GEO audit

Answer one question with evidence: when an AI answer engine looks for an answer in this page's topic, can it find, read, trust and quote this page? Score it, show the evidence, and list the fixes in the order that moves the score most.

When to use

• "How does ChatGPT / Perplexity see this page?", "GEO audit", "AEO audit", "AI readiness".
• Before publishing or redesigning a page that should get cited.
• As the diagnostic step when ai-visibility-report shows a page never gets cited.

Not for: Google crawl and speed problems (technical-seo-audit), writing the content (answer-engine-content), or generating the llms.txt itself (llms-txt).

Inputs

• URL — one page. With --site, also audit the home page, the robots.txt, llms.txt, the sitemap, and up to five pages the user names or the sitemap ranks first.
• Target query (optional) — the question this page should be the answer to. If absent, derive it from the H1 and say so.
• Brand (optional, Era) — with --brand, pull the brand's citation data so the audit knows whether the engines already cite this domain and which sources they prefer.

Steps

1. Fetch what the engines fetch. Get robots.txt, llms.txt, llms-full.txt, the sitemap, and the page itself. Run scripts/check_ai_access.sh <url> when a shell is available — it prints the AI-crawler rules, the status of the three files, and the page's response headers. Otherwise fetch them with the web tool and read the same things by hand. Note the HTTP status, canonical, noindex, nosnippet, max-snippet, and whether the main content is present in the initial HTML or only rendered by JavaScript.
2. AI crawler access. Compare robots.txt against the agents in references/ai-crawlers.md. For each engine family, state: allowed / blocked / unlisted. Flag the common mistake of blocking the training crawler and the search/answer fetcher together (GPTBot vs OAI-SearchBot, ClaudeBot vs Claude-User, Google-Extended vs Googlebot).
3. Answer shape. Read the page as an engine would: does the first screen answer the target query directly, in one to three sentences, before context? Are the H2/H3s questions or claims the engine can map to sub-queries? Are definitions, numbers, steps and comparisons in extractable units (short paragraphs, lists, tables) rather than buried in prose? Count: direct-answer paragraphs, question headings, tables, lists.
4. Citability. What would an engine quote? Count concrete statistics with a stated source and date, named entities (products, people, standards, places), quotable one-sentence claims, and original data (benchmarks, surveys, first-hand tests). Note vague claims ("industry-leading", "fast") that no engine can cite.
5. Entities and structured data. Check for JSON-LD (Organization, Article, Product, FAQPage, HowTo, BreadcrumbList), sameAs links, a consistent brand and product name across title, H1 and schema, and dateModified. Validate what exists; list what is missing for this page type.
6. Trust signals. Author with a bio and credentials, organisation identity, contact page, last-updated date visible on the page, outbound citations to primary sources, consistency between what the page claims and what third-party sources say.
7. Era context (when --brand is given and Era is connected). list_brands → brand; list_cited_domains(brand_id, days=90) and get_top_citations to see whether this domain is cited at all and which third-party domains the engines prefer in the category; list_search_queries(brand_id) to check whether the target query is one the engines actually run. Fold this into the score as the Authority category evidence.
8. Score and write. Apply references/scoring.md (six categories, 100 points). Each deduction is a line with the evidence. Then produce the report below.

Output

GEO audit — {url}
Score 58/100  · Access 16/20 · Answer shape 9/20 · Citability 8/20 · Entities & schema 10/15 · Authority 9/15 · Freshness 6/10
Target query: "{query}"  (derived from H1 / given by user)
Verdict: {one sentence — e.g. "Readable by every engine, but nothing on the page is quotable: no numbers, no sources, no direct answer."}

Then:

• Access — table: engine family · crawler · rule · effect. Plus the three files (robots.txt, llms.txt, sitemap): present / missing / malformed.
• Answer shape — what the first screen says, the question headings found, the extractable units counted.
• Citability — the quotable sentences found (quote them), the statistics with and without sources, the vague claims to replace.
• Entities & schema — the schema present (validated), missing types for this page type, naming inconsistencies.
• Authority & freshness — author, organisation, dates, outbound citations, and (with Era) whether the engines already cite the domain.
• Fixes, ranked — a numbered list; each item says the expected score gain, the effort (minutes / hours / days), and which sibling skill does it (llms-txt, schema-markup, answer-engine-content, on-page-seo, technical-seo-audit).

With --site, add a per-page table (url · score · top fix) and a site-level section for robots.txt, llms.txt and the sitemap. When the geo-page-auditor subagent is available (it ships with this plugin), run one instance per page in parallel and merge their blocks; otherwise audit the pages in sequence with the same rubric.

Guardrails

• Only fetch public URLs the user gave or that the sitemap lists. Never submit forms or follow login flows.
• Report what the fetch returned, not what the CMS claims. If the page needs JavaScript to show its content, say that the engines' fetchers may see an empty page, and score accordingly.
• Robots rules are read literally; do not assume an engine honours or ignores them beyond what the rules say.
• A score is a diagnosis, not a ranking prediction. Say so once in the Method line.

Era tools used (optional)

• list_brands (read) — resolution.
• list_cited_domains, get_top_citations (read) — is this domain cited, and which sources win.
• list_search_queries (read) — is the target query one the engines run.

Without Era

Everything except step 7 runs on the public page alone. Skip the Era line in Authority and note that citation data would show whether the engines already trust the domain.