How Claude chooses citationsWhat Anthropic documents — and what is honestly unknown.
Anthropic documents three Claude bots, robots.txt compliance, and Claude's capability set including context-window sizes and retrieval-mode behaviour. It does not publish a citation ranking algorithm. This page describes the documented surface and the structural signals operators control — without claiming knowledge of internal scoring we do not have.
By Martin Yarnold · UpdatedDocumented vs inferred
Documented
Inferred
Frequently asked questions
Does Anthropic publish Claude's citation algorithm?
No. Anthropic publicly documents the three Claude bots (ClaudeBot, Claude-User, Claude-SearchBot), their robots.txt compliance, and Claude model capabilities (context windows, citation behaviour modes). It does not publish a citation ranking algorithm. Specific claims about "Claude prefers X over Y" beyond the documented bot-level and model-level behaviour are inference, not vendor contract.
When does Claude cite sources at all?
Primarily when invoked with web access (Claude.ai with web mode), document context, or specific system prompts that request citation. Pure model-knowledge responses without retrieval typically do not include inline citations. Anthropic documents Claude's citation behaviour as mode-dependent — citations appear when the invocation pattern provides retrieval context. Outside those modes, citation discipline is inconsistent and not vendor-guaranteed.
Which structural signals influence whether Claude reaches your content?
The signals Anthropic documents and operators control: ClaudeBot, Claude-User, and Claude-SearchBot allowed in robots.txt; server response health (sub-300ms is target); valid JSON-LD structured data; complete metadata (title, description, canonical); substantive content. These do not guarantee citation; collectively they remove the common blockers. Specific ranking signals beyond these are not publicly documented.
Does Claude cite differently when given documents as context?
Yes — observably. When Claude is invoked with documents (PDFs, web pages provided in context) and the right system prompt or invocation mode, it tends to produce inline citations that reference specific sections of the provided sources. This is observably consistent in Claude.ai web mode and via the API with appropriate prompting. It is not the same as guaranteed attribution: model output still varies between releases and prompt patterns.
Do different Claude model versions cite differently?
Yes — model behaviour shifts between versions and tiers. Anthropic publishes release notes describing per-version capability changes. Claude Opus, Sonnet, and Haiku have different cost/latency tradeoffs and somewhat different attribution discipline in retrieval modes. The reliable approach: test against the specific Claude version you target operationally; do not generalise from one model to all.
How do I measure Claude citations against my queries?
Pick a fixed query set, run them in Claude.ai with web access enabled, and record whether your domain appears in the cited sources. The output is a citation-rate-per-query time series. Promagen Sentinel automates this on a fixed query set; manual measurement uses a structured spreadsheet plus weekly habit. Track per-version where possible — Claude behaviour shifts when Anthropic releases new model variants.