How to get cited by Claude: fewer sources, a higher bar

Three bots, three different switches

Anthropic formalized a three-crawler structure in early 2026, splitting what used to be a single ClaudeBot into three agents with three distinct jobs, reported by Search Engine Journal on 25 February 2026 as mirroring the three-bot split OpenAI made to GPTBot in late 2024. ClaudeBot collects material for model training. Claude-SearchBot indexes content to power search-style answers. Claude-User fetches a specific page in real time when a person's question sends Claude looking for it. Anthropic's own documentation, last updated 7 April 2026, states the consequences separately for each: blocking Claude-SearchBot "prevents our system from indexing your content for search optimization, which may reduce your site's visibility," while blocking Claude-User "prevents our system from retrieving your content in response to a user query." Blocking ClaudeBot alone touches neither. A robots.txt written before this split, or one that only lists ClaudeBot the way our own crawler guide still does, is not wrong, but it is no longer complete: allowing training while accidentally leaving Claude-User blocked, or the reverse, is now a real and easy way to misconfigure access without noticing.

What a citation actually looks like

Claude's citations are not inline text-span annotations the way Gemini's grounding citations are, where each cited phrase carries its own start and end index. Claude's API documentation describes a separate structured object attached to a response: a web_search_result_location block naming the source url, its title, and up to 150 characters of the exact cited_text being drawn on. The documentation states citations are always enabled for web search results, not optional, and that the citation fields themselves do not count against a request's token usage, so there is no incentive for a shorter answer to skip naming its source. In the consumer app, Anthropic's help documentation describes the same idea in plainer terms: a response can include "direct citations to sources" and "source links for further reading" inline in the conversation, with no separate panel a reader has to click through to see what was used.

How many sources Claude actually reaches for

The API gives a developer a dial, not a fixed number. A max_uses parameter caps how many searches a single request can run, and Anthropic's own guidance is that "simple factual queries typically use 1 to 3 searches; comparative or multientity research can use 10 or more." That is a ceiling, not an average. The only measurement of what Claude actually does in practice comes from a study by Bill Widmer published on Orbit Media's blog on 2 September 2026, tracking 13,184 citations across ChatGPT, Claude, Gemini and Perplexity for 72 B2B prompts across three brands over roughly three months ending 23 August 2026. Claude averaged 3.6 sources per answer, the lowest of the four engines in that sample. Fewer sources per answer means each one that does get named matters more, and a page competing for one of 3 to 4 slots is a different problem than competing for one of 8 or 10.

Mentioned is not cited

The same study found Claude mentioned the tracked brands in 70% of its answers but cited a source for only 40% of them, a 30-point gap between being talked about and being credited. That distinction is exactly what the AEO Visibility Audit scores separately, because a brand a model already knows from training can get described accurately with no live source attached at all, which means no outbound link, no referral opportunity, and no way for a reader to verify the claim against your own site. A large mention rate can look like AEO success from inside a chat window and still convert into zero measurable citations in an audit built to tell the two apart.

What Claude actually links to

Where Claude does cite, the same study describes a distinct sourcing pattern: it favors review sites and curated listicles, with Clutch turning up as its single most-cited third-party domain in the sample, and it never cited Reddit once across the hundreds of answers tracked. That is a sharp contrast with what this site has already documented about Reddit's dominance of ChatGPT and Perplexity citations, and it means a Reddit-heavy AEO strategy built for those two engines will not transfer to Claude. The same study also measured how often Claude keeps citing the same URL when an identical question runs again two days later: only 30% of citations held, versus 65% for Perplexity and 38% for Gemini. A citation earned today is not a citation banked; Claude's answers reshuffle their sources faster than the other three engines measured.

What this means for your site

Treat the three crawler tokens as three separate decisions rather than one. If being found through Claude-powered search matters to you, confirm Claude-SearchBot and Claude-User are both allowed, independently of whatever you decided about ClaudeBot and training. Do not assume a citation record built for ChatGPT or Perplexity carries over: Claude cites fewer sources, leans on review and comparison sites over forums, and has shown zero appetite for Reddit in the only study that has measured it. And do not treat one citation as durable. With roughly a 70% chance the same query drops your URL within two days, a single win is a snapshot, not a foothold, which is the whole argument for tracking this monthly rather than checking once and moving on, the discipline AI Citation Monitoring runs on a schedule instead of by memory.

A quick check

Pull your robots.txt and look specifically for ClaudeBot, Claude-User and Claude-SearchBot as three separate lines, not one. Then ask Claude your five most valuable buyer questions directly and note, for each answer, whether your brand is named at all and whether a source link actually appears next to that mention, since the 30-point gap this page opens with means those are two different outcomes. Run the same five questions again in 48 hours and see how many of the citations held. If most of them changed, that is not a bug in your content, it is the baseline this page describes.

Sources

The three-crawler split and its consequences come from Anthropic's own crawler documentation, last updated 7 April 2026, with the formalization date and OpenAI comparison from Search Engine Journal, 25 February 2026. Citation mechanics and the max_uses guidance come from Claude's Platform Docs for the web search tool. Consumer-facing behaviour is described in the Claude Help Center article on enabling and using web search, and the feature's own rollout is dated in Anthropic's web search announcement. The 3.6-source, 70/40 mention-versus-citation, Clutch, zero-Reddit and 30% retention figures all come from Bill Widmer's 13,184-citation study on Orbit Media, published 2 September 2026. Ahrefs' own large-scale citation-overlap research, 15,000 long-tail queries published 11 August 2025, covers ChatGPT, Gemini, Perplexity and Copilot but explicitly does not include Claude, which is the gap the paragraph below names directly.

What we could not verify

Every other engine-specific guide on this site leans on a study measured in the thousands or tens of thousands of queries. No such study exists yet for Claude. The 3.6-sources and 70/40 figures above come from a single practitioner's tracking of 72 prompts across three brands, real and dated, but a fraction of the scale Ahrefs ran for the other four engines, and Ahrefs' own research explicitly did not include Claude at all. Until a study of comparable size exists, the honest position is that Claude's citation behavior is documented mechanically, through Anthropic's own API and crawler pages, far better than it is measured empirically, and this page treats those as two different kinds of evidence rather than blending them into one number.

Common questions