Method
Diagnose bottom up, or you are guessing.
Audits run in four layers, in order. Each layer constrains what the layer above it can achieve, so a finding produced out of sequence cannot be justified.
The sequence
AEO
Retrieval tested repeatedly against real buyer queries. Reported as a distribution, never as a rank.
Link profile
Generated answers reward agreement across sources, not authority on one domain.
Content
Retrieval works on passages. Whether an answer survives being pulled out of context.
Technical
What a machine actually receives. Cheapest layer to fix, and it blocks the most.
Layer 01 · Technical
Whether a machine can fetch the page, receive its content without executing JavaScript, and get a coherent response back. Crawl behaviour, rendering parity, status codes by agent, and where crawl budget is actually spent.
This layer produces the findings that cost the least to fix and block the most. A client-side rendered application returning an empty document to a plain fetch is not an AEO problem, and no amount of content work will move it.
Layer 02 · Content
What is on the page and whether it can be extracted. Retrieval works on passages, so the question is not whether a page covers a topic but whether a specific answer sits in a chunk that survives being pulled out of context.
Also assessed here: internal contradiction across the site, which on large catalogues is a larger problem than per-page quality, and whether the entity behind the brand is stated consistently or paraphrased differently on every page.
Layer 03 · Link profile
Corroboration. Generated answers reward agreement across sources rather than authority on a single domain, which puts weight on surfaces you do not control. This layer maps who else describes you, how, and whether that description matches the one you would choose.
Layer 04 · AEO
Built on the three below it. Retrieval testing against real buyer queries, sampled repeatedly rather than once, because a single result in a probabilistic system is a draw from a distribution and not a measurement.
The output is a frequency distribution and a set of structural findings, not a position number. Anyone reporting AEO as a rank is reporting noise. The reasoning is in the difference between SEO and AEO.
What this practice will not claim
- That revenue can be attributed to a model citation. It currently cannot, and the problem is epistemic rather than technical.
- That a number of citations or a position in AI Overviews can be guaranteed. Neither is in anyone’s control.
- That a single answer-engine test result means anything on its own.
Saying this up front costs some engagements. It is also the reason the findings that are claimed can be checked.
Next step
Send the problem. Get a scoped proposal.
Within 24 hours you get a proposal with a fixed scope and a delivery window, or a note saying this is not the right practice for it. Quoted per engagement in euros. No subscription, no default retainer.