Claude Pulls Its Links from Brave: 89 out of 100 Citations Sit in the Top 20 of That Search Engine

You spent years building up positions on Yandex and Google, and then a customer asks ChatGPT about your niche and sees your competitors. Sound familiar? The reason is simple: the AI does not open your favorite SEO report. It goes to the search engine that its developer connected to it.

How an AI Finds Links

When an assistant searches the internet, it does not google the user's question word for word. It rewrites it into several short queries, so-called subqueries, sends them to search, reads the pages from the results, and places under the answer the ones that were put to use. That is the citation. If your site is not in the results of the relevant search engine, there is simply nothing to cite.

This leads to an unpleasant conclusion: positions for the original question tell you almost nothing. The AI writes the subqueries itself, anew each time. According to work by Mahi Chen and Nick Kudas, ChatGPT cites pages found by its own search, and they rarely match the results of Google or Bing.

Who Searches Through Whom

Here is the picture by service:

• Claude uses Brave results. Peec checked 8,500 URLs across 1,000 prompts: 89 out of 100 cited pages were found in the Brave top 20. Profound, on 400 queries, got 79.2% matches in the top 10.
• ChatGPT relies on Bing, but not only on it. In HIPAA workspaces, search runs on OpenAI's own index.
• Alisa AI uses Yandex Search and sources that rank highly. For it, familiar positions remain the foundation.
• Perplexity has its own index and the PerplexityBot crawler. For GigaChat, the results provider has not been confirmed.

The source domains of Claude and ChatGPT barely overlap: according to the Profound measurement, they share only 8% on average. A single "visibility in AI" does not exist.

What to Check in Practice

Start not by editing texts, but with positions in the relevant search engines.

For ChatGPT, open Bing Webmaster Tools. You can sign in with a Microsoft or Google account, and sites are imported from Search Console. The AI Performance report shows grounding queries, that is, the queries the AI sent to search, and the pages it cited. Since October 13, 2026, it includes a Topics Portfolio beta, but smaller sites may not have that section.

For Claude, you need the Brave Search API. One request returns up to 20 results, and the region and language are set with the country and search_lang parameters. A script using curl and jq is enough: send a query such as "CRM for an online store 2026", filter the output by your domain, and look at the line number. Empty output means you are not in the top 20. There is no free tier: $5 in credits per month, which is about 1,000 requests, and a bank card is required.

Phrase your test queries the way the AI does. According to Profound, Claude adds the year to 94% of subqueries, ChatGPT to 17%. So for Claude, rewrite the customer's question as a short query with the year. Brave's index is noticeably smaller than those of Google and Bing, so Peec advises submitting important pages there manually.

Where the Numbers Fall Short

Both measurements concern only Claude and Brave. They cannot be transferred to ChatGPT and Alisa AI. Profound itself stipulates that a match with the results does not prove that Claude chooses sources because of Brave positions. The samples and thresholds differ between studies, so the percentages do not add up.

There is another caveat: Claude does not always search. In the Profound measurement, it did without search on 63.4% of prompts, and there Brave positions decide nothing at all. If you store Brave results, note that the API terms prohibit saving the data received, and it is better to clear this with support before cross-checking.

The takeaway for a product manager and founder: add two more columns to your organic report, Bing and Brave. If your site is not there, no content will make ChatGPT or Claude remember you. An open question remains: what share of ChatGPT citations matches Bing results for Russian-language queries. There is no figure comparable to 89% yet.