Claude picks its sources from Brave's top 10. Is your gap search, rank or selection?

KEY TAKEAWAYS

1. Claude cites a page only after three gates. It must search, it must retrieve your page, and it must choose it.
2. Retrieval tracks Brave closely (87.8% of retrieved URLs in Brave's top 10), but Claude searched on only 36.6% of one study's prompts.
3. Score each target prompt by trigger rate and Brave position. It shows where the next citation is cheapest.
4. Brave has no submission route. Googlebot access, Google rankings and real visitors are how pages get in.

On 1 September 2026, Complete SEO published the largest public test yet of where Claude's web search gets its sources. Across 1,481 matched queries collected over ten weeks, 87.8% of the URLs Claude retrieved also sat in Brave's top 10 for the same query. Google's organic results, used as the control, held 45.6%. On 1,455 of those queries (98.2%), Claude's source set was closer to Brave's than to Google's.

The common reading is simple: rank in Brave and Claude will cite you. The data supports something narrower. Claude cites a page only after three separate gates, and Brave rank governs one of them. Before spending on Brave visibility, it pays to know which gate each of your target prompts is failing.

How do we know Claude's search runs on Brave?

Anthropic has never said so in its product documentation. The web search tool docs describe access to current web content and stop there. The Claude help page on web search does not mention Brave, though it notes that image results come from Bing.

The evidence is indirect but consistent. TechCrunch reported in March 2025 that Brave Search had been added to Anthropic's subprocessor list. Simon Willison found that Claude's search function schema was titled BraveSearchParams, and one query's ten citations matched Brave's results exactly. The 2026 studies then measured the overlap at scale.

One open question remains. Xponent21 found a second web search subprocessor, turbopuffer, added to the list on 6 May 2026. It sells vector and full-text search, and its role for Anthropic is not public. It may cache pages, rerank them or do nothing that affects citations. Nobody outside Anthropic knows yet.

What are the three gates?

A Claude citation needs all three of these to go your way.

1. Search. Claude has to decide the prompt needs the web. Anthropic's help page says Claude searches when current information would help. Profound found Claude searched on 36.6% of its tested prompts and answered the rest without searching.
2. Retrieval. Your page has to be in the result set Claude gets back. This is where Brave matters. The 87.8% figure measures this gate.
3. Selection. Claude has to choose your page from that set and cite it. Complete SEO's client data shows this is not automatic, even inside the top 10.

Most guides merge the second and third gates. That is why the published figures look like they disagree.

Does Claude just copy Brave's top 10, or does it choose?

Two camps exist. Profound's first test in March 2025 concluded Claude used Brave's top results with little or no reranking. That rested on 13 of 15 results across three queries on one day. Jonathan Clark, relaying a later Profound session in Search Engine Land, said Claude "doesn't re-rank search results." RivalHound and several other guides repeat the 86.7% figure as settled.

The newer data qualifies this. Complete SEO compared client Brave rank with whether Claude cited the client, across 1,408 prompt observations:

If Claude cited the top 10 as delivered, positions 4 to 10 would not trail positions 1 to 3 by 22 points. And Blyskal's 2026 analysis of about 35,000 URLs across 400 queries found 20.8% of Claude's cited URLs sat outside Brave's top 10.

Both camps are measuring something real. The 87.8% counts URLs Claude retrieved. The 79.2% in Blyskal's data counts URLs Claude cited. Those are different units. Citations from outside the top 10 likely come from Claude's other searches in the same answer, since the API lets one request run several. The evidence supports a middle position: Brave largely decides the shortlist, and Claude still picks from it.

Dynamic filtering, which Anthropic shipped in February 2026, adds a reason to expect selection. Claude now writes code to filter results before they reach its context. Anthropic's post on it reports accuracy gains but says nothing about how sources are ranked.

How often does Claude search at all?

This is the gate most Claude guides skip. If Claude answers from memory, your Brave rank is irrelevant to that answer.

Search Engine Land's account of the Profound session gives trigger rates by prompt type. "Best" and recency prompts triggered a search 81% of the time. Ranking prompts triggered 67%, location prompts 55% and comparison prompts 51%. ChatGPT searched on about 90% of prompts. The sample and prompt set were not published, so treat these as directional.

Two practical points follow. First, buyer prompts with "best" or a year in them are where Brave rank does the most work. Second, prompts Claude answers from training depend on what was in its training data. Anthropic says ClaudeBot collects content that could be used for training. Blocking it is a fair choice, but it is a choice about this gate. Our guide to which AI crawlers to allow or block covers the full trade-off.

Blyskal also found Claude put a year into 94% of its search queries, against 17% for ChatGPT. Pages whose titles carry the current year may match Claude's queries more closely, though he reports a similarity gain, not a citation gain.

A worked example: where is the next citation cheapest?

The method combines the two datasets. For each target prompt, multiply the trigger rate for its type by the citation rate for your Brave band. The sum is the expected number of citations per run of the prompt set. The brand and positions below are hypothetical, and pairing two studies this way is our assumption, not a published model.

A B2B software brand tracks 10 prompts. It checks its position in Brave for the query Claude would likely run.

The total is about 3.26 citations per run of 10 prompts. Ignore the search gate and the same positions predict about 5.06. That is 55% higher than the gated figure, which is why reports that skip trigger rates overpromise.

Now compare two fixes. Moving the two pages at positions 15 and 12 into the top 10 adds about 0.43. Moving the three absent prompts into positions 11 to 20 adds about 0.50. The second fix is worth slightly more, and it often costs less, because getting into Brave's index at all is a cheaper job than climbing it.

Run the prompts several times before trusting any single reading. Our note on measuring AI visibility when answers keep changing sets out how many runs a stable number needs.

Which gate is failing, and what fixes it?

Complete SEO found ranking gaps outnumbered selection gaps among its clients, 218 to 56. That is one agency's client base, but it suggests most brands should start with the index and rank rows.

How do pages get into Brave's index?

There is no Brave Search Console and no sitemap upload. Brave's own help page says its Submit URL tool is for asking Brave to re-fetch a page, for example after a noindex change. RivalHound's advice to submit your site through Brave webmaster tools does not match what Brave documents.

Brave's crawler page sets the rules that do exist:

• The crawler uses no distinct user agent. It will not crawl a page that Googlebot cannot crawl, so a robots.txt or firewall rule aimed at Googlebot also shuts out Brave.
• To remove a page from the index, Brave asks for a noindex directive, not a robots.txt block.
• Discovery leans on the Web Discovery Project, an opt-in, desktop-only feature of the Brave browser.

Brave's help centre adds the detail that matters for new pages. A URL must be visited by at least 20 different people before the browser sends it. The project also reports searches users run on other engines and the results they click. A page that already ranks in Google and draws real visitors is far more likely to reach Brave than a fresh page nobody has opened.

Brave says its index holds over 30 billion pages and adds or refreshes over 100 million a day. Those are Brave's own figures. Its API documents a site: operator, marked experimental, which is a quick first check of what Brave holds for your domain.

Do Claude's sources look like ChatGPT's?

No, and that is the reason to measure Claude separately. Blyskal found Claude and ChatGPT shared 8% of cited domains on the same prompts. In his data, listicles made up 36.4% of Claude's citations against 19.7% for ChatGPT, and forums 0.9% against 15.8%.

Yext reports the opposite on user content. Its analysis of 17.2 million citations says Claude cites user-generated sources at 2 to 4 times the rate of other models, and nearly 10 times Gemini's rate in food and beverage. The two studies define the category differently and test different prompts. Yext's sectors are local, where reviews dominate. Blyskal's prompts lean toward products and explainers. Check your own category before copying either pattern.

Why bother at all? Goodie's GA4 panel, mostly B2B brands, found Claude rose from 1.4% of AI referral traffic in mid-2025 to 18.5% in March and April 2026. Yet Claude drew only about 1.3% of visits to the five AI platforms Goodie compared on SimilarWeb. The panel is confidential, but the gap points the same way. For B2B buyers, Claude sends more traffic than its size suggests.

What to take from this

Claude visibility is part of the AI visibility pillar of digital authority, and it is the pillar where engines differ most. Score your prompts by gate. Fix indexing before ranking, and ranking before passage polish. Keep Claude-SearchBot and Claude-User open. When a rival wins a top 10 slot you hold, look at the pages Claude cites and earn a mention in them.

Ethos Engine is building monthly per-engine tracking of this kind into its digital authority reports. Join the waitlist to be told when it opens. Questions go to esh@ethosengine.io.

Questions

Does Claude use Brave Search?

Almost certainly, though Anthropic's product docs do not name a provider. Brave Search was added to Anthropic's subprocessor list in March 2025, and a September 2026 study of 1,481 queries found 87.8% of the URLs Claude retrieved were in Brave's top 10, against 45.6% for Google. A second web search subprocessor, turbopuffer, was listed in May 2026 with no stated role.

If I rank in Brave's top 10, will Claude cite me?

Often, not always. In one agency's client data, pages at Brave positions 1 to 3 were cited 92.0% of the time and positions 4 to 10 were cited 69.8% of the time, when Claude searched. Claude also answers many prompts without searching at all, so a top 10 position only counts on the prompts that trigger a search.

How do I get my site into Brave's index?

Brave has no sitemap submission or webmaster console. Its crawler follows Googlebot's access rules, so a page Googlebot cannot crawl stays out. Brave also learns about pages from opt-in browser users, and a URL must be visited by at least 20 different people before it is sent. Ranking in Google helps because Brave watches which results those users click.

Should I block ClaudeBot or Claude-SearchBot?

They do different jobs. Anthropic says ClaudeBot collects content that could be used for training, while blocking Claude-SearchBot may reduce your visibility and accuracy in Claude's search results. Claude-User fetches pages when a user's question needs them. Block training if you want to, but keep the two search agents open if you want to be cited.

Sources

1. Complete SEO, Nearly nine in ten URLs Claude retrieves are in Brave's top 10 completeseo.com
2. Josh Blyskal, The state of AEO in 2026. Claude is not ChatGPT joshblyskal.com
3. Search Engine Land, Does Claude visibility now hinge on Brave Search rank searchengineland.com
4. Claude Platform Docs, Web search tool platform.claude.com
5. Claude Help Center, Does Anthropic crawl data from the web, and how can site owners block the crawler support.claude.com
6. Claude Help Center, Enabling and using web search support.claude.com
7. Claude, Increase web search accuracy and efficiency with dynamic filtering claude.com
8. Brave Search Help, Brave Search crawler search.brave.com
9. Brave Help Center, What is the Web Discovery Project support.brave.app
10. Brave Search API, Search operators api-dashboard.search.brave.com
11. Brave, How does the Brave Search API compare to other web search API options brave.com
12. TechCrunch, Anthropic appears to be using Brave to power web searches for its Claude chatbot techcrunch.com
13. Simon Willison, Anthropic used Brave simonwillison.net
14. Profound, What is Claude web search, explained tryprofound.com
15. Xponent21, Anthropic lists two web search subprocessors behind Claude xponent21.com
16. Yext, How ChatGPT, Perplexity, Gemini and Claude actually decide what to cite yext.com
17. Goodie, AI search traffic report 2026 higoodie.com
18. RivalHound, Claude runs on Brave Search. Your Bing playbook won't get you cited rivalhound.com

NEXT STEP

See where your brand stands.

Your Digital Authority Score, who AI recommends for your target questions, and the placements that close each gap. Your first report is free.

Get your free report →