The Claude Retrieval and Citation Study, round one: retrieval and selection in the Claude API's web search tool

Fifty B2B SaaS buyer questions were each put three times through the Claude API's web search tool, 150 runs in total, collected on 7 September 2026 against claude-sonnet-5. Every run logs the domains the search returned separately from the domains the answer cited, so retrieval and selection can be told apart. Main results. Where the query itself named a brand, that brand's own site was among the retrieved domains in 36 of 108 run-brand pairs, 33.3 per cent, and in 3 of 30 on alternatives phrasings. Twelve of the 25 named brands were never retrieved in any run. How-to questions triggered a search in 1 of 24 runs, while recommendation, comparison and alternatives questions searched in every run. A median of 12 domains was retrieved per searching run and 7 cited, a per-run mean citation share of 62.7 per cent. Two runs of the same query overlapped at a mean Jaccard of 0.555 on the cited domain set, 0.435 on comparison queries. Across the study 458 unique domains were retrieved and 333 cited, with 2,138 citations; 107 of the cited domains, 32 per cent, appeared in exactly one run. Matching rule. A brand counts as retrieved when a retrieved host equals one of its known domains or is a subdomain of one. Three narrower rules give 30.6, 27.8 and 25.0 per cent; the finding holds across the whole band. Defects. Two collection defects were found and corrected before publication. Twenty runs whose searches failed silently were re-run, and 42 runs where searches issued exceeded searches billed, meaning partial retrieval, were re-run about twenty hours later in three batches of 14. Every run in this deposit has searches issued equal to searches billed. Because a re-run is a fresh live search, the dataset spans two collection times on the same day. Both defects and their effect on every published figure are recorded in notes/defect-record.md. Total collection cost, read from the Anthropic Console, was 0.92 US dollars.