), subprocessor list, Peec and Profound samples |
| ChatGPT | Bing plus undisclosed sources; a private OpenAI index in HIPAA workspaces | Bing's webmaster blog, TechCrunch, OpenAI docs |
| Alice | Yandex Search | Yandex Webmaster help |
| Perplexity | Its own index and the PerplexityBot crawler | Perplexity's blog |
| GigaChat | Not confirmed | — |
The Claude–Brave numbers do not collapse into one figure, because the samples differ. Peec compared 8,500 URLs from subqueries: 89 of 100 pages Claude cites sat in Brave's top 20. Profound, on another sample, found 79.2% of cited pages in Brave's top 10 and says the overlap does not prove Claude picks sources because of rank. Claude and ChatGPT barely share sources: across 600-plus queries they averaged about 8% of domains in common.
Why it matters
There is no single AI-answer visibility. A site missing from Bing or Brave barely exists for ChatGPT or Claude, no matter how it ranks on Yandex. Alice is the reverse: Yandex's docs say it leans on highly ranked Search results, so Webmaster reports stay the baseline.
Teams usually edit copy against one report while the citation is born in another index, on another phrase. Profound says Claude adds a year to 94% of subqueries and ChatGPT to 17%. Checking the customer's original question in "your" search engine measures the wrong shelf.
In practice
- For ChatGPT, use Bing Webmaster Tools. Average position there is for ordinary Bing search, not a separate chat rank. The AI Performance report shows the queries the model sends to ground an answer and which pages were cited. A beta Topics Portfolio has existed since 13 October 2026; small sites may not see it.
- For Claude, test the subquery, not the original sentence: a short query that includes the year, in Brave. The Brave API returns up to 20 results per call. There is no separate free tier; about a thousand calls a month come from a $5 credit, and a card is required. The index is smaller than Google or Bing, so Peec suggests submitting important URLs to Brave by hand. The API terms forbid storing the result data.
- For Alice, keep the check in Yandex Webmaster.
- Do not copy Peec and Profound percentages onto ChatGPT or Alice. Both samples are about Claude and Brave.
Takeaway
The link under an answer is a page from one provider's results, for a phrase the model wrote itself. Find out whose search that is before you look for your URL.
Comments