SEO as cognitive architecture: the full model

SEO is a discipline of information retrieval and of behavior, and the five-layer model describes the path between a piece of content existing and a person acting on it.

Where the model’s starting observation comes from

When the interface returns an answer instead of a list, click capture shrinks and the work shifts from competing for position to competing for recognition. The observation has consequences for the object of the discipline, and the tactical consequences come after that.

While the object was a document’s position in a list, a working model based on crawling, indexing and relevance was enough. Once the object includes the decision that happens after an answer, the model has to reach behavior, and behavior has a literature of its own.

The five-layer model, layer by layer

The table below summarizes the five layers with the question each one answers and its own failure mode.

The failure column is what gives the table practical use. It lets you enter wherever the problem shows up, instead of walking through the model from the start, which is how it tends to be used once it becomes a diagnosis.

What is cumulative between the layers and what is not

The relationship between the layers has an asymmetry that needs to be stated precisely, because it is the source of most misuses of the model.

• A failure in the first layer cancels the rest. Inaccessible content never gets structured, interpreted, remembered or chosen.
• Solving the first three does not guarantee the fourth or the fifth. Flawless retrievability can coexist with a complete absence from the decision.

What is my own contribution and what is literature

The line between what I formulated and what I applied needs to be explicit in every public presentation of the model, including the stage, the classroom and the sales proposal.

• Established literature: the recognition heuristic, formulated by Goldstein and Gigerenzer in the field of cognitive psychology.
• My own contribution: the five-layer architecture and the application of the recognition heuristic as the action layer.

Concepts associated with the model

Three concepts travel with the model, and each has a distinct origin that needs to come along whenever the concept is cited.

The origin column exists so the borrowing stays visible. Two of the three concepts are my own formulation and one comes from established literature, and citing all three without that distinction would mean claiming what belongs to someone else.

How the five-layer model should be used

• Use it as a reading key to locate which layer a project’s problem sits in.
• Use it as a diagnostic structure, with the caveat that the fourth layer depends on third parties.
• Use it to organize a curriculum, because the order of dependency helps teaching.
• Never present a layer as an execution phase with a deadline and a deliverable, because that promises control over something that cannot be controlled from inside your own domain.
• Never promise layer 5 as a contracted result, for the same reason.

Every sales deck that turned the five layers into five deliverables sold the fourth one, which depends on third parties. Good luck putting other people’s memory on a Gantt chart.

How to apply the model in a diagnosis

The practical way to use the model in an audit is the recognizable brand x-ray, published alongside my SEOcamp talk in September 2026. It translates the layers into eighty checks grouped into six symptoms, each tied to the layer where the failure happens. Inconsistent entity, absence across platforms and non-recurring presence sit in the memory layer. Data that agents cannot read sits in the structure layer, vector distance from the audience in the interpretation layer and unverifiable trust in the action layer.

The first level of each symptom runs in the browser, with no paid tool, which lets you start the diagnosis in the first meeting. The reading follows the model’s order of dependency: a symptom in the structure layer is solved before any conclusion about memory, because the earlier failure contaminates the reading of the next one.

To track progress after the diagnosis, the AI search KPIs follow the same five-layer split, with eighteen operationally defined indicators. The execution sequence for anyone who wants to be cited, from crawlability to external corroboration, is in the guide on becoming a preferred reference source for AI.

Where GEO and AEO fit in the model

GEO, short for Generative Engine Optimization, and AEO, for Answer Engine Optimization, circulate as names for a new discipline. In the model, the practices gathered under those acronyms are distributed across the layers that already exist.

Allowing an AI system’s crawler is layer 1 work. Writing a block that answers on its own, with the entity named and the provenance inside the excerpt, is layer 2 and 3 work. Showing up in the third-party sources that feed the answer is layer 4. What the acronyms usually leave out is layer 5, where the question becomes whether the cited brand is the recommended one.

That is why I treat GEO and AEO as market names for parts of the same work, useful in a sales conversation and insufficient as a method. Every new acronym comes with its own conference, and the htaccess stays exactly where it was.

What remains open in the model

The methodological frontier is recorded on purpose. It still needs to be specified whether recognition operates in the model’s parametric layer, built into the weights, or in the retrieval layer, depending on a document found at query time.

An invitation to critique the model

A model that cannot be refuted is of no use. The conditions below would bring down the current formulation, and publishing them is part of presenting the model as a scientific contribution, open to refutation.

• Evidence that brand recognition does not influence the composition of generated answers, in a design that controls for the presence of a retrievable document.
• Evidence that solving the first three layers, on its own, produces stable participation in decisions.
• A demonstration that the separation between memory and action has no distinct practical consequence, which would reduce the model to four layers.