Method in one clause: 1,000 prompts compared against Brave results (8,500 URLs), then 734 fetched pages and 1,568 quoted sentences from Claude API responses used to test which passage-picking method best matches Claude's real choices. Figures: Claude's results covered by Brave at "89.1% across the top 20 results"; Claude used about 1,900 characters of a typical 14,600-character page; plain keyword matching (BM25) picked the right passage 41.6% of the time, against 21.9% for random and 34.1% for always taking the top of the page. The authors say the findings are "not officially confirmed by Anthropic", and that a retrieved passage can still go uncited. Peec sells an AI visibility tracker, the prompts are not described by sector, and no small businesses are named.
The Brave link has outside support. Anthropic's subprocessor list added Brave Search on 19 March 2025 (Simon Willison, 21 March 2025, OLDER, matched Claude's citations to Brave's results for one query), and Xponent21 (Isaac Marcuson, 2 July 2026, OLDER, read first-hand) reports a second web search provider, turbopuffer, added 6 May 2026, with its role not public. The subprocessor list itself was not read today. Brave's crawler help page (undated, read first-hand) says that if Googlebot cannot crawl a page, "Brave Search's bot will not crawl it either", and that its crawler has no separate user agent.