Claude's search engine runs on Brave: what it changes for getting cited

Claude's web search does not go through Google. Anthropic has listed Brave Search on its public list of subprocessors, with the label "Web Search", since March 2025. The entry is still there. Brave runs its own index, built by its own crawler, with no call to Bing since April 2023. Ranking first on Google and on Bing therefore does not automatically put you in what Claude sees of your business. Two doors open separately: the Brave index and Anthropic's crawlers.

On this page

All three figures come from Brave's own publications and vary from one post to the next, sometimes a few weeks apart. We give them with their date, and not as independent measurements.

An index almost nobody watches

Brave Software is first of all a browser maker. Its search engine, launched in 2021, followed an unusual path: instead of reselling the results of a giant, the company built its own index and detached itself from Bing in stages. On April 27, 2023, it announced the end of the operation by removing the share of results that still depended on a call to the Bing API, about 7% according to its announcement. Since then, every web result displayed by Brave Search comes from its own index.

Josep M. Pujol, Chief of Search at Brave, summed up the company's position in April 2024: "Brave Search is uniquely positioned in this landscape, with the only independent search index besides Bing and Google." In April 2026, JP Schmetz, head of advertising for Brave Search, used a slightly more cautious wording with MediaPost: "Brave Search is one of the three largest independent indexes in the Western world."

The index fills up in two ways. First through classic crawling, with an in-house crawler. Then through the Web Discovery Project, an opt-in system, off by default, that users of the Brave browser turn on in their settings and that contributes anonymized data to the building of the index.

How we know Claude goes through Brave

The nature of the evidence needs to be stated precisely, because that is where most articles on the subject go off the rails.

Like any company subject to European data protection rules, Anthropic publishes the list of subprocessors that handle data on its behalf. On March 19, 2025, at the very moment Claude received web search, Brave Search appeared on that list. Simon Willison noticed it on March 21 and quoted Anthropic's changelog word for word: "As of March 19, 2025, we have made the following changes: Subprocessors added: Brave Search". TechCrunch published the same day. The entry has not moved since, and we consulted it on September 6, 2026: "Brave Search, Web Search, All products, United States".

A second technical clue was spotted at the time: the presence of a parameter named BraveSearchParams in Anthropic's API. And observers noted that the results served by Claude reproduced those of Brave on the same queries.

On Brave's side, the language is just as cautious. Brian Brown, Chief Business Officer at Brave Software, wrote on July 16, 2025: "The Brave Search API currently supplies most of the top 10 AI LLMs with real-time Web search data, and for some of them, Brave is in fact the only search engine index supporting their AI answers." The claim is strong, and it comes from a named executive. It names no customer, however. When Brave names customers, in its posts and on its product page, they are Snowflake, Cohere, Mistral AI, Kagi or You.com. Anthropic does not appear there.

Brave did publish a post devoted to Claude in June 2026, about an integration of its engine through the MCP protocol in an Amazon Bedrock environment. That is a documented integration you turn on yourself, and not a confirmation of the internal plumbing of Claude's web search. The two things look alike and are not the same.

Why it is probably not Brave alone

Two observations are a reason not to write the equation "Claude equals Brave".

The first comes from the list of subprocessors itself. A second vendor carries exactly the same "Web Search" label: TurboPuffer, a vector database. Its addition was spotted in spring 2026 by the agency Xponent21, without any announcement from Anthropic. What TurboPuffer does exactly in the retrieval path is documented nowhere, and it would be unwise to invent it. Its presence simply indicates that retrieval runs through more than one building block.

The second observation comes from Anthropic's documentation. The help article on web search never names Brave, but it does name Bing: "Claude selects images from web search results, powered by Bing, and displays them alongside its text response." Bing is therefore used for images. That is enough to establish that the pipeline draws on several vendors depending on the type of result.

One last detail for practitioners. Since March 2026, the web_fetch tool in Anthropic's API has accepted a use_cache parameter, whose existence reveals that a page cache is kept on Anthropic's side. Claude therefore does not systematically go and fetch a page live, which has a practical consequence: a corrected page can keep being cited in its previous version for a while.

The robots.txt trap

This is the configuration we run into most often in audits, usually on environments that have just been rebuilt. One technical clarification comes on top, and Brave's page states it bluntly: "robots.txt is not used to prevent a page from being indexed." To remove a page from the Brave index, the noindex tag is what you need. A robots.txt block prevents crawling, not necessarily inclusion in the index, and that confusion produces wrong diagnoses.

Anthropic's three crawlers

Anthropic documents three crawlers, on a page updated April 7, 2026. The page specifies that they honour the standard robots.txt directives and support the non-standard Crawl-delay extension, and it publishes its IP address ranges for verification. The point that matters for a site owner is simple: blocking one crawler does not block the others, and each decision carries a different cost.

The existence of Claude-SearchBot deserves to be underlined. Anthropic documents a crawler whose role is to index, at the very moment it depends on an outside search vendor. The verb "index" on a support page does not prove that a general-purpose web index exists, and we will not claim it does. But it is the most tangible sign of an indexing layer that belongs to Anthropic. The decision to block each of these crawlers or not is covered in full in Business owners: should you let AI crawlers access your content? and the same trap exists at Google, as shown in Blocking Google-Extended does not remove you from Google Search.

What we actually do

What you lose compared with Google

It has to be said plainly, because this is a real constraint and not a detail of method.

With Google, you have a console, coverage reports, a history of positions, documentation that states what counts and, since June 2026, an impressions report for generative AI surfaces. With Microsoft, Bing Webmaster Tools has exposed since February 2026 the grounding queries used to retrieve your content and, since June 2026, your citation share on those queries.

For Claude, nothing comparable exists. No Brave console for site owners, no indexing report, no citation measurement on Anthropic's side. You have a URL submission form and your own eyes. And no, the llms.txt file changes nothing here: what it does and what it does not do is detailed in llms.txt: what the file does not do, and what the work behind it actually does.

The consequence for method is direct: you do not track a rank, you track a citation rate on a sample of questions that stays stable over time, recorded at regular intervals. It is less comfortable, and it is the only honest measurement available. The detail of what each tool sees is in Measuring your AI visibility: what each tool actually sees and the line between the two disciplines is in SEO and GEO: what really differs between the two.

What we do not know

Frequently asked questions about Claude's search engine