Anthropic publishes no documentation on how Claude selects citations, but Claude's web search is Brave-backed, and a small analysis by Profound found 86.7% overlap (13 of 15 results) between Claude's citations and Brave's top organic results. If that pattern holds at scale, then optimising for Claude is mostly optimising for Brave Search: a ranking surface almost no marketing team works on. This post is the thesis, the evidence, its limits, and a test you can run yourself this week.
Key Takeaways
- Anthropic documents three crawlers: ClaudeBot (training), Claude-User (user-initiated), Claude-SearchBot (improving search result quality), and publishes verifiable bot IPs at
claude.com/crawling/bots.json. - Anthropic publishes no documentation on citation selection. Everything below the crawler layer is inference.
- Profound's analysis found 86.7% overlap (13 of 15 results) between Claude citations and Brave's top organic results, with 100% match on two of three test queries.
- That is n=15 across 3 queries. It is suggestive, not proven. Anyone telling you Claude "is" Brave rankings is overclaiming from a very small sample, including anyone quoting this post.
- If the thesis holds, Brave Search SEO becomes a live channel: Brave runs an independent index with its own webmaster surface, and competition there is close to zero.
- The test is cheap. Twenty queries, an afternoon, and you will know more about your own footprint than any vendor study can tell you.
Why This Question Is Worth Asking At All
Every AI visibility conversation I have with a founder eventually reaches the same wall. We can measure ChatGPT roughly. We can measure Copilot properly, because Bing Webmaster Tools gives first-party citation data. We can infer Perplexity from its documented retrieval architecture.
Claude is the black box. Anthropic has published careful documentation on crawler behaviour and nothing at all on how the model decides which of the retrieved pages to cite. For a model with serious adoption among developers, analysts and increasingly ordinary knowledge workers, that is a large blind spot.
So the interesting question is not "how do I rank in Claude." It is "what is Claude actually retrieving from, and can I influence that upstream?"
What Anthropic Does Document
Credit where it is due: Anthropic's crawler documentation is clean, and its published IP list is genuinely useful.
ClaudeBot
The training crawler. Collects publicly available content that may be used in model training. Blocking it is a data-policy decision, not a visibility decision, the same separation OpenAI maintains between GPTBot and OAI-SearchBot.
Claude-User
Fires when a user asks Claude to fetch a specific URL, or when Claude's web-browsing tool retrieves a page during a conversation. This is the closest analogue to a human clicking your link.
Claude-SearchBot
Documented as crawling to improve search result quality. This is the one that plausibly touches the retrieval corpus Claude draws from.
The bots.json file
Anthropic publishes its crawler IP ranges at claude.com/crawling/bots.json. Use it. User-agent strings are trivially spoofed; IP verification is how you find out whether the "ClaudeBot" hits in your logs are actually Anthropic or a scraper wearing its name. In my experience a non-trivial share are the latter.
Anthropic's crawler and support documentation is the authoritative reference here: check support.anthropic.com rather than trusting third-party summaries, including this one.
What Anthropic Does Not Document
How citations are selected. Whether there is a re-ranking layer. How freshness is weighted. Whether domain trust plays a role. What triggers a search at all versus answering from parametric knowledge.
None of it. There is no Claude equivalent of Google's Search Central guidance, no ranking-signals page, no webmaster surface.
That silence is what makes the Brave overlap finding interesting, and what makes it important not to fill the silence with confident guesses.
The Profound Finding, Stated Honestly
Claude's web search is Brave-backed. Profound, a vendor in the AI visibility space, ran a comparison between Claude's cited sources and Brave's top organic results for a set of test queries.
The finding: 86.7% overlap, 13 of 15 results. Two of the three test queries matched at 100%. Profound's conclusion was that Claude appears to surface Brave's top-ranked organic results without heavy re-ranking.
Now the part most write-ups leave out
That is fifteen results across three queries. It is a vendor-run study. Three queries is not a sample from which you can generalise to a model handling millions of daily searches across every conceivable topic, language and query intent.
I want to be precise about what this does and does not support:
- It supports: the hypothesis that Brave's organic ranking is a meaningful input to Claude's citation set, strongly enough to justify spending a day testing it.
- It does not support: the claim that Claude citations are Brave rankings, that re-ranking does not occur, or that the relationship is stable across query types.
Three queries could easily be three queries where Brave's index happened to be unambiguous. Navigational and factual queries would produce high overlap under almost any retrieval architecture. The interesting cases, contested comparisons, commercial-intent queries, freshness-sensitive topics, are exactly the ones a three-query sample is least likely to have covered.
Why the Thesis Is Still Worth Acting On
Here is the asymmetry that makes this worth your time even at low confidence.
If the thesis is right, you have found a ranking surface with an independent index, a real webmaster interface, and effectively no competition, that feeds a major AI assistant. That is an enormous edge.
If the thesis is wrong, you have spent some effort making sure your site is well-indexed in an additional independent search index. The downside is a few hours and no harm done.
Low cost, high variance, positive expected value. That is a bet worth taking even when the evidence is thin, provided you are honest with yourself and your stakeholders that it is a bet.
Brave is genuinely independent
This is the part that surprises most marketers. Brave Search runs its own index, built from its own crawler, not a rebrand of Bing or Google results. It has its own webmaster surface for submitting and checking site coverage. Almost nobody in the SEO world does deliberate work there, because until AI assistants started grounding on it, the traffic did not justify the effort.
The traffic case may have changed. The competitive case certainly has.
How to Test This Yourself, The Actual Protocol
Do not take Profound's word for it. Do not take mine. Twenty queries, one afternoon.
Step one: pick queries that matter to you
Twenty queries from your actual category. Mix the types deliberately:
- Five informational ("how does X work")
- Five commercial-comparison ("best X for Y", "X vs Z")
- Five long-tail specific questions your buyers actually ask
- Five freshness-sensitive ("X changes 2026")
The mix matters. If your overlap is high on informational and collapses on commercial, that is a far more useful finding than a single blended percentage.
Step two: capture Claude's citations
Run each query in Claude with web search enabled. Record every cited URL in order. Use a fresh conversation per query so context does not bleed.
Step three: capture Brave's top organic results
Run the same query in Brave Search. Record the top ten organic URLs in order. Exclude ads, widgets and answer boxes, organic results only.
Step four: compute overlap two ways
Compute set overlap (what share of Claude's citations appear anywhere in Brave's top ten) and rank correlation (do they appear in a similar order). Set overlap tells you whether Brave is the source. Rank correlation tells you whether re-ranking is happening.
That second measure is the one Profound's headline number does not capture, and it is where the real answer lives. High set overlap plus low rank correlation would mean Brave supplies the candidate pool but Claude reorders it, a materially different strategic picture from "Brave's top three win."
Step five: log what you find, including the null result
If your overlap comes back at 40%, that is a finding. Publish it. The space badly needs more people running tests and reporting honest numbers instead of recycling one small vendor study into confident advice.
If the Thesis Holds: What You Would Actually Do
Assume for a moment your test comes back high. What follows?
Get properly indexed in Brave
Brave's webmaster surface lets you check coverage and submit. Start there. An excellent page that Brave has never crawled cannot be cited by anything downstream of Brave.
Treat Brave rankings as a tracked metric
If Brave organic position predicts Claude citation, then Brave position becomes a leading indicator worth putting on a dashboard. Nobody is tracking this. That is the opportunity.
Do not abandon anything else
Brave-focused work does not replace Google, Bing or the general quality work that helps everywhere. It is additive. Most of what makes a page rank in an independent index, clear structure, genuine substance, crawlability, sensible internal linking, is the same work that helps in every other index.
Keep Claude's crawlers unblocked
Specifically Claude-SearchBot and Claude-User. Whatever the retrieval architecture turns out to be, blocking the retrieval crawlers guarantees the answer is no.
What Would Falsify This
Good theses come with disconfirming conditions. Here is what would tell me the Brave thesis is wrong or has stopped being true:
- Overlap testing at n=100+ across mixed query types coming back below roughly 50%.
- Claude consistently citing pages that do not appear anywhere in Brave's top results for the same query.
- Anthropic publishing documentation describing a distinct retrieval or re-ranking layer.
- A visible change in citation behaviour with no corresponding change in Brave's rankings.
Watch for those. If you see them, drop the thesis. Holding a position after the evidence turns is how practitioners lose credibility, and this whole space has too much of that already.
Frequently Asked Questions
Does Anthropic confirm Claude uses Brave Search?
Claude's web search being Brave-backed is established. What Anthropic does not document is how citations are selected from what is retrieved, including whether any re-ranking is applied.
How reliable is the 86.7% overlap figure?
It comes from a vendor-run analysis of 15 results across 3 queries. Treat it as a reason to test, not as a measurement you can plan a budget around.
Should I stop doing Google SEO and switch to Brave?
No. Brave work is additive and cheap, not a replacement. Google still drives the overwhelming majority of search traffic for almost every business.
How do I submit my site to Brave Search?
Brave provides a webmaster surface for coverage checks and submission. Confirm your site is crawled and indexed there first. That is the prerequisite for everything else.
Which Anthropic crawler should I make sure is unblocked?
Claude-SearchBot and Claude-User, at minimum. ClaudeBot is the training crawler and is a separate policy decision.
How do I verify that a ClaudeBot hit in my logs is genuine?
Check the requesting IP against Anthropic's published list at claude.com/crawling/bots.json. User agents are spoofable; IP ranges are not.
Will Claude citations show up in my analytics?
Only partially. Referral data from AI assistants is inconsistent, and answers that do not generate a click leave no trace at all. Assume you are undercounting.
Does this mean Brave SEO is the same as Google SEO?
Substantially overlapping but not identical. Brave has its own crawler, index and ranking system. Core quality fundamentals transfer; specific tactical assumptions from Google may not.
Is anyone actually optimising for Brave right now?
Very few people, which is precisely why it is worth a look. Low competition on a surface that may feed a major assistant is an unusual combination.
What sample size would make you confident in this thesis?
Personally, a hundred or more queries spanning informational, commercial and freshness-sensitive intents, run by someone with no product to sell, with both set overlap and rank correlation reported. Until that exists, this stays a thesis.
If you run this test, I would genuinely like to see your numbers: especially if they contradict the thesis, because that is more useful than another confirmation. I have spent 4+ years in marketing helping edtech and startup brands grow organically, including the work behind Masai School's Instagram going from 26K to 117K and LinkedIn from 50K to 160K. You can see more of what I have worked on and get in touch through the contact form at younusfardeen.com. Always happy to compare notes with people actually running experiments.