The Complete GEO Audit Checklist (60 Points)

A GEO audit works from the ground up: an AI engine can only cite what it can reach, index, retrieve, and extract — and it prefers sources with authority. This checklist walks six layers in dependency order. Fix failures top to bottom, because each layer depends on the ones above it. A perfect passage means nothing if a WAF rule blocks the crawler that would read it.

Layer 1 — Crawl access

• robots.txt explicitly allows OAI-SearchBot, ChatGPT-User, PerplexityBot, and Claude-SearchBot — these bots produce citations and must never be blocked.
• Training bots (GPTBot, ClaudeBot, Google-Extended) are allowed if you want parametric memory — recommended for B2B SaaS wanting brand recall.
• CDN/WAF layer verified separately — Cloudflare "Block AI Scrapers" and Vercel's bot firewall override robots.txt silently.
• Each critical UA returns HTTP 200 via curl -A "<UA>" -I — a live fetch is the only proof that access actually works.

Layer 2 — Indexing

• Site verified in Bing Webmaster Tools — non-negotiable, because ChatGPT and Copilot lean on Bing's index.
• Sitemap submitted with accurate lastmod values so engines can rank freshness correctly.
• IndexNow key file hosted at site root and pinging on publish for instant Bing push.
• Google Search Console confirms money pages are indexed, not just discovered.

Layer 3 — Retrievability

• Content present in the initial HTML response (SSG or SSR) — GPTBot and PerplexityBot do not execute JavaScript.
• Core Web Vitals in range: LCP under 2.5s, INP under 200ms (INP replaced FID in March 2024), CLS under 0.1.
• One clean H1 per page with hierarchical H2/H3 — headings are the retrieval hooks engines chunk on.
• llms.txt published per the llmstxt.org spec as a prioritized site map for assistants.

Layer 4 — Extractability

• Every key section answers its heading in the first 40-80 words so it survives being quoted alone.
• H2s are question-shaped ("How does X work?") to mirror real user queries.
• Comparisons use real GFM tables — engines lift tables into answers wholesale.
• At least three concrete facts per page (numbers, dates, named sources); the GEO paper by Aggarwal et al. (2024) measured a 30-40% visibility lift from adding statistics and citations.

Layer 5 — Authority

• Named author with a bio page and sameAs links — engines demote anonymous content in trust-sensitive spaces.
• Organization schema with a single @id referenced by every page's JSON-LD.
• Consistent entity naming across the site, social profiles, and third-party listings.
• Presence on high-authority external sources (Wikipedia, Wikidata, industry review sites) that engines treat as corroboration.

Layer 6 — Measurement

• A fixed prompt set of 20-40 real buyer questions is run on a schedule across ChatGPT, Perplexity, Claude, and Gemini.
• Share of voice and citation tracking are baselined before you change anything.
• AI referral traffic is captured in analytics via referrer rules and a dedicated channel group.
• Wins and losses are reviewed against model updates, so a ranking drop is diagnosed rather than guessed at.

Run the full sweep quarterly and after any major model release. A pass on all 60 points does not guarantee citations, but a failure on any single point in layers one through three guarantees you cannot earn them. Menra automates the measurement layer so the audit stays honest between reviews.