Claude GEO

Claude has the thinnest evidence base of any major AI assistant, and that shapes what honest advice looks like. Anthropic has never publicly named the search index behind Claude’s web search, and no credible large-sample study of what Claude cites exists.

Anyone selling you a Claude GEO package is therefore working from inference. Some of that inference is reasonable, and some of the numbers being quoted at you were invented by a content marketer.

What follows is what Anthropic has documented, what independent researchers have established, and where the honest answer is that nobody knows.

How Claude gets web pages

Web search reached Claude.ai as a research preview for paid US accounts on 20 March 2025, then opened to all plans worldwide on 27 May 2025. Developers got a web_search tool on the Messages API on 7 May 2025, billed at $10 per 1,000 searches on top of tokens.

Anthropic runs three crawlers with separate jobs, and blocking one leaves the other two working.

All three honour robots.txt, including Crawl-delay, and each has to be disallowed by name, per subdomain. Anthropic doesn’t publish crawler IP ranges and says IP blocking isn’t a reliable substitute, since its bots run from public cloud address space.

Citations are always on. In the API, every web search result carries the source URL, the title & up to 150 characters of the passage Claude drew from, and those citation fields don’t count toward your token bill.

The search index nobody has confirmed

Anthropic has never stated which index backs Claude’s web search, in documentation, a blog post or a press statement. That gap has been filled by circumstantial evidence rather than a company answer.

On 19 March 2025, one day before web search launched, Anthropic’s Trust Center subprocessor list added Brave Search under the Web Search category. The independent researcher Simon Willison then matched Claude’s returned results against Brave’s own output and found a BraveSearchParams field inside Claude’s tool-call schema.

That evidence is good. It’s still inference, and treating it as confirmed is how bad advice gets built. If Brave is the retrieval layer, then Brave’s index coverage of your site is the constraint, which is a different optimisation target from Google’s.

Not sure which of this matters for your site?

I am Eric. Give me 30 minutes and I will look at your site and tell you which of these things would move anything for you, and which you can skip.

5.0 from 12 Google reviews

Book a free call →

Why there is no Claude citation data

ChatGPT and Perplexity have been measured at scale by several data vendors. Claude has not, and the reason is mundane: it holds a smaller share of consumer search behaviour, so the tools that scrape AI answers point at bigger targets.

Figures do circulate. Claude averages 5.67 citations per response. It overlaps 86.7% with Brave’s top results. Structured databases account for 68% of its source mix. Every one of those numbers traces back to an agency blog with no disclosed sample, no methodology and no date.

The “experts” quoting them to you have not run a study. They have read another agency’s blog post, which read another one.

What has been measured across Claude is stability, and the result should make anyone selling AI rank tracking uncomfortable. SparkToro and Gumshoe ran roughly 3,000 identical prompts across Claude, ChatGPT and Google’s AI with about 600 volunteers, and the odds of the same brand list appearing twice came in under 1 in 100.

What carries over from ordinary SEO

Retrieval systems that read the open web reward the same things, whichever index sits underneath. A page has to exist, be crawlable, answer the question in plain language and be linked to from places that matter.

The measured correlations from the larger engines are the best available proxy. Across 75,000 brands, Ahrefs found branded web mentions correlated at 0.664 with AI citation, against 0.217 to 0.254 for raw backlink count, so being talked about beats being linked to on its own.

The tactics with no evidence behind them have none here either. A controlled test on 1,885 pages found schema markup moved AI citations by an amount indistinguishable from zero, and Ahrefs found 97% of llms.txt files got zero requests of any kind in May 2026. Nothing about Claude changes that arithmetic.

What to do about it

Decide the three Claude bots deliberately rather than in one blanket rule. Blocking ClaudeBot keeps your pages out of future training; blocking Claude-User and Claude-SearchBot removes you from answers people are actively asking for. The same split logic is walked through on the Common Crawl bot page.

There’s a live reason to think about training separately from search. Anthropic settled a US authors’ copyright class action for $1.5 billion, covering an estimated 482,000 works, with final court approval on 20 July 2026. Publishers weighing a licensing position have something concrete to point at.

Beyond that, do the work you’d do anyway: pages that answer the question, links from real publications & a brand that gets mentioned where your category is discussed. For where the engines genuinely diverge, the ChatGPT GEO, Gemini GEO and Perplexity GEO pages have the specifics.