The model in one view
Ask ten marketers whether a brand is “visible in AI search” and you get ten confident, incomparable answers: one pasted a prompt into ChatGPT, one read a screenshot thread, one is quoting a tool that prints a proprietary score with no method behind it. None of that survives a paying client’s three reasonable questions: what is happening, how do you know, and what should we do.
The four-layer model exists to make those questions answerable. Reading the diagram top to bottom: technical eligibility feeds evidence, evidence feeds representation, representation feeds outcomes, and each connection is a gate that the layer above must pass before the layer below means anything.
Layer one: technical eligibility
Before any answer engine can use your content, its systems have to reach and read the pages that matter: robots directives, response codes, rendering, index controls, canonical signals, facts present in rendered text. Boringly technical, completely decisive.
The failure mode is silence. A disallowed crawler produces no error a marketer ever sees, just absence, and absence gets misdiagnosed as “the algorithm does not like us” when the real cause is a header nobody has inspected since the last migration. This is also the only layer where the platforms publish hard rules: OpenAI’s crawler documentation states that sites disallowing OAI-SearchBot will not be shown in ChatGPT search answers, and Google’s AI features documentation says a page must be indexed and snippet-eligible, with no additional AI-specific requirements. Which crawler governs what is covered in the OAI-SearchBot versus GPTBot guide.
Eligibility questions have observable, checkable answers. That is exactly why they come first.
Layer two: evidence
Reachable is not the same as used. The second layer asks whether answers to relevant buyer questions actually draw on your sources: linked, cited, or paraphrased. Your pages are candidates here, but so is everything else the engine can retrieve about you: directories, reviews, old press, a competitor’s comparison table.
This is where public GEO conversation collapses, because people sample one prompt, once, on one surface, and generalize. Answer engines are probabilistic and personalized; a single observation is an anecdote. The professional version is a controlled sample: a frozen prompt panel derived from buyer decisions, repeated runs, every observation preserved, exactly as specified in the measurement protocol. Small, boring, repeatable. That is what turns “I saw” into “we observed.”
A layer-two failure means the retrieval pool around your buyer’s questions is owned by other people’s documents. The fix is evidence work, not markup tricks.
Layer three: representation
When the brand does appear, is it described accurately? Wrong pricing, dead product names, a competitor’s feature attributed to you, a three-year-old positioning statement recited as current fact. Representation errors are more commercially dangerous than invisibility, because a buyer who reads a confident wrong answer does not know it is wrong, and neither do you until you check.
Representation is classifiable, which is what makes it reportable: every material statement in a sampled answer is compared against an approved fact ledger and labeled accurate, incomplete, outdated, conflicting, or unsupported. Classification against approved facts is what makes the finding client-safe rather than an opinion about tone.
Layer four: outcomes
Finally the commercial layer: does any of this move enquiries, demos, trials, sales? It is the layer everyone wants to discuss first and the layer with the least direct evidence. Referral clicks are partially observable: OpenAI’s publisher FAQ documents the utm_source=chatgpt.com parameter on ChatGPT links, and Google keeps AI Overviews and AI Mode activity inside Search Console’s Web search type. Assisted influence, the answer that shaped a decision without a click, is mostly not observable today.
The honest position: report outcomes as directional, tie them to the evidence in layers one through three, and refuse to invent a number where none exists. Clients do not leave because you told them measurement has limits. They leave because you promised what you could not show.
The gating is the point
The layers form a dependency chain, and the chain is what makes this an audit instead of a debate. There is no evidence without eligibility. Representation problems are only findable once real answers exist to inspect. Outcome claims without the first three layers are astrology with a dashboard.
Two practical consequences follow. First, diagnosis always starts at the earliest unresolved layer, because work on a later layer is unfalsifiable while an earlier one fails: content investment cannot be evaluated while a crawler block is live, and no amount of representation polish matters on pages no engine reads. Second, the layers must never be averaged. A composite score that blends an eligibility pass with a representation failure manufactures a middling number out of two findings that demand opposite actions.
Running a layer-by-layer diagnosis
Work the table top to bottom and stop at the first failing row. The finding is the layer, the evidence, and the smallest action that would clear the gate.
| Order | Layer | Core check | Layer passes when |
|---|---|---|---|
| 1 | Technical eligibility | Response codes, robots rules, index controls, rendered text on priority pages | Every priority page is reachable, indexable, and shows its material facts in rendered text, with dated evidence |
| 2 | Evidence | Frozen prompt panel, repeated runs, citation and source logging | Sampled answers draw on owned or accurate third-party sources at a rate you can state with its denominator |
| 3 | Representation | Material statements compared against the approved fact ledger | No material conflicting or outdated claims remain unaddressed in the sampled answers |
| 4 | Outcomes | Attributable referrals and qualified actions in analytics | Observable outcome data is reported with its limits, separate from answer observations |
The sentence this produces, “your problem is at layer two, here is the evidence, here is the fix,” is worth more to a client than any visibility score on the market. The 47-point audit checklist turns this diagnosis into a scored protocol, and the report template gives each layer its own section so the findings stay separated all the way to the client.
Grade your own protocol layer by layerThe checker scores the same gates as separate sections, so a complete technical review cannot cover for a panel you never froze. It grades your method, not the brand, and it never estimates visibility.
Open the free checkerThis framework is the thinking layer of the AI Search Visibility Audit System, which turns it into a full working service: the scoped prompt protocol, the observation method, the representation baseline, and the client report, proposal, and 90-day roadmap that package it.