Claude — where llms.txt and honesty win

What is Claude SEO?

Claude SEO is the practice of structuring web content so Anthropic's Claude assistant cites the page as a source. Claude has three citation surfaces: trained knowledge (the model's pre-training corpus), search-augmented answers (Claude reads live results from its own index when the user enables web search), and on-demand browse (Claude fetches a specific URL when the user asks).

Two defining facts distinguish Claude from other AI engines. First, Anthropic is the most explicit llms.txt adopter — the crawler treats /llms.txt as a primary source-prioritization signal. Shipping a clean llms.txt has its strongest effect on Claude specifically. Second, Constitutional AI training makes Claude unusually strict about factual grounding — pages with weak attribution, hedged claims, or mismatched dates cite less reliably even when classic SEO is otherwise strong.

The Claude bot family is the largest of any AI engine — five distinct crawlers (ClaudeBot, Claude-Web, Claude-SearchBot, Claude-User, anthropic-ai). Each serves a different Claude product surface. Operators who allow only the main ClaudeBot typically miss the search-augmented and on-demand browse paths, which is where commercial-intent citations live.

Which Claude bots do what?

Anthropic ships five distinct crawlers covering training, on-demand browse, search, and legacy infrastructure. Allow all five explicitly for full Claude coverage.

Which Claude SEO signals matter most?

Ordered by leverage. Signal 1 (llms.txt) compounds with the bot accessibility floor in signal 2; everything else layers on top.

1. 1. llms.txt — Claude's strongest single signal

   Anthropic Claude is the most explicit llms.txt adopter in the industry. The crawler reads /llms.txt as a primary source-prioritization signal when ranking sources for citation. Shipping a clean, curated llms.txt (top 20-50 authoritative URLs with informative descriptions, Markdown format, under 10 KB) is the single highest-leverage Claude-targeted action available. Less than 1% of indexed sites have one — the early-adopter window is open.

2. 2. Allow the full Claude bot family

   Anthropic operates five distinct bots: ClaudeBot, Claude-Web, Claude-SearchBot, Claude-User, anthropic-ai. Blocking any of them reduces Claude's ability to reach your content via the surface that bot serves. The minimum-viable robots.txt has explicit Allow: / rules for all five — explicit posture is what AEO analyzers reward and what guarantees coverage across Claude product surfaces.

3. 3. Factual grounding — Constitutional AI penalizes weak sources

   Claude's Constitutional AI training penalizes hallucination and unsupported claims more aggressively than other engines. Pages with high factual density (attributed statistics, named-expert quotes, links to peer-reviewed or .gov sources) are preferentially cited because the answer composer can ground its response without speculation. Vague, unattributed assertions get marginalized even when content depth is otherwise strong.

4. 4. Long-form depth over breadth

   Claude's context window is among the largest in production (200k tokens for Claude 3.5 Sonnet, up to 1M for some variants). The answer composer prefers ONE deeply researched page over multiple shallow ones — it can ingest the full deep article in context. Pages with 2,000-5,000 words of substantive single-topic coverage out-cite cluster pages on the same topic for Claude specifically.

5. 5. Honest dateModified — Claude verifies dates

   Claude's answer composer cross-references visible dates against schema dateModified. Pages where the visible 'last updated' stamp does not match the schema dateModified are treated as lower-trust — Constitutional training flags the inconsistency. Get the dates right or omit them; do NOT publish mismatched dates.

6. 6. Cite your own sources — Claude weighs citation-density recursively

   A page that itself cites authoritative external sources (DOI links, .gov, .edu, recent news) is more likely to be cited by Claude. The signal is recursive: trustworthy sources cite other trustworthy sources. Two to four attributed external citations per important page is the rough threshold where citation-weight measurably shifts.

7. 7. Clean Markdown-friendly structure

   Claude was trained heavily on Markdown-formatted content and its extractor prefers content that converts cleanly to Markdown — clear heading hierarchy, real bullet/numbered lists (not styled divs), code blocks for technical content, paragraph-level topic sentences. JS-rendered visual flourishes that break Markdown conversion hurt extraction.

8. 8. Author transparency and credentials

   Anthropic's training places explicit value on transparent authorship — named authors with verifiable credentials get weighted higher than anonymous or 'editorial team' bylines. Author bio pages with LinkedIn / .edu / peer-reviewed publication links cite more reliably. The bar is higher than for ChatGPT or Perplexity; matching the bar lifts Claude citation share.

How does SeoChatAI audit Claude readiness?

• llms.txt presence, validity, and content-quality scoring
• Five-bot accessibility probing — all of ClaudeBot / Claude-Web / Claude-SearchBot / Claude-User / anthropic-ai
• Factual-density scoring with attributed-citation detection
• Date consistency between visible 'last updated' and schema dateModified
• Long-form content depth scoring (Claude favors single-topic depth)
• Markdown-extraction friendliness — heading hierarchy + list usage + code blocks

Claude SEO frequently asked

• Is Claude SEO meaningfully different from other AI engine SEO?

  Yes, in three specific ways. (1) llms.txt has its strongest effect on Claude — Anthropic is the most explicit adopter. (2) Constitutional AI training makes Claude more strict about factual grounding; weak unattributed content cites less even when classic SEO is strong. (3) The 200k+ token context window makes Claude prefer deep single-topic pages over cluster strategies. Operators optimizing for ChatGPT and Perplexity by default can leave Claude-specific lift on the table.

• How many Anthropic bots do I actually need to allow?

  All five for full coverage: ClaudeBot, Claude-Web, Claude-SearchBot, Claude-User, anthropic-ai. Skipping any one risks silent failure on the Claude surface it serves. The robots.txt cost is trivial; the citation cost of a partial allow is non-trivial. Allow all five explicitly with separate User-agent stanzas.

• Does Claude have its own search engine?

  Yes, sort of. Claude has a web-search-augmented answer surface that uses an Anthropic-managed index built by Claude-SearchBot. It is not a standalone search engine like Perplexity; it is a feature inside the Claude product. The bot accessibility logic is similar to ChatGPT Search — without Claude-SearchBot access you do not exist for the search-augmented surface.

• How do I optimize for Claude's Constitutional AI training specifically?

  Three concrete tactics. (1) Attribute every factual claim — link the source, name the expert, cite the dataset. (2) Match dates exactly — visible 'last updated' must match schema dateModified to the day. (3) Write in declarative, hedge-light prose for factual sections — Claude favors clear claims over qualified ones for the parts of its answer that need to be true. Hedging is fine for opinion sections; for facts, commit.

• Does Claude respect llms.txt?

  Yes — more explicitly than any other major AI engine. Anthropic has publicly endorsed the llms.txt standard and treats it as a primary source-prioritization signal. A site with a clean curated llms.txt jumps in Claude's source ranking measurably. Most sites still do not have one, so adopting it is the single highest-ROI Claude-targeted action available in 2026.

• How do I measure my Claude citation share?

  Three methods. (1) Analytics referrer log — claude.ai is the canonical referrer for outbound clicks from Claude-cited sources. (2) Manual probing — run 20-50 category queries in Claude (with web search enabled) and count citation appearances. (3) Third-party trackers (Otterly, Profound) probe Claude as part of their multi-engine coverage.

• What is the difference between Claude 3.5 Sonnet, Claude Opus, and the Claude Sonar variants?

  Claude 3.5 Sonnet and Claude Opus are Anthropic's general-purpose conversational models. Claude-Sonar-Pro is a Perplexity-managed variant of Sonar (Perplexity's Llama-based model) that users can pick inside Perplexity Pro Search — NOT an Anthropic product. The two are distinct: Anthropic's bots crawl for Anthropic-owned Claude surfaces; Perplexity's bots crawl for Perplexity surfaces including Sonar-Pro. Allow both bot families for full coverage.

• Will Claude cite my brand if it is already in trained knowledge?

  Sometimes — but unreliably and without an inline link. The trained-knowledge layer surfaces brand mentions without source attribution, and Constitutional training makes Claude wary of mentioning sources it cannot verify. For commercial intent you almost always want the search-augmented surface (Claude-SearchBot + Claude-Web reachable), which generates inline citation links the user can click. Optimize for the live search surface; let trained knowledge handle itself.

Keep going