Claude citation

Claude citations are the source attributions produced by Anthropic's Claude when its web search tool returns real-time web content for grounding answers. Citations appear inline as source chips or pill-shaped tags in the consumer surfaces (claude.ai web app, Claude Desktop, mobile apps for iOS and Android, Claude Code), and as structured web_search_result_location objects in responses from the Anthropic API¹. The feature was announced 2025-03-20 in feature preview for US paid users and rolled out globally to all Claude plans on 2025-05-27².

Claude citations are a distinct measurement target from Perplexity citation, Microsoft Copilot citations, AI Overview citation, AI Mode, and AI dev tool citations: each engine appears to use different retrieval and citation behavior, and the same domain can be cited heavily by one engine while absent on another. Per-engine measurement is the defensible strategy for tracking presence in the Claude source pool specifically.

Status in 2026

Anthropic's documentation describes the web search tool verbatim as giving "Claude direct access to real-time web content, allowing it to answer questions with up-to-date information beyond its knowledge cutoff," with the explicit statement that "the response includes citations for sources drawn from search results"¹. Five product surfaces matter for citation tracking:

• claude.ai web app: the primary consumer surface. Citations render as small domain-labeled chips attached to the relevant portion of the answer; clicking opens the source URL in a new tab.
• Claude Desktop (macOS / Windows) and mobile apps (iOS / Android): appear to use broadly similar source-chip or pill-style citation rendering, adapted by platform. Citations appear inline; tapping or clicking opens the source URL.
• Claude Code: Anthropic's CLI agent. In Claude Code sessions where web search is available, cited sources may appear as URLs in terminal output. Reached general availability 2025-05 after a 2025-02 preview launch.
• Anthropic API: developers integrate Claude's web search via tool use. Two documented tool versions are available: web_search_20250305 (original; basic web search) and web_search_20260209 (newer; adds dynamic filtering where Claude writes and executes code to post-process search results before they enter the context window). The newer version supports Claude Mythos Preview, Claude Opus 4.7 and 4.6, and Claude Sonnet 4.6, and requires the code execution tool to be enabled. Citations are returned as structured web_search_result_location objects with url, title, encrypted_index, and cited_text (up to 150 characters) fields. Anthropic's documentation states verbatim that "Citations are always enabled for web search."
• Cloud deployments: web search is available on the Anthropic API directly, Microsoft Foundry, and Claude Platform on AWS. Vertex AI supports only the basic tool version (no dynamic filtering). Web search is not available on Amazon Bedrock.

Detection methodology by surface: referrer-based detection works differently across the surfaces.

Server-log filtering captures only the claude.ai web subset reliably; the other four surfaces require active probing or downstream-app cooperation.

Anthropic charges the web search tool at $10 per 1,000 searches in the API, in addition to standard token costs. Web search results retrieved during a conversation count as input tokens; citation fields (cited_text, title, url) do not count toward token usage.

How to apply

Three layers of practitioner action for optimizing inclusion in the Claude web search source pool:

• Be in the retrieval index Claude reaches at search time: Anthropic does not document the exact retrieval backend for the web search tool. Anthropic operates ClaudeBot with documented sub-categories for different uses (refer to Anthropic's crawler documentation for current purpose declarations), but it is not vendor-confirmed whether Claude's web search results are sourced entirely from a Claude-owned index, from a partner search backend (Brave Search has been reported as a backend in some industry discussion), or from a combination. The practitioner action is to allow ClaudeBot in robots.txt plus ensure broad search-engine indexability (clean sitemap, allow major crawlers, no JavaScript-only content), so the page is reachable by whichever backend Claude queries.
• Optimize for citation-friendly content shape: the Claude API documentation includes the policy statement that "When displaying API outputs directly to end users, citations must be included to the original source." This means Claude WebSearch citations are durable: when a developer builds an app on Claude's web search, the citation back to your domain must remain visible. Optimizing for clean answer-block structure, cite-ability, single-question-single-answer paragraphs, source-supported claims, and primary-source attribution makes a page easier to retrieve, summarize, and cite. Whether Claude's retrieval explicitly weights these features is not vendor-published; the practitioner discipline applies regardless.
• Track Claude citations separately from other engines: Claude's citation patterns can diverge from Perplexity, ChatGPT, Microsoft Copilot, and Google AI surfaces. Aggregating "AI citation rate" into a single number hides per-engine signal; the attribution rate entry covers the measurement decomposition.

What to skip:

• Paying for Claude-specific citation tracker tools in month 1. There is no publisher-side dashboard from Anthropic; tracking is via active probes (asking the consumer app questions and observing citations) and downstream-app cooperation. In the first month, manual probing plus clean answer-block discipline may be sufficient for small sites before paying for Claude-specific tracking.
• Submitting your domain directly to Anthropic. No such submission portal exists; Claude reaches public content via its crawl plus any partner search backend.
• Treating Anthropic API tool-use citations as a publisher-visible analytics signal. Server-to-server API calls produce no identifiable referrer; the measurement gap is structural.

What remains contested or unverified

• The exact composition of Claude's web search retrieval backend. Anthropic documents the tool API surface but does not name the underlying search index or partner mix. Whether Claude crawls primarily through ClaudeBot, queries a partner search API (Brave Search has been reported in industry discussion but is not vendor-confirmed), or blends sources by query type, is not publicly documented.
• How web_search_20250305 (basic) and web_search_20260209 (with dynamic filtering) compare in citation outcomes for matched queries. Both return the same web_search_result_location schema, but whether the cited URL sets differ meaningfully has not been systematically studied.
• The boundary between Claude's web search tool citations (real-time web content) and Anthropic's separate Citations feature (for grounding on user-provided documents). They are different API surfaces with different intended uses; consumer-app citation behavior may blend both depending on context, and the user-visible distinction is not always explicit.
• Whether allowing ClaudeBot in robots.txt is necessary for inclusion in the web search pool, given that the retrieval backend is not fully disclosed. Two scenarios with different practical implications: if Claude's web search routes primarily through a partner index (Brave Search has been reported as one such partner), ClaudeBot's robots.txt status may have limited effect on whether your content is reachable; conversely, if Claude uses a proprietary index, partner crawlers' permissions would be irrelevant and ClaudeBot allow becomes the controlling signal. Practitioner consensus has not converged on which scenario dominates in 2026.

How it relates to other concepts

• Parallel surface category to Perplexity citation, Microsoft Copilot citations, AI Overview citation, AI Mode, and AI dev tool citations: each is a distinct citation-surface family with its own measurement story. Aggregating across them into one "AI citation rate" loses per-engine signal that matters for prioritization.
• Crawl-side dependency on AI crawler bots: ClaudeBot is Anthropic's declared crawler, but the web search retrieval backend may also include partner indexes. Allowing ClaudeBot is one prerequisite for the declared crawler, but neither necessary nor sufficient to guarantee Claude's web search reaches a page.
• Measurement input to attribution rate and citation-share frames: Claude is one of the five major engines those frames track, with its own per-query citation pattern that does not generalize from any other engine.
• Cross-cutting with AI dev tool citations: Claude Code, Claude Desktop, and Anthropic API integrations all sit at the intersection of "Claude as a vendor surface" (this entry) and "AI dev tools as a category" (the dev-tool entry). The two entries cover different cuts of overlapping product behavior.