seo-ai-crawlers

seo-ai-crawlers (M14)

Controls whether AI search engines can crawl and cite the page, and whether they can read it without JS. The training-vs-search-vs-fetch distinction is everything. Reference: references/ai-crawlers.md.

Audits

Working from the PageSnapshot (rendered_dom if present, else raw_html) plus the site robots.txt:

1. Citation access: are retrieval/citation bots — OAI-SearchBot, Claude-SearchBot, PerplexityBot, Bingbot — actually allowed (not caught by a broad Disallow: / or a wildcard block)? Confirm Googlebot is not blocked and the Googlebot (search) vs Google-Extended (Gemini training control) split is correct.
2. User-agent classification: bucket every AI agent in robots.txt into training (GPTBot, ClaudeBot, Google-Extended, Applebot-Extended, CCBot), search/retrieval (OAI-SearchBot, Claude-SearchBot, PerplexityBot), and user-triggered fetch (ChatGPT-User, Claude-User, Perplexity-User). Match user-agents case-insensitively; treat the table in references/ai-crawlers.md as a starting set, not exhaustive.
3. Renderability for non-JS crawlers: pull the M4 (seo-crawl-render) render result — most AI crawlers do not execute JS. If primary content only appears in rendered_dom and is absent from raw_html, flag it as invisible to AI retrieval.
4. llms.txt / llms-full.txt (also covers M21): presence at the site root, valid Markdown structure (H1 title, summary blockquote, sectioned link lists), and that linked URLs resolve. Follow references/ai-crawlers.md.

Fixes

• AUTO (fixable: auto): a citation-friendly robots.txt preset, choice-gated — the user picks allow-citations (default: allow search/retrieval, opt out of training), allow-all, or block-all. Deterministic, additive, verifiable; emitted as a diff for fix.
• AUTO (fixable: auto), disclosure-gated and scored 0: llms.txt / llms-full.txt, generated from the site's own structure only on explicit request (fix --category llms), shown as a diff before writing. Additive and deterministic, but never sold as proven ranking value — the disclosure that it is low/uncertain impact is shown every time.
• ADVISORY (fixable: advisory): edge/WAF block for bots that ignore robots.txt (e.g. Bytespider) — the tool never writes infra config.

Never fabricate sitemap URLs, contact emails, or link targets — ask the user or leave a clearly-marked TODO placeholder.

Verification

• node ${CLAUDE_SKILL_DIR}/../../scripts/parse-robots-sitemap.mjs --url <u> (robots_parse) — parses robots.txt, resolves the effective directive for each AI user-agent, and confirms it.
• dom_assert against the M4 render result for the renderability check (content present in raw_html).
• When the required data tier is unavailable (e.g. robots.txt unfetchable, no M4 render result), status is needs_api — never a false pass.

Findings

Findings conform to schema/finding.schema.json. evidence.observed quotes the page/robots line; verification.reproduce is the runnable command above; expected_impact is banded + confidence-tagged (no naked %). Examples:

• M14.citation_bots.blocked — Disallow: / reaches OAI-SearchBot/Claude-SearchBot/PerplexityBot (status fail, severity 4, fixable: auto, axis ai, confidence established).
• M14.render.content_js_only — primary content in rendered_dom but absent from raw_html, invisible to non-JS AI crawlers (status warn, severity 4, fixable: advisory, axis ai, confidence directional).
• M14.llmstxt.missing — no /llms.txt at site root (status warn, severity 1, fixable: auto, axis ai, confidence speculative, scored 0).

Honesty

• Blocking a training bot does NOT block the matching search bot — they are separate user-agents (GPTBot ≠ OAI-SearchBot, ClaudeBot ≠ Claude-SearchBot). Many "block AI" guides get this wrong.
• Bytespider and some agents frequently ignore robots.txt; a robots rule is best-effort. Real enforcement needs an edge rule / WAF (advisory).
• robots.txt controls crawling, not indexing — to keep a page out, use a noindex meta tag and don't also Disallow it.
• llms.txt has only partial vendor support (Anthropic and Perplexity honor it in retrieval; Google does not use it for AI Overviews/AI Mode; OpenAI uncommitted) and is also useful as IDE/coding-agent context. Still low/uncertain — scored 0, never presented as proven ranking value.