Research

How Claude's web search works: what Anthropic documents, and the evidence it runs on Brave

Anthropic is the most guarded of the AI companies about its search stack. It documents Claude's web search tool but never names the search engine underneath it. The evidence nonetheless points hard at Brave: Anthropic's own subprocessor list added Brave one day before launch, a Brave-specific parameter appears in the tool, and independent testers reproduced Brave's results as Claude's citations. Here is what is documented, what is only inferred, and what to do with either.

Published 2026-07-228 sources, first-party plus independent testing9 min read

Michael NefedovFounder

With Perplexity you can read the blueprints, and with ChatGPT you can at least read which vendor it rents. With Claude you get neither. Anthropic tells you the tool exists and what it does for the user, and stops there. So this post is a different kind of piece: it separates what Anthropic has actually put in writing from what has to be inferred, labels each honestly, and then works out what a site owner can do given that the most important fact is one Anthropic will not confirm.

What Anthropic documents#

The documented part is thin and functional. Anthropic's web search tool “gives Claude direct access to real-time web content, allowing it to answer questions with up-to-date information beyond its knowledge cutoff,” and “the response includes citations for sources drawn from search results.”[1]The loop, in Anthropic's words: Claude decides when to search, “the API runs the searches and provides Claude with the results,” and Claude answers with cited sources.[1]

The dates and price are on the record too. Web search launched on the Anthropic API on May 7, 2025, for “Claude 3.7 Sonnet, the upgraded Claude 3.5 Sonnet, and Claude 3.5 Haiku at $10 per 1,000 searches plus standard token costs.”[2] That is the whole of what Anthropic describes about the machine. Notice what is absent.

The one thing Anthropic will not say#

Nowhere in Anthropic's web search documentation or launch announcement is a search provider named.[1] Perplexity publishes its index scale; OpenAI names Bing and Shopify outright. Anthropic names no one. For a feature whose entire job is to fetch and cite the live web, the identity of the search engine underneath is the single most important architectural fact, and it is precisely the fact Anthropic omits. That omission is not an oversight you should paper over. It is the reason the next section is labeled evidence, not documentation.

The evidence it runs on Brave#

What Anthropic will not state, its own disclosures and independent testing strongly imply. Three threads, reported and reproduced by the developer Simon Willison and corroborated in the tech press, point to Brave Search.[4]

First, Anthropic's own Trust Center subprocessor list added an entry for “Brave Search” on March 19, 2025, one day before Claude's in-app web search went live. Second, Claude's function-calling schema exposed a parameter named BraveSearchParams. Third, running the same query against Claude and against Brave returned matching citations.[4]Brave, for its part, sells access to what it calls “a completely independent index of the Web,” an index of “over 30 billion pages.”[5][6]

That is about as strong as circumstantial evidence gets, a legal disclosure plus a leaked parameter name plus reproducible matching results. But the tech press that broke it got no on-record confirmation from Anthropic, and even the investigator who assembled the case hedges with “presumably.”[4]So we report it the honest way: the evidence points to Brave, and Anthropic has never confirmed it. If you need a fact you can stand on without an asterisk, this is not one, and no amount of research fixes that, because the gap is Anthropic's silence, not missing data.

Anthropic's own crawlers#

One area Anthropic does document clearly is its crawlers, and it turns out Anthropic is not a pure renter. Its support page lists three bots with distinct jobs: ClaudeBot, which collects “web content that could potentially contribute to” training; Claude-User, which fetches pages for live user requests; and Claude-SearchBot, which “navigates the web to improve search result quality.”[3]

Claude-SearchBot is the interesting one. It means Anthropic runs some of its own search crawling and indexing, even while the live web search tool appears to lean on an outside provider. The two are documented separately and Anthropic does not explain how they relate, so treat them as two distinct pieces of infrastructure rather than assuming one accounts for the other. Worth knowing alongside this: on Claude.ai, image search is separately and explicitly “powered by Bing.”[7] The stack is more of a patchwork than a single engine.

What to do about it#

Uncertainty about the backend does not leave you without moves. It just means the moves are grounded in what is documented, and honest about what is not.

1. Allow the crawlers that gate live answers

Claude-User retrieves your content for a live query and Claude-SearchBot indexes it for search quality.[3] Blocking either removes you from the paths Claude uses to answer with your page. ClaudeBot is a separate, training-only decision you can make independently.

2. Treat Brave presence as a plausible lever

If the evidence is right and Claude's web search reads Brave's index,[4]then being present and well-ranked in Brave's independent 30-billion-page index[6] is a lever almost no one is pulling. It is a reasonable, low-cost bet precisely because it is uncrowded, offered here as an inference from the evidence, not a guarantee.

3. Verify by measuring, because you cannot verify the backend

Since Anthropic will not confirm what powers the search, the only reliable signal is the outcome: ask Claude the questions your buyers ask, repeatedly, and watch whether it cites you. That measurement is where most tools stop, and where the actual work starts: finding why Claude skips you, shipping the fix, and checking that the citation changed. That is the part Noetio does, and the sourced comparison against the monitoring platforms is on the compare page. For how Claude sits alongside the other engines and their indexes, see the index machines behind ChatGPT, Claude, Perplexity, and Google.

Sources

1. The web search tool gives Claude “direct access to real-time web content”; the API “runs the searches and provides Claude with the results”; no provider is named: Anthropic, “Web search tool” documentation.
2. Web search launched on the Anthropic API on May 7, 2025 at “$10 per 1,000 searches plus standard token costs”: Anthropic, “Web search on the Anthropic API”, May 7, 2025.
3. Three crawlers: ClaudeBot (training), Claude-User (live requests), and Claude-SearchBot, which “navigates the web to improve search result quality”: Anthropic Support, “Does Anthropic crawl data from the web?”.
4. Brave Search added to Anthropic's subprocessor list on March 19, 2025 (a day before launch); a BraveSearchParams field in Claude's tool schema; Claude's citations matching Brave's results; no on-record Anthropic confirmation: Simon Willison, “Anthropic's new subprocessor: Brave Search”, March 21, 2025.
5. Brave Search API is “powered by Brave Search, a completely independent index of the Web”: Brave, “Brave Search API”.
6. Brave's “independent index of over 30 billion pages”: Brave, “What sets the Brave Search API apart”.
7. On Claude.ai, image search is “powered by Bing”: Anthropic Support, “Enabling and using web search”.
8. Corroborating coverage that Anthropic appears to use Brave for Claude's web search, with no on-record confirmation from Anthropic: TechCrunch, March 21, 2025.

Claims about what Anthropic documents quote Anthropic's own docs, blog, and support pages, retrieved and verified on July 22, 2026. The Brave dependency is reported as evidenced, not confirmed: it rests on Anthropic's own subprocessor disclosure plus independent, reproducible testing, and Anthropic has never stated its search provider. Brave's index figures are cited from Brave's own pages.