GEO Audit

Inventory a site's Generative-Engine-Optimization surface and emit a prioritized gap report. Work top-to-bottom; report findings with concrete file paths / URLs and a fix for each. Grounded in the geo-kit playbook (embedded below so this skill is self-contained).

Step 1 — Crawler access (do this FIRST; it gates everything)

The #1 failure mode is an edge/CDN silently blocking AI crawlers. Verify from outside the app:

for ua in GPTBot OAI-SearchBot ChatGPT-User ClaudeBot Claude-SearchBot Claude-User PerplexityBot CCBot; do
  echo "$ua -> $(curl -s -o /dev/null -w '%{http_code}' -A "$ua" https://TARGET/)"
done

• 403/503 on the retrieval/user agents (OAI-SearchBot, Claude-SearchBot, PerplexityBot, ChatGPT-User, Claude-User, Perplexity-User) = critical blocker — the site cannot be cited by that engine. Usual cause: Cloudflare "Block AI Scrapers / AI Crawl Control", or a robots.txt Disallow for those agents. Flag it as blocker #1.
• On Cloudflare, check the zone-wide policy (Security → Settings → "AI Crawler and Agent access policies"), not just the per-crawler toggles in AI Crawl Control → Security, which the policy locks. "Training: Disallow" has blocked Claude-User (live in-answer fetches), because Cloudflare files it under Training. "Bot Preference Sync" is what injects the managed robots.txt block.
• A faked user-agent only tests the rule. CDNs verify real bots by IP, so confirm with real bot hits in server logs after any change.
• Fetch robots.txt and check: is there ONE coherent file? Correct sitemap URL? Are the retrieval agents allowed? Any managed/injected block (e.g. Cloudflare "Content-Signal ai-train=no")?

Step 2 — Indexability & hygiene

• SSR check: fetch a key page and strip <script> — is the primary content in the raw HTML? AI crawlers don't run JS; client-only content is invisible.
• Sitemap: does /sitemap.xml exist and cover every citable route (blog, location pages, key routes)? Any orphan pages absent from it?
• Bing: is the site in Bing Webmaster Tools? (ChatGPT grounds via Bing.) Is IndexNow set up?
• OG image / canonical: default OG image resolves (not 404)? Canonicals set?

Step 3 — Structured data

Grep the rendered HTML for application/ld+json. Inventory schema types present and missing:

• Expected where relevant: Organization + WebSite (home), BreadcrumbList, FAQPage, Article/NewsArticle (blog), plus domain schema (Event/Product/LocalBusiness).
• Check Organization for sameAs, founder, areaServed, knowsAbout.
• Flag fabricated/invalid data (e.g. an Event with a made-up startDate).

Step 4 — On-page structure (sample the top pages)

For each sampled page, does it have:

• An answer-first summary in the first screen (40–70 words)?
• Question-styled headings?
• Sourced statistics?
• An FAQ section emitting FAQPage JSON-LD?
• Visible EEAT (author, credentials, primary sources) and an updated date?

Step 5 — Off-site & entity (report, don't fix)

• Wikidata/Wikipedia entity present? Google Business Profile (for local)? Branded mentions in local/niche press? YouTube presence?
• Search the brand name unscoped in an AI engine's own search tool (Claude's WebSearch, standard mode). If an aggregator (Eventbrite, Meetup, a directory) takes the results instead of the homepage, the engine doesn't know the brand.

Step 6 — Each engine's view (if time allows)

Whether agents find, open and cite the site depends on their own search index, not Google's. Check it directly with Claude's WebSearch tool:

• allowed_domains: ["TARGET"] with a broad word → indexed at all?
• 3–5 terms where the site is top 10 on Google, unscoped → does it rank there?
• The same terms in extended mode → a ranking-depth problem? Agents almost always use standard.

"Indexed but never ranks" is an entity/authority gap, not a content gap. For a full measurement, use the geo-fanout skill.

Output — the gap report

Rank findings by leverage, not order found:

1. Blockers (crawler 403s, client-only rendering) — nothing else matters until fixed.
2. High (missing schema, no answer-first structure, sitemap gaps, no IndexNow/Bing, brand unknown to an engine's index).
3. Medium (EEAT, freshness, OG, entity signals).
4. Explicitly skip — call out any llms.txt/speakable/mass-thin-page work as folklore (no engine consumes llms.txt; page count barely correlates with AI visibility). Don't recommend them.

For each finding: what, where (path/URL), the fix, and the expected impact. If geo-kit's playbook/ is available, cite the relevant section; otherwise the rules above are self-sufficient.