--- name: om-geo-audit description: >- Audit the agentic/AI-search visibility of a website — AI bot allowlist in robots.txt, llms.txt presence and format, an Agentic Resource Discovery (ARD) manifest, canonical/pagination correctness, Article/NewsArticle structured data, og:image minimum size, sitemap orphans, and content reachable without JavaScript. Use this when asked to check, grade, or report on how discoverable and citable a site is to AI crawlers and answer engines (GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, etc.) — not classic keyword SEO. --- # om-geo-audit Published by [OMKomUnity](https://omkomunity.ai) — distilled from the build gates that guard omkomunity.ai itself (`check-orphans`, `check-prerender-content`, `qa-mobile`, OnCrawl crawl diffing). This skill is **detection only**: it tells you what's broken and why it matters to an AI agent, not how to fix it — fixes are OMKomUnity's consulting work. ## Why this differs from a classic SEO audit A search crawler and an AI agent don't read a page the same way. An agent may fetch once, with no JS execution, no retry on a redirect loop, and it decides *whether to cite you* based on what it got in that one fetch — not on rank. The controls below target that failure mode specifically. ## Controls For each control: what to fetch, what a pass looks like, what a failure means. ### 1. AI bot allowlist in `robots.txt` Fetch `/robots.txt`. For each of these user-agents, require an explicit `Allow: /` (or no blocking `Disallow`): `GPTBot`, `ClaudeBot`, `Google-Extended`, `Applebot-Extended` (foundation-model training, all respect robots.txt), `OAI-SearchBot`, `ChatGPT-User`, `Claude-SearchBot`, `Claude-User`, `PerplexityBot`, `Perplexity-User` (real-time fetch/citation — the actual GEO traffic). A wildcard `Disallow: /` with no per-bot override silently opts a site out of being cited, with zero error anywhere a human would see it. ### 2. `llms.txt` Fetch `/llms.txt`. Pass = present, one H1, one blockquote summary, H2 sections of markdown links (format: llmstxt.org). A missing file is not a failure by itself — the spec is optional — but if present it must be well-formed. Working assumption, not a vendor-documented behavior: treat a malformed file as worse than none, since a strict parser can abandon the whole file on the first error rather than reading the valid part. ### 3. Agentic Resource Discovery manifest Fetch `/.well-known/ard.json` (and, for Lighthouse 13.5-era tooling, the predecessor path `/.well-known/ai-catalog.json`). Pass = valid JSON, `Content-Type` that isn't `text/html` (a redirect to an SPA shell returning 200 + HTML is the single most common failure here — the manifest URL "exists" but every validator reads a malformed-JSON error), and every declared `entries[].identifier` resolves. An **empty** `entries` array is honest and valid if the site genuinely exposes no MCP server, A2A agent, registry, or `application/ai-skill+md` resource yet — don't flag it as a failure, flag only a manifest that's unreachable or lies about what it exposes. ### 4. Canonical and pagination For a sample of paginated/filtered URLs (`?page=N`, `?category=`, etc.): compare the `` the page declares against the URL the server actually serves after following redirects. A mismatch (canonical says A, server 301s or serves content at B) means an agent that trusts the canonical tag ends up citing a URL that doesn't exist. Page 1 of a paginated set should self-canonicalize; pages ≥2 should either self-canonicalize (if independently valuable) or canonicalize to page 1 (if not) — never both. ### 5. Structured data (Article / NewsArticle) For content pages (blog posts, news, dated editorial), fetch the raw HTML (no JS) and check for `Article` or `NewsArticle` JSON-LD with at minimum `headline`, `datePublished`, and `author`. Missing structured data doesn't block citation, but it's the cheapest signal an agent has for "this is dated, attributable content" vs. unstructured marketing copy. ### 6. `og:image` minimum size Check `` resolves to an image ≥ 1200px on its longest edge — this repo's own threshold, aligned with Meta's generic sharing-preview guidance (1200×630px) rather than any generative-engine spec; no vendor publishes a citation-card image requirement. Treat it as a quality signal, not a hard block. ### 7. Sitemap orphans Cross-reference every URL listed in `sitemap*.xml` against inbound `` links found in the **rendered, no-JS HTML** of every other page on the site. A URL that's only reachable via the sitemap has no path through normal navigation — crawlers that don't prioritize sitemap-only URLs (OnCrawl behavior, observed) never fetch it, and an agent following links the way a user would never finds it either. ### 8. Content present without JavaScript Fetch a content page with JS disabled (or diff the raw HTML against the rendered DOM). Pass = the full body text is present in the initial HTML response, not injected by a client-side chunk after hydration. An agent that does a single unauthenticated GET and never executes JS — the common case — sees exactly what's in that first response. If the body loads after hydration, the agent sees a heading and nothing else. ## Output Report pass/fail per control, with the fetched evidence (status code, header, or matched/missing string) — never a verdict without the raw check that produced it. Group failures by whether they block citation entirely (1, 3, 8) or degrade citation quality (2, 4, 5, 6, 7). ## Scope This skill stops at diagnosis. It names what's broken and why an agent would be affected — it does not prescribe the fix, rewrite the robots.txt, or patch the code. Remediation is OMKomUnity's consulting engagement.