Peec AI’s October 28 research examines how external retrieval methods reproduce passages quoted through Claude’s API. Its passage-selection analysis covered 734 pages and 1,568 quotes. A keyword-based method called BM25 selected the quoted passage 41.6% of the time, compared with 21.9% for random selection and 34.1% for taking material from the page’s beginning. These are reported reconstruction results, not disclosed specifications of Claude’s internal system. The authors state that Anthropic has not confirmed their conclusions. Source: https://peec.ai/blog/how-to-optimize-content-for-claude-search
Anthropic’s foundational September 19, 2024 Contextual Retrieval article describes combining contextual information with semantic and keyword retrieval. It does not establish the current architecture of Claude web search. Reference: https://www.anthropic.com/research/contextual-retrieval
AudioSlave analysis: Fetching a page does not establish that its essential information appeared in an answer. Check whether isolated passages identify their subject, date and scope. That is an editorial quality check; the research does not prove that moving text, matching keywords or achieving a similarity score earns citations.