seo-ai-crawlers (M14)

Two separate questions live in this module, and only one of them has vendor documentation behind it:

1. Google's AI features. A page can appear in AI Overviews / AI Mode only if it is indexed and eligible to be shown in Google Search with a snippet. Those gates are documented, so M14.ai_eligibility.* carries the only established findings here.
2. Third-party AI crawlers. Who may fetch and cite the page — OpenAI, Anthropic, Perplexity, Apple, Amazon, Meta, Mistral, DuckDuckGo, Cohere. Access is documented per user-agent; the citation effect is not, so those findings stay directional.

The training-vs-retrieval-vs-user-fetch distinction decides every robots recommendation. Reference: references/ai-crawlers.md (UA table with sources, Content-Signal, eligibility gates, presets). AI discovery files (llms.txt, agents.md, /.well-known/ucp, agentic sitemap) are module M21 — see seo-ai-discovery; agentic-commerce readiness is M18 (seo-ecommerce).

Inputs

Work from the PageSnapshot named in your dispatch envelope: read parsed from <run_dir>/pages/<slug>.json plus robots_directives, headers (x-robots-tag), and render{needed,used,delta}; Grep pages/<slug>.html for verbatim evidence. <run_dir>/site/robots.json carries, for the audited path, verdicts, effective_access{[bot]:{class, fetch, url_allowed, via, rule, signals, doc_url, robots_reliability}}, content_signals{present, placement, global, by_group[], malformed[], cloudflare_managed_hint}, groups, and ua_table_version. When the run included a user-agent diff, <run_dir>/ua-diff.json holds per-UA status/title/h1/word_count/canonical/robots/jsonld. Deterministic findings already emitted by audit.mjs are listed in <run_dir>/findings.deterministic.json — do not re-emit those ids; add model-judged findings only. If invoked directly with a URL/path and no snapshot exists, first run node "${CLAUDE_PLUGIN_ROOT}/scripts/snapshot.mjs" <target> --out "${CLAUDE_PLUGIN_DATA}/runs" and use the printed snapshot path.

Audits

Working from the PageSnapshot (parsed_rendered when render.used is not none, else parsed) plus site/robots.json:

1. Google AI-feature eligibility (the established gate). Collect every <meta name="robots"> / <meta name="googlebot"> tag and every X-Robots-Tag header value (header and meta both count; the most restrictive wins). Read noindex/none, nosnippet, max-snippet:N, and the data-nosnippet attributes — for data-nosnippet, measure how much of the main content it covers and specifically whether it wraps the H1 or the lead passage. Also confirm Googlebot itself is not disallowed for this path. Record the scope: a directive coming from a template or a site-wide header is scope: template|site; one hand-written into a single document is scope: page.
2. Effective per-bot access. Read effective_access for the audited path — not for /. A site can allow / and disallow /blog/, so root-level posture is not an answer. Check the retrieval/citation bots (OAI-SearchBot, Claude-SearchBot, PerplexityBot, Amazonbot, DuckAssistBot, Google-CloudVertexBot), the classic engines (Googlebot, Bingbot, Applebot), the training tokens (GPTBot, ClaudeBot, Google-Extended, Applebot-Extended, CCBot, Meta-ExternalAgent, cohere-ai, Bytespider), the user-fetch tokens (ChatGPT-User, Claude-User, Claude-Web, Perplexity-User, Meta-ExternalFetcher, MistralAI-User), and OAI-AdsBot. Current OpenAI tokens ship as GPTBot/1.4 and OAI-SearchBot/1.4; match user-agents case-insensitively on the token, never on the full UA string. Anthropic publishes its crawler IP ranges at claude.com/crawling/bots.json. Quote ua_table_version in the report so a stale table is visible.
3. Content-Signal posture. content_signals reports the search / ai-input / ai-train keys, whether the line is global (Cloudflare's managed robots.txt appends it after all groups) or inside a User-agent group, and any malformed pairs. A group line overrides the global line for that group. A Disallow always wins over a signal: Disallow: / blocks the bot whatever ai-train says, and Allow + ai-input=no means the page is still fetchable while consent is withheld — say exactly that, never "blocked".
4. User-agent divergence. Compare what the origin serves to default, googlebot, gptbot, oai-searchbot, and claude-searchbot: differing status, title, H1, word count, canonical, robots directives, or JSON-LD. Separate a bot challenge (403/429/503, interstitial, CAPTCHA markup, near-zero word count) from a genuine content difference. The search-axis cloaking judgment — Googlebot vs a browser being served different content — belongs to M2 (M2.cloaking.ua_content_divergence); M14 owns only the AI-bot access side.
5. Renderability for non-JS crawlers. Use the M4 render block. If the H1, the primary body, or the JSON-LD exists only in rendered_html_path and is absent from raw_html_path, the page is not reliably readable by AI crawlers that do not execute JavaScript.
6. Google-Extended sanity. If Google-Extended is disallowed, state plainly that this does not remove the site from AI Overviews or AI Mode — it limits Gemini training and grounding only.

Fixes

• AUTO (fixable: auto): a citation-friendly robots.txt preset, choice-gated — the user picks allow-citations (allow search/retrieval, opt out of training), allow-all, or block-all. Presets in references/ai-crawlers.md; each ends with a Content-Signal: line that restates the same intent in that vocabulary. Deterministic, additive, verifiable; emitted as a diff for fix.
• PROPOSED (fixable: proposed): removing a nosnippet / max-snippet:0 / noindex that the user confirms is unintended, and lifting data-nosnippet off the primary content. Never auto-written: a snippet control is often deliberate (paywall, licensing).
• ADVISORY (fixable: advisory): edge/WAF rules — unblocking a challenged AI bot, or blocking an agent that ignores robots.txt (e.g. Bytespider). The tool never writes infrastructure config. Never fabricate sitemap URLs, contact emails, or link targets — ask the user or leave a clearly-marked TODO placeholder.

Verification

• node "${CLAUDE_PLUGIN_ROOT}/scripts/parse-robots-sitemap.mjs" --url <final_url> --path <audited path> (method robots_parse) — resolves the effective directive per user-agent against the audited path and returns content_signals, effective_access, and ua_table_version. Use --robots <robots_url> or --file <path> for an offline robots file.
• node "${CLAUDE_PLUGIN_ROOT}/scripts/ai-eligibility.mjs" --snapshot <pages/<slug>.json> (or --url <u>) — every meta-robots/X-Robots-Tag directive, max-snippet value, and the data-nosnippet coverage share, including whether it wraps the H1/lead (methods dom_assert + header_check).
• node "${CLAUDE_PLUGIN_ROOT}/scripts/ua-diff.mjs" --url <final_url> --ua default,googlebot,gptbot,oai-searchbot,claude-searchbot (method ua_diff) — per-UA response fields plus challenge detection.
• render_diff against the M4 result for the renderability check (primary content present in raw_html_path).
• When the required tier is unavailable (robots.txt unfetchable, no M4 render result, no UA diff run), status is needs_api or manual_review — never a false pass.

Findings

Findings conform to schema/finding.schema.json. evidence.observed quotes the page, header, or robots line verbatim; verification.reproduce is one of the runnable commands above; expected_impact is banded + confidence-tagged (no naked %). Severity policy: 5 is reserved for catastrophic, eligibility-killing facts at site/template scope; 4 major · 3 moderate · 2 minor · 1 cosmetic · 0 informational — and only an established severity-5 fail in an active category can cap a score (references/scoring-model.md).

Google AI-feature eligibility — axis ai, confidence established except where noted; sources: Google AI features and your website and the robots-meta / X-Robots-Tag spec:

• M14.ai_eligibility.not_indexable — noindex/none, or Googlebot disallowed for this path (fail, severity 5 at site/template scope, 4 for a single page, fixable: proposed).
• M14.ai_eligibility.nosnippet — nosnippet in a meta tag or header (fail, severity 5 site/template, 4 single page, fixable: proposed).
• M14.ai_eligibility.max_snippet_zero — max-snippet:0 (fail, severity 5 site/template, 4 single page, fixable: proposed).
• M14.ai_eligibility.max_snippet_low — a positive max-snippet:N below our 50-word cutoff (warn, severity 2, fixable: advisory, confidence directional — the cutoff is ours, Google publishes no threshold).
• M14.ai_eligibility.data_nosnippet_primary — data-nosnippet wraps the H1 or the lead passage (fail, severity 4, fixable: proposed).
• M14.ai_eligibility.data_nosnippet_partial — data-nosnippet present but only around ancillary blocks (pass, severity 1).
• M14.ai_eligibility.ok — indexable and snippet-eligible, no suppressing directive (pass, severity 3).
• M14.google_extended.blocked_info — Google-Extended disallowed; informational, because it does not affect AI Overviews or AI Mode (pass, severity 1).

Third-party crawler access — axis ai:

• M14.citation_bots.blocked — a Disallow reaches OAI-SearchBot / Claude-SearchBot / PerplexityBot (or another retrieval bot) on the audited path (fail, severity 4, fixable: auto). Confidence is established only when at least one blocked token is vendor-documented including its robots.txt behaviour (lib/bots.mjs robots_reliability: 'documented' with a doc_url); when the block reaches only legacy or undocumented tokens (Claude-Web, MistralAI-User, or a limited one like Meta-ExternalFetcher) they are still listed as evidence, marked as such, and the confidence drops to directional. The citation loss itself is stated as directional in rationale either way.
• M14.retrieval.allowed — every retrieval/citation bot may fetch the audited path (pass, severity 4, confidence directional).
• M14.render.content_js_only — primary content, H1, or JSON-LD only in rendered_html_path (warn, severity 4, fixable: advisory, confidence directional).
• M14.access.ai_bot_challenged — an AI user-agent gets a challenge/interstitial or a near-empty body where the default UA gets the page (warn, severity 4, fixable: advisory, confidence directional).
• M14.access.googlebot_blocked_at_edge — the spoofed Googlebot UA is refused at the edge (warn, severity 2, fixable: advisory, confidence speculative — refusing an unverified spoof is legitimate defence; verify with reverse DNS before acting).

Content-Signal — axis ai, never established (no vendor documents compliance; the IETF draft expired 2026-04 and defines vocabulary only):

• M14.content_signal.ai_input_no — ai-input=no (warn, severity 3, fixable: advisory, directional): consent withheld for answer-engine input while the page stays fetchable.
• M14.content_signal.search_no — search=no (warn, severity 3, fixable: advisory, directional).
• M14.content_signal.ai_train_no — ai-train=no (pass, severity 1): a legitimate, low-cost preference that costs no citations.
• M14.content_signal.malformed — unknown key, non-yes|no value, or an unparseable pair (warn, severity 2, fixable: auto).
• M14.content_signal.absent — no Content-Signal line (not_applicable, severity 0 — absence is not a defect).

Probes (owned by the geo command, listed here because they carry the M14 prefix): M14.probe.web_presence and M14.gsc_ai.impressions_present|zero_impressions|unavailable are severity 0, live under report.json.probes, and never move or cap a score.

Honesty

Google's own statements bound every claim in this module:

• A page reaches AI Overviews / AI Mode by being indexed and snippet-eligible in Search. There is no AI-specific file, token, schema type, or rewriting step that adds eligibility.
• Google-Extended does not affect AI Overviews or AI Mode. It limits training and Gemini/Vertex grounding. Blocking it is a training choice, not an AI-visibility choice.
• Google Search ignores llms.txt. It is reported under M21 at weight 0 and never moves the M14 category.
• Search Console's generative-AI report exposes impressions only, and the Search Analytics API exposes no AI dimensions (checked 2026-08) — the tool never claims API-level AI citation data.

Further caveats:

• Content-Signal is metadata, not enforcement. No AI vendor documents compliance with it; every M14.content_signal.* finding is directional at most. A Disallow line does the work — never present a signal as a substitute, and never call an ai-input=no page "blocked" when it is still served.
• UA-diff results are spoofed-UA results. We send the token, not a verified crawler: a site may legitimately refuse an unverified Googlebot string, and a CDN may serve a challenge to any unknown client. Treat divergence as a lead to verify (reverse DNS, server logs, Search Console URL Inspection), never as proof of cloaking. That is why M14.access.googlebot_blocked_at_edge is speculative.
• Blocking a training bot does not block the matching search bot — GPTBot ≠ OAI-SearchBot, ClaudeBot ≠ Claude-SearchBot. Most "block AI" guides get this wrong and cost the site citations.
• Bytespider and some user-fetchers (Meta documents that Meta-ExternalFetcher may bypass robots.txt) do not reliably honor robots.txt; a robots rule is best-effort and real enforcement needs an edge rule (advisory).
• robots.txt controls crawling, not indexing — to keep a page out of the index use noindex and do not also Disallow it, or the crawler never sees the directive.