| Media kit page | For journalists and PR: the findings as story angles |
|---|---|
| Original blog post | How Claude Turns Brave Search Results Into Citations |
How I tested it
I sent 1,100 prompts to Claude Sonnet 5 with Anthropic’s web search tool enabled. Each asked for the top sites or providers for a buyer scenario, covering 100 industries with 11 prompt variants each. Within 24 hours, I ran all 2,282 unique queries Claude wrote through the Brave Search API. I also ran the same prompts without web search 10,538 times to see what Claude recommended from memory. Finally, I downloaded all 34,491 result URLs and checked every recommended brand against the page text the tool delivered. I collected the data in early July 2026.
What I found
- The results make the connection to Brave clear. Of the 21,203 results Claude’s tool delivered, 95.8% appear in Brave’s top 20 for the same query, URL for URL. Matched results have identical titles 99.3% of the time. Anthropic’s Trust Center also lists Brave Search as a subprocessor for web search.
- Claude writes familiar keyword queries: 2.1 searches per response, averaging 9.2 words each. Of those searches, 51.4% start with "best" and 17.4% with "top". A year appears at the end of 20.9%, and 31.0% end with either a year or a word such as "reviews", "comparison" or "rankings".
- Claude gets about 9 of Brave’s 20 results. Brave’s ranking determines most of that selection, with a preference for newer pages. Among the top 5 results, pages older than 24 months reached Claude 74.7% of the time, compared with 86.5% for pages under 6 months old.
- Brave’s FAQ, discussion and video sections added just 5 exclusive URLs out of 21,203 results.
- Claude reads about 4 KB of text from each page it receives. All 2,290 tool calls were web searches. It never followed a link.
- Claude cites 51.2% of the pages it receives. When a page appears in two or more searches for the same response, the citation rate rises to 80.5%. For pages found in one search, it is 48.9%.
- Third-party lists, reviews and comparisons receive 59.8% of citations. Only 37.4% of the recommended brands had one of their own pages delivered to Claude.
- The delivered pages name or link to 93.0% of recommended brands. In responses where every delivered page downloaded, only 1.5% of recommendations look like they came purely from memory.
Which Brave results reach Claude
| Brave rank | Result delivered to Claude |
|---|---|
| 1 | 89.4% |
| 5 | 82.1% |
| 10 | 46.1% |
| 12 | 26.0% |
| 15 | 12.5% |
| 20 | 4.0% |
Where the brand names come from
For 54.4% of the brands Claude recommends, there is no trace in the Brave data for that response. So I compared the answers with the runs without search. Of the recommendations, 52.9% also appear when Claude answers from memory, and 47.1% do not. Search brings those brands into the answer. It also pushes others out: 52.3% of brands consistently recommended from memory disappear once search results arrive.
The page text helps explain what is happening. Look at brands that appear neither in Brave’s data nor in ten runs without search: 86.4% are named or linked in the text Claude received. Of those brands, 83.7% are named in the text and only 41.7% are linked. Since Claude never followed links in this test, the mention itself is what reaches the model.
On a list page, being in the first 10 matters more than being first. Items in positions 1 to 5 were recommended 22% to 25% of the time. Positions 6 to 10 were recommended 17.5% of the time, and positions beyond 10 only 7.6%.
What this means for your brand
- Work toward Brave’s web top 10 for "best" and "top" queries in your category. A result at rank 10 reaches Claude 46.1% of the time. At rank 20, that drops to 4.0%.
- Keep your pages dated and current. At the same rank, pages older than 24 months had a delivery rate 10 to 12 points lower.
- Cover related queries. Pages found in several related searches had an 80.5% citation rate, compared with 48.9% for pages found in one.
- Build relationships with the third-party pages that already rank, and get your brand mentioned there. An unlinked mention counts as much as a linked one in this analysis.
- Check what Claude recommends with web search off. If your brand appears, it is in the model’s memory, but search can displace it. If it does not appear, it needs to reach Claude through the pages fetched during search.
What the study does not show
- I tested one model, Claude Sonnet 5, and one family of prompts. These citation rates describe this dataset; they do not apply to every use of Claude.
- The Brave and Claude searches were 24 hours apart. The indexes changed a little in between.
- I could not download 13.4% of result URLs. A small set of sites blocks crawlers outright.