The SEO-AEO-GEO
audit checklist

A. Crawler access and technical foundation

If a model cannot fetch the page, nothing else on this list matters. This phase is also where the most expensive single mistake lives: a setting aimed at AI scrapers that quietly blocks Google.

1. robots.txt present, reachable and returning 200A missing or erroring robots.txt is read as unrestricted by some crawlers and as blocked by others. The ambiguity is the problem.
2. AI crawlers explicitly named and permittedMost sites block AI crawlers by accident rather than by decision. Explicit permission removes the ambiguity for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, anthropic-ai, PerplexityBot, Google-Extended and CCBot.
3. Multi-purpose crawler trap checked at the edgeCloudflare judges multi-purpose crawlers by the strictest rule that applies. Blocking Training also blocks Googlebot, Bingbot and Applebot. A tick-box aimed at AI can cost you traditional search entirely.
4. Edge and WAF bot rules reviewedBot management, rate limits and challenge pages often block agents before robots.txt is ever read. The site owner rarely knows this is happening.
5. Platform default changes checked against current policyDefaults move without notice. Cloudflare began blocking Agent and Training bots on ad-bearing pages by default in September 2026 for new and free zones.
6. sitemap.xml valid, complete and submittedDiscovery should not depend on a crawler finding its way. Submitted to both Google and Bing, because ChatGPT search grounds substantially on the Bing index.
7. Key content renders without JavaScriptMany agents fetch without executing JavaScript. Content that only appears after hydration is invisible to them regardless of how good it is.
8. Response codes, redirect chains and soft 404sEvery hop costs crawl budget and some agents abandon after one. Soft 404s poison the index with pages that look real and say nothing.
9. Canonical tags consistent and non-conflictingConflicting canonical, hreflang and sitemap signals make a model uncertain which URL is authoritative, and uncertainty suppresses citation.
10. Time to first byte and payload weight for crawl efficiencyAgents operate under time and token budgets. A slow or heavy page gets partially read or dropped.

B. Entity clarity and structured data

A model has to resolve who you are, what you sell and who you serve without contradiction. Inconsistent naming across sources splits one business into two in a model's view.

1. Organization schema complete on the home pageThis is the anchor record. Legal name, URL, logo, contact, slogan, knowsAbout and makesOffer covering every offering.
2. Person schema for the founder or named expertsModels weight identifiable expertise when choosing whom to cite. An entity with no person behind it is harder to trust and easier to skip.
3. Service or Product schema on every offering pageOne page per offering, each declaring what it is, who provides it and where it is available.
4. BreadcrumbList on interior pagesTells a model where a page sits in the structure rather than making it infer hierarchy from URLs.
5. FAQPage markup matching visible copy exactlyThe most reliably parsed structured data available. A mismatch between markup and visible text is a spam signal.
6. No self-serving Review or aggregateRating markupReviews about your own business, collected on your own site, are disallowed for rich results and risk a manual action.
7. Naming invariance across every propertyOne character of inconsistency splits an entity. The legal name, product names and capitalisation must be identical everywhere a model can find them.
8. sameAs coverage audited and extendedThis is how a model corroborates that two mentions are the same entity. Sparse sameAs means weaker resolution.
9. All structured data validates with zero errorsA malformed block can invalidate the rest of the graph. Errors here are silent.

C. Content structure for extraction

Extraction pulls passages, not pages. Sentences that depend on the paragraph above them cannot be quoted, so they do not get cited.

1. One h1 per page, heading order never skipping a levelThe heading tree is the outline an agent walks. A broken tree means a misread structure.
2. Title, meta description, canonical, Open Graph and Twitter card on every pageThese are the fields most often used verbatim when a model summarises or a platform previews you.
3. One page competes for one primary intentA page covering five topics gets cited for none. Five offerings means five URLs, not one page with five sections.
4. Key claims written as self-contained passagesExtraction pulls passages. A sentence that depends on the one before it cannot be quoted, so it does not get cited.
5. FAQ answers 40 to 60 words and standaloneLong enough to be complete, short enough to be lifted whole.
6. No FAQ text duplicated across two URLsDuplicate FAQ markup risks one copy being discarded, and you do not control which.
7. Pricing published, or a defensible range and structureAn agent asked what you cost will answer from something. If you publish nothing it infers a number, and the number will be wrong.
8. Comparison and alternatives content existsComparison is one of the most common buyer prompts. With nothing to cite, a model uses a competitor's comparison page — written to make you lose.
9. Audience and use-case content resolves who this is forHalf of buyer prompts are qualification questions. If a model cannot tell whether you serve a 15-person firm, it will not recommend you to one.
10. Decisive facts moved out of PDFs and images into HTMLA price list in a PDF from four years ago is the single most common cause of an agent quoting a wrong number.
11. Tables marked up as tablesA comparison table rendered as an image is invisible. As HTML it is one of the most extractable structures on the web.
12. llms.txt present, accurate and currentA plain-text summary of the offering with links to each page. Cheap to maintain, and increasingly requested.
13. Content freshness and visible datingModels discount undated content and heavily discount content they can tell is stale.

D. Prompt testing and baseline

The only direct measurement available. We ask the engines what they say about you today and record it, so there is something to compare against later.

1. Buyer-intent prompt basket definedThe basket is the instrument. Twenty to fifty prompts phrased the way a buyer actually types them, not the way the business describes itself.
2. Brand-recognition prompts runDoes the model know you exist, and does it describe you correctly when named directly.
3. Category prompts run — unprompted mentionThe one that matters. Asked who to hire for the problem you solve, without your name in the prompt, are you named at all.
4. Comparison prompts run against the named competitive setHow the model frames you next to specific competitors, and whether the framing is accurate.
5. Objection and risk prompts runWhat a model says when asked what is wrong with you. Buyers ask this and almost nobody tests it.
6. Pricing and terms prompts runWhere wrong answers do the most commercial damage, and where the source of the error is usually findable.
7. Coverage across the major enginesEngines behave differently enough that optimising against one gives a misleading picture. ChatGPT, Claude, Perplexity, Gemini, Copilot and Grok.
8. Baseline scored and storedWithout a dated baseline there is no way to demonstrate movement later, and no way to tell drift from effect.
9. Factual errors traced to their sourceAn error is only fixable once you know where the model got it. Usually an old page, a directory listing or a third-party article.

E. Competitive citation analysis

Roughly 85% of AI citations come from third-party sources. Knowing where competitors are cited from tells you where the gap actually is — usually off your own site.

1. Competitive set identified from the prompts, not from assumptionThe competitors a model names are frequently not the ones the client names. That gap is itself a finding.
2. Citation sources mapped for each competitorWhether they are cited from their own site or from third parties determines what you would actually have to do to compete.
3. Third-party source gap identifiedRoughly 85% of AI citations come from off-site sources. This is usually where the real work is, and it is the part clients least expect.
4. Directory, listing and profile gapsCheap to close, disproportionately effective, and frequently the reason a competitor is named and you are not.
5. Review and community platform presenceModels lean on consensus signals. Absence from the platforms your category is discussed on is a structural disadvantage.

F. Findings, priorities and handoff

A list of 200 issues sorted by a tool is not a deliverable. Priority order, effort against impact, and a document you own.

1. Findings separated into applies, does not apply, and already correctA scanner cannot tell the difference between missing and not applicable. Separating those is most of what you are paying for.
2. Remediation ranked by impact against effortA list of 200 issues sorted by a tool is not a plan. Ranking is where an audit either becomes actionable or becomes shelfware.
3. Quick wins separated from structural workSome of this ships in an afternoon and some needs a quarter. Conflating them stalls the whole plan.
4. Measurement infrastructure confirmed liveNothing can be demonstrated later without Search Console, Bing Webmaster and analytics running now.
5. Re-measure scheduledModel behaviour drifts. A position earned in one quarter is not a position held in the next.
6. Source-of-truth document delivered to the clientYou own the findings and the plan regardless of who implements them. That is deliberate.

Take it with you

Two designed PDFs. The checklist, with room to score your own site against it. And a sample deliverable — a four-page findings report for a fictional B2B software company, so you can see exactly what an audit produces before you buy one.

Tell us which you want and we will send them. We will also send occasional notes on how answer engine behaviour is changing. Nothing else, and you can leave at any time.

Rather not share your email? The full checklist is on this page, in full, and always will be. The PDFs are a convenience, not the content.

What you actually receive

The checklist is what we examine. This is what lands at the end of it — a four-page findings report for a fictional B2B software company, showing the baseline, every check classified as applies, already correct or not applicable, and a remediation plan split into quick wins and structural work.

It is deliberately short. A sixty-page PDF is a way of hiding that only nine things matter.

Questions