Why a dependency order rather than a checklist
Most AI SEO advice arrives as a flat list of tactics, which hides the thing that decides whether any of it works: the tactics depend on each other. A page a retrieval agent cannot reach is not improved by better writing. Better writing about a business a system cannot identify does not accumulate to anything. Money spent on layer six while layer one is broken is money spent on nothing.
INFERRED The ten layers below are this publication’s working model. It is not a proprietary framework and it is not a platform’s statement. Each layer is labelled with how well it is established. The order is the useful part: it says what to check first.
The ten layers
1. Crawl and fetch access
DOCUMENTED Can search and AI agents reach the pages, and does what they receive contain the content? Every platform documents its agents and Google says explicitly that crawling must be allowed “in robots.txt, and by any CDN or hosting infrastructure”. Usually cheap to fix; usually the first thing that is wrong. The crawler access guide covers the check.
2. Retrievable passages
INFERRED Does each page answer its question directly, near the top, in language a person would use? Answers are composed from passages, so the section is the unit that gets used. A page that answers in paragraph nine competes with one that answers in paragraph one.
3. Entity resolution
INFERRED Can a system state what the business is without hedging? One canonical name, one address form, one description, explicit relationships between the organisation, its people and its services. The layer most established businesses have never examined. Entity SEO is the reference.
4. Structured facts
DOCUMENTED Is the structured data accurate, validated and consistent with the visible page? Google’s instruction is that markup must match visible text, and that no special markup is needed for AI features. Markup that contradicts the page is worse than none. The structured data guide covers it.
5. Topic coverage
DOCUMENTED Does the site cover the subject, including the comparison, cost and suitability questions, or only the pages that sell? Google documents that its generative surfaces fan one question out into related searches, which is a documented reason coverage of sub-questions matters. A single page tells a system a business sells something; a connected set tells it the business knows the subject.
6. Original evidence
TESTED Does the site publish anything that does not already exist elsewhere: real methodology, first-hand observation, data? The GEO paper’s benchmark found that content changes altered generative visibility, with effects varying by domain; what it did not find is that restating what is already indexed earns anything. Information gain is the hardest layer to fake.
7. Corroboration
INFERRED Do independent sources agree with the business’s facts? Registries, licensing bodies, industry publications, reputable directories. It carries weight on this model precisely because the business does not control it. Knowledge panels and corroboration covers the sources.
8. Reputation
HYPOTHESIS What does the open web say about the business, and is it consistent with what the business says? Review platforms and community threads are indexed and independent. Whether and how they shape recommendations is a hypothesis this publication intends to measure, not a claim it makes.
9. Prompt coverage
INFERRED Which questions shape decisions in the category, and how are they answered today? People ask assistants situational questions that look nothing like keyword phrases. Mapping them, and noting which return an answer naming no business at all, turns AI search from a guess into a plan.
10. Measurement
DOCUMENTED What can honestly be observed? Google folds AI feature traffic into Search Console’s Web search type; GA4 groups assistant referrals into an AI Assistant channel; everything else is repeated-run sampling. The layer sits last and starts first, because without a baseline there is nothing to compare against. Measuring AI search visibility covers it.
Why the order matters
The most common expensive mistake is starting at layer five or six: deciding AI search matters, commissioning twenty articles, and seeing nothing, because a search agent had been disallowed since 2023, or the site rendered client-side, or the business name on the site did not match the registry. The articles were fine. Nothing beneath them was.
Each layer makes the one above it worth doing. Access makes content reachable. Entity resolution gives content something to attach to. Structured facts make the attachment machine-readable. Coverage turns pages into a subject. Original evidence makes the subject worth citing. Corroboration makes it credible. Cost runs the other way: layers one to four are days or weeks of bounded technical and editorial work; layers five to eight are open-ended. The cheap layers unblock the expensive ones, which is the practical argument for the sequence.
What the model does not do
None of the layers forces a system to name a business. They improve access, clarity and evidence, which is a weaker claim than the ones sold elsewhere and the accurate one. Outputs vary by session, region and model version; a source that appears in one answer may not appear in the next, and that variance is a property of the systems, not a fault in the work. A layer that says “do this and you will be cited” is not on this list because no platform documents such a thing.
What is documented and what is inferred
- Documented: access controls; markup matching visible text; query fan-out; measurement surfaces.
- Tested: content changes altering generative visibility in the GEO benchmark.
- Inferred: the passage as unit; entity resolution; corroboration weighting; prompt coverage.
- Hypothesis: reputation’s effect on recommendations.
- Not established: any weighting between layers, and any Perth-specific frequency.