What Does Claude Cite In Search Results

• Claude's citations come from the web search tool Anthropic documents in its API documentation, which returns a final…
• Anthropic has not officially named the search provider behind Claude's web results.
• In the datasets examined, community and social platforms are rarely cited.
• Claude does not search every query.
• Anthropic's three crawlers are ClaudeBot, Claude-SearchBot, and Claude-User, as summarised by ALM Corp from Anthropic's documentation.

Claude's citations come from the web search tool Anthropic documents in its API documentation, which returns a final response with cited sources drawn from search results. Three versions exist:

• Web_search_20250305
• Web_search_20260209 (adds dynamic filtering)
• Web_search_20260318 (adds response inclusion control)

The documentation confirms the mechanism but not the full list of source types Claude favours. That gap is filled by independent analyses, which agree on a pattern: Claude leans toward specialist and practitioner content rather than mainstream headlines. Treat any single percentage with caution, because each study samples different domains and time windows. The reliable takeaway is structural: Claude cites pages that answer a specific question in extractable form.

These figures come from different samples and categories, so they should not be added together or compared as if they measured the same thing.

Anthropic has not officially named the search provider behind Claude's web results. Rankry assembles an evidence chain pointing to Brave Search, drawing on subprocessor lists, exposed API parameters, and independent citation-overlap tests. According to Profound's analysis, cited by Choco Media Agency, 86.7% of URLs Claude cites overlap with Brave's top organic results, compared with 26.7% for ChatGPT against Bing's top results.

That is strong correlation, not a confirmed contract. Rankry also flags TurboPuffer as a second search-related subprocessor since May 2026, with its role not publicly documented. Until Anthropic states otherwise, the accurate framing is that Claude's retrieval appears to track Brave's index, and the TurboPuffer role remains an open question.

In the datasets examined, community and social platforms are rarely cited. OtterlyAI recorded zero Reddit citations and 0.9% social share overall, with LinkedIn making up 43.1% of that small social slice. Oltre.ai's analysis of 2,170 cited URLs found no Reddit, LinkedIn, YouTube, Medium, Quora, or Hacker News pages in its dataset.

Large news and Q&A platforms do appear in cross-assistant research: a 2024 analysis published in Nature found over 40% of citations across major models referenced top-100 news sites, Wikipedia, or large Q&A platforms. Claude's pattern therefore differs from that broader mix. For a brand, a community post is unlikely to be the citable asset; a standalone, self-contained page on your own domain is.

Claude does not search every query. In a test run on 23 September 2026, ITGrows found a search-enabled model answered 10 of 20 German business questions without searching and cited no one. Where it did search, the model cited primary sources on 9 of 10 questions. These sources included laws, ministries, OECD, and large accounting firms. Questions about stable facts may never trigger retrieval. Time-sensitive or evidence-seeking questions are more likely to do so.

Claude's Research feature is a more deliberate retrieval path. According to Anthropic, Research uses live web results powered by Brave Search to return answers with citations. A page that targets a clear, current, verifiable question gives Claude a concrete reason to fetch it.

Anthropic's three crawlers are ClaudeBot, Claude-SearchBot, and Claude-User, as summarised by ALM Corp from Anthropic's documentation. ClaudeBot collects training data, Claude-SearchBot indexes content for Claude's search, and Claude-User fetches pages in real time when a user's question requires it. All three honour robots.txt Disallow rules.

Proud Brands reports that blocking Claude-SearchBot and Claude-User may reduce visibility in Claude's user search results, while blocking ClaudeBot keeps content out of future model training. These are distinct decisions. Many publishers want citation visibility but do not want training use, so they allow the two search-facing agents and disallow ClaudeBot. Check your robots.txt for all three user agents, because a broad wildcard block can silently remove you from Claude's retrieval.

Dynamic filtering arrived with web_search_20260209 and later, and it applies to Claude 4.6 and newer models. According to Anthropic's web search documentation, Claude writes and runs code to filter search results before they enter the context window. This reduces token use.

The practical consequence is that a page's content is not guaranteed to be read in full. Claude may keep only the passages its code judges relevant. Pages with a clear answer in the opening lines, labelled headings, and tight paragraphs survive filtering better than pages where the key claim is buried under navigation text or long preambles.

Each section should be written so the relevant sentence stands alone, because filtering can isolate that sentence from the rest of the page. Pages that follow this structure give the filtering step a clean unit to keep.

According to Oltre.ai's analysis, 56% of Claude citations came from /blog/ paths, only 3% were homepages, and 47% used listicle-style URL patterns such as best, top-10, vs, and alternatives. These are correlations in one dataset, so treat them as patterns to test rather than rules.

The more durable principle is that Claude cites pages that resolve a specific question. A comparison page, a dated how-to, or a documented method gives the model a discrete, quotable unit. A generic homepage gives it nothing to extract. Structure the page around the question, state the answer first, and keep each section self-contained.

The engines draw from different indexes and cite different things. According to OtterlyAI, only 13% of cited domains overlapped between Claude and ChatGPT, and about 4.2% at the URL level.

Claude's retrieval appears to track Brave's index, while ChatGPT's tracks Bing's, which is why the same query can produce entirely different source lists. Claude also leans toward practitioner and documentation content. According to eSEOspace's 2025 GEO study, 63% of Claude citations pointed to niche SaaS blogs, documentation pages, or practitioner articles.

Google AI Overviews draw more heavily on established search rankings. Optimising for one engine does not transfer automatically to the others. Measure citations separately per engine, and treat Claude as its own retrieval environment with its own crawlers and source preferences.

The specific figures and claims on this page are grounded in the following sources, reviewed at the time of writing:

• Citations on the Anthropic API
• Claude Citation Study Reveals What Actually Gets Cited
• Web search tool - Claude Platform Docs
• How Claude Picks Sources: A Technical Breakdown of What Claude Cites and Why
• How Do You Get Your Business Cited by Claude?
• How Claude Finds and Cites Sources