TL;DR: Generative Engine Optimization (GEO) is the discipline of making your site discoverable, retrievable, and citable by AI search — Google AI Overviews, AI Mode, ChatGPT Search, Perplexity, Claude. These 37 checks run in dependency order: crawler access first, measurement last. They are not ranking factors, and completing all 37 guarantees nothing — but failing the early ones guarantees invisibility. Accessible ≠ eligible ≠ retrieved ≠ cited.
I am Akira. This is the framework we run before any GEO engagement — ordered by dependency, because fixing Layer 3 content while Layer 1 access is broken is polishing a car with no engine. If you want the machine-readable version run against your own domain, that's the GEO Scorecard (free) or the full GEO Audit (48-hour baseline).
Layer 1 — Access: 7 Checks
1. Googlebot can access the target page. No robots.txt block, login wall, or firewall rule. Verify with Search Console's URL Inspection.
2. OAI-SearchBot, GPTBot, and ChatGPT-User are separate decisions. OAI-SearchBot surfaces you in ChatGPT Search. GPTBot is training crawling — a policy choice, not a visibility requirement. ChatGPT-User is user-triggered fetching. Allowing GPTBot is not required to appear in ChatGPT Search.
3. PerplexityBot ≠ Perplexity-User. PerplexityBot builds the index that surfaces you. Perplexity-User fetches on user request and generally ignores robots.txt. Check both, separately.
4. Anthropic's bots are three gates, not one. ClaudeBot (training), Claude-User (user fetches), Claude-SearchBot (search retrieval). Align each with your policy.
5. Your WAF/CDN isn't blocking what robots.txt allows. An Allow rule means nothing if the CDN returns 403/429 to the crawler — the most common technical SEO blind spot we find in GEO audits. Audit server logs and actual HTTP responses, not just config files.
6. Important content renders. Google processes JavaScript; don't assume every AI crawler does. Test representative pages and check actual retrieved output where possible. Never serve crawlers materially different core content.
7. Sitemap current — but not worshipped. A sitemap aids discovery; it guarantees neither indexing nor citation. And no — Google does not require llms.txt, AI-specific Markdown mirrors, or any special machine file for AI Overviews eligibility. Optional extras, not the foundation.
Layer 1.5 — Search Eligibility: 2 Checks
8. The page is indexed by Google. No index, no AI Overview supporting link. This is technical SEO before it is GEO: fix indexing before touching a single paragraph. Content refinement cannot compensate for absence from the index.
9. The page is snippet-eligible. Review nosnippet, data-nosnippet, max-snippet, noindex — and the Search Console "Search generative AI" control (globally available; inclusion is default, but child properties can inherit exclusion). Ignore the myth that a page "must rank top 10 organically" to appear in AI Overviews. Google has published no such threshold.
Remember the chain: accessible ≠ eligible ≠ retrieved ≠ cited. Each gate is separate; each failure has a different fix.
Layer 2 — Entity Clarity: 5 Checks
10. Organization names are consistent. Official name, short name, legal name — one convention across key pages.
11. Structured-data entity identifiers are stable. Same @id for the same entity, everywhere.
12. sameAs links point only to pages that genuinely identify the entity. Not every social profile qualifies.
13. Author information is real and consistent. Named humans, stable bios. AI engines are entity-resolution machines; anonymous "Admin" bylines are dead weight.
14. Business and contact details match across pages. NAP consistency is boring until it's the reason an AI can't confirm you're one company.
None of these are published ranking factors. They're ambiguity removal — and ambiguity is the silent killer of citation.
Layer 3 — Content Clarity: 8 Checks
15. Answer important questions directly. First sentence of the section answers the heading's question. This is the single highest-leverage habit in GEO — and it's the discipline AI search inherited directly from good SEO.
16. Headings describe their sections. If a heading could sit on any page of any site, it's decoration, not structure.
17. Important statements have a clear subject. "It improves performance" — what is "it"? Pronoun soup is unretrievable.
18. Conditions sit next to results. "Cuts costs 40%" without "for teams under 20" is a claim AI engines learn to distrust.
19. Evidence sits next to claims. Source, date, method — adjacent, not three scrolls away.
20. Concept relationships are stated. Show how numbers and ideas connect; retrieval lifts fragments, and fragments must carry their own context.
21. Tables and lists have clear labels. Headers that mean something out of context.
22. Don't shred content into micro-sections "for AI." Google explicitly doesn't require artificially fragmented content. Write for comprehension, not for a rumored parser.
Layer 4 — Evidence: 4 Checks
23. Original evidence anchors important topic clusters. First-party data, small studies, real benchmarks. Commodity content gets summarized; evidence gets cited.
24. Important numbers carry scope, date, and method. "74.5% trigger rate" means nothing without the sample, window, denominator, and exclusion rules.
25. Platform facts come from current official sources. Crawler policies shift. Cite the documentation as of a date, and revisit it.
26. Uncertainty is stated. If language and topic are confounded in your sample, say so. AI engines — and sophisticated buyers — reward calibrated claims over confident ones. We call the accumulated cost of unverified claims about your brand Truth Debt. It compounds.
Layer 5 — Structured Data Consistency: 4 Checks
27. Structured data matches visible content. Markup that contradicts the page is a liability, not a boost.
28. Only schema types that fit the page. FAQ markup on a page with no FAQ is spam-shaped.
29. Author and dateModified reflect reality. Fake freshness is detectable and corrosive — Perplexity's recency weighting makes this expensive.
30. Schema validation ≠ citation. Passing the validator means the syntax parses. Nothing more.
Layer 6 — Link Architecture: 3 Checks
31. Cluster content links back to its pillar. 32. Pillars link out to important clusters. 33. No near-duplicate intent pages competing with each other — keyword cannibalization is the oldest SEO own-goal, and AI retrieval makes it more expensive, not less. Internal-link architecture is classic SEO doing GEO work: it's how you teach both Google and LLM retrievers which page owns which question — the principle behind how we structure the GEO service mesh.
Layer 7 — Bilingual Layer: 2 Checks
34. Canonicals point to the same-language page. 35. hreflang is reciprocal. For Hong Kong sites running English and Traditional Chinese, this is where bilingual implementations quietly break — English and Chinese pages are not translations of one intent; they're two research jobs. (This is core to our Hong Kong SEO and GEO work.)
Layer 8 — Measurement: 2 Checks
36. The prompt set and protocol are versioned. Fixed prompts, platforms, markets, languages, run counts — changed deliberately, with version notes.
37. A baseline and normal-variation range exist before optimization. Without them, "improvement" is unfalsifiable storytelling. This is the check most agencies skip, because it makes them accountable.
How to Run the 37 Checks
In three passes, run like an SEO audit with a GEO lens: access and eligibility first (checks 1–9 — hard gates), content and evidence second (15–26 — where citations are actually won), measurement always on (36–37 — the loop that keeps you honest). Entity, structured data, links, and bilingual layers slot in by dependency.
FAQ
Where should Generative Engine Optimization start? With the baseline — check 37, paradoxically first. You can't prove improvement without a starting measurement.
Will blocking Google-Extended remove us from AI Overviews? No. Google-Extended governs training and grounding uses. AI Overviews runs on Googlebot under the nosnippet family of controls. Different switch, different scope.
If all 37 checks pass and there's still no citation, what's missing? Usually conventional visibility — the page isn't in the retrieval pool for that query at all. That's an SEO problem, and no passage polish fixes it.
Which checks can be skipped if time is short? None of Layer 1. Ever. If forced to triage: 1–9, 15, 23–24, 36–37.
Stop guessing what AI sees. Start measuring it. Good SEO made you findable; GEO makes you citable — in 2026 you need both. Run the free GEO Scorecard for a first read, or book the GEO Audit and get the full 37-check pass with a per-engine baseline. If the outcome you're buying is citations, the destination page is Get Cited by AI.
Mercury Technology Solutions: Accelerate Digitality.
