Claude finds you through Brave, not Google

Claude's web search is a retrieval step that, by all public evidence, pulls candidate pages from Brave Search's independent index, so the pages Claude cites are largely the pages Brave already ranks, not the pages Google ranks.

Most AI visibility plans quietly assume one index underneath everything. Rank well in Google, get picked up by Bing, and the assistants will follow. For Claude that assumption fails in a measurable way, and the fix involves a search engine most SEO teams have never checked.

#The evidence that Claude searches Brave

Anthropic has not published a statement naming its search provider in the product. The evidence comes from three places.

The first is paperwork. Simon Willison noted that Anthropic's Trust Center added Brave Search as a subprocessor as of March 19, 2025, with the entry linked to web search documentation. In the same post he ran a search for "interesting pelican facts" through Claude and found the ten results matched the same query on Brave exactly, and that the web search tool definition includes a parameter object named BraveSearchParams.

The second is a small early test. Profound compared Claude's citations with Brave's results in March 2025 and found 13 of 15 cited results (86.7%) matched Brave's top non-sponsored results across three queries. Fifteen results is too few to generalize from, so treat it as a first signal.

The third is the large study. MERJ's July 2026 analysis, built on roughly 35,000 Claude citations across 400 queries using Profound data from June 2026, found that 79.2% of the URLs Claude cited sat in Brave's top 10 for the matching query. Only 34% sat in Google's top 10.

That gap is the headline. A page in Google's top 10 is not a strong predictor of a Claude citation. A page in Brave's top 10 is.

#Claude is its own citation market

Josh Blyskal's state of AEO 2026 research adds context from a separate prompt set of more than 600 queries. Claude and ChatGPT shared only 8% of citation domains on average. Claude's cited domains overlapped 64% with Google's top 50, against 37% for ChatGPT.

Those two Google figures look like they contradict MERJ, but they measure different things. Blyskal compares domains against a top 50. MERJ compares exact URLs against a top 10. A large site can rank somewhere in Google's top 50 for a query and still have a different page, or no page at all, in Brave's top 10. Domain-level overlap flatters everyone. URL-level overlap shows which page actually got picked.

Blyskal also flags two limits worth repeating. Claude ran a search for only 36.6% of the test prompts, so most answers were produced without retrieval at all. The authors also state the findings are observational and do not by themselves establish a causal retrieval dependency. The Brave link is well supported, but the exact share will move with the prompt mix.

#Brave discovers pages differently

If Claude inherits Brave's ranking, Brave's discovery rules become your problem. They differ from Google's in ways that catch sites off guard.

There is no submission route. According to MERJ, Brave has no search console, no sitemap submission and no IndexNow support. Its submit URL form only re-fetches a page.

Discovery comes from browser users. Brave's crawler help page says the crawler is partially powered by users who opt in to the Web Discovery Project in Brave browser settings, which is off by default. MERJ describes two channels. A page report needs roughly 20 opted-in users on separate networks to visit a URL before it can be decrypted. A query report fires when a single opted-in user sees your page in Google, Bing, Yahoo or DuckDuckGo results, after which the page is re-fetched anonymously following a delay of 1 to 20 minutes.

In practice, a page that nobody visits and that ranks nowhere else has very few routes into Brave. Ranking in other engines helps, but indirectly: it creates the chance of discovery, not a Brave position.

The crawler inherits your Googlebot rules. Brave's help page states: "If a domain or page is not crawlable by Googlebot, then Brave Search's bot will not crawl it either." It also says the crawler does not advertise a differentiated user agent, so that sites which allow only Google do not shut it out.

Robots.txt does not remove you. The same page says robots.txt is not used to prevent indexing. Delisting requires a noindex directive, and Brave has to re-fetch the page before the change applies.

The index is smaller. Brave's Search API page describes an independent index of over 30 billion pages. It is built with Brave's own ranking models, so it is not a repackaged copy of Google or Bing results, and the ordering differs accordingly.

#Where sites lose Claude without noticing

Three failure patterns follow directly from those rules.

A WAF or bot manager that allowlists named crawlers and challenges everything else can block Brave's fetcher, because there is no Brave name to allow. Robots.txt testers stay green while the edge refuses the request.

A page that is disallowed for Googlebot, perhaps a staging rule that leaked or a faceted path excluded for crawl budget reasons, is also off limits to Brave, and by extension less likely to reach Claude.

A new page with little traffic can sit outside Brave for a long time. Pinging IndexNow reaches Bing and its partners. It does nothing for Brave.

#What to check

Search your priority queries on Brave. Take the commercial and informational prompts you already track for Claude citations, run the underlying queries on search.brave.com, and record whether your target URL appears in the top 10. MERJ's URL-level finding makes this the closest cheap proxy available.

Compare the gap against Google. Pages that rank in Google's top 10 but are missing from Brave's are the ones where Google-based reporting overstates your Claude exposure.

Audit edge rules for unnamed crawlers. Confirm that legitimate fetches without a recognizable bot token are not challenged by default on the paths you want cited.

Check Googlebot access on those paths. Brave inherits it, so any Googlebot disallow is also a Brave disallow.

Use noindex, not robots.txt, to delist. If you want a page out of Brave, and so out of Claude's search results, the directive has to be on the page.

Report Claude separately. With an 8% domain overlap against ChatGPT in Blyskal's data, a blended AI visibility score hides the engine that behaves most differently. Track Claude citations on their own line, and read them next to Brave rankings rather than Google rankings.

One index rarely explains every assistant. For Claude, the index to check first is Brave's.