If you want to know whether Claude will cite your page, the first thing to check is Brave Search. Peec AI’s study “How to optimize content for Claude search”, by Tomek Rudzki and published on 28 October, found that 89.1% of the pages Claude cites also appear in Brave’s top 20 for the same queries. Russian SEO media picked it up this weekend. A Habr-sourced write-up on 8 November added figures from Profound and told Yandex-focused teams to add Bing and Brave columns to their reports. We think every non-Google SEO should take the same advice.
What Peec measured
Peec used the Claude API, which exposes the fan-out queries Claude runs and the chunks it selects. Across 1,000 prompts, 1,000 Claude queries and 8,500 retrieved URLs, it reran Claude’s fan-outs in Brave and measured how many of the URLs Claude used also appeared there. The figure was 89.1% across the top 20. Peec also says Anthropic’s trust centre names Brave as Claude’s search provider.
The second finding matters more for how you write pages. A typical page in Peec’s sample ran to about 14,600 characters, but Claude passed only about 1,900 of them to the model. That is roughly an eighth of the page. To work out how that eighth is chosen, Peec cut 734 pages into chunks and tested which methods predicted the passage Claude actually quoted, across 1,568 quotes:
- BM25, the 1990s keyword-scoring formula, picked the right chunk 41.6% of the time. Picking two chunks at random scored 21.9%, and always taking the top of the page scored 34.1%.
- Asked which of two chunks Claude would prefer, BM25 was right 74 times out of 100, the same as a neural reranker. Embeddings managed 68.
- Of 12 public embedding models, Voyage and Gemini matched Claude best, but a model Peec trained on Claude’s citations beat both. That suggests Anthropic’s system is custom.
Peec says plainly that none of this is confirmed by Anthropic and that the pipeline could change at any time.
The extra numbers from Profound
The 8 November write-up adds figures it attributes to Profound. On 400 queries, 79.2% of Claude’s sources matched Brave’s top 10. Claude didn’t search at all on 63.4% of prompts. Claude and ChatGPT shared only about 8% of source domains, and Claude added a year to 94% of its sub-queries against 17% for ChatGPT. We haven’t seen Profound’s underlying data, so treat these as reported figures. The overall picture still holds: different AI assistants use different search indexes, and they don’t overlap much.
Why this matters for Brave SEO
This puts a number on something we’ve been saying since Brave started selling search to AI agents. Brave’s index is now a route into AI answers in its own right, and it is much smaller than Google’s or Bing’s. It is also the index the New York Times and Tribune papers are suing over, and the one Brave users can now reshape with Rerank. Practical steps:
- Check whether you are in Brave’s index at all. Search Brave for your key URLs, and use Brave’s URL submission page for important pages that are missing, as Peec suggests.
- Track Brave positions for fan-out-style queries. These are short queries that often include the year, not the user’s full question. Remember that the Search API is a paid product, and its terms restrict storing results.
- Write passages that work on their own. Because only about an eighth of a page reaches the model, the passage that answers the query should be near the top, use specific terms such as names, numbers, years and audiences, and make sense without the rest of the page.
One caveat: 89% coverage shows that Claude retrieves from Brave. It doesn’t prove that ranking higher in Brave makes Claude more likely to cite you. Even so, if your page isn’t in Brave’s results at all, Claude almost certainly won’t cite it.