What can AI-visibility measurement actually prove?

// 01 · the problem

Why does a page like this exist?

No engine publishes positions. Anyone showing you one made it up.

The GEO market has a credibility problem: per-query answers from ChatGPT, Perplexity and Google AI are non-deterministic, personalized and cached — the same prompt asked by two people, on two different days, can return two different citation sets, and neither one is "wrong."

There is no search console for ChatGPT or Perplexity — no dashboard a vendor can screenshot to show you a position. Google is the partial exception: Search Console reports aggregate AI-surface impressions for your own property, but nothing for a competitor and no per-prompt citation view — not the thing a vendor would need to show you a "ranking."

A vendor who promises you positions in AI answers is describing something they cannot actually observe.

// 02 · what is measurable

What can AI-visibility measurement actually show you?

Named engine, named prompt, named day — or it is not a reading.

Three things, all directly observable: whether a specific engine — ChatGPT, Perplexity, Google AI Overviews and AI Mode, Copilot or Gemini — cites you for a specific prompt, on a specific day, from a specific locale; the delta between two such readings taken weeks apart, which is a trend, not a single snapshot; and the technical layer underneath — whether AI crawlers can reach and render your pages at all, and whether your name, description and facts agree with each other across the open web.

Two more things this market sells as levers are recorded here as facts rather than promises. Structured data is checked because it is ordinary SEO hygiene, not because it buys citations: Google's own AI-optimization guide states that structured data is not required for generative AI search and that there is no special schema.org markup to add for it. An llms.txt file is checked the same way. The same Google guide says machine-readable AI files neither help nor harm visibility in Google Search because Search ignores them, and server-log studies published through 2026 report that AI crawlers almost never request the file at all — so the audit records whether you have one and whether it is valid, and never counts it as a reason you were cited.

Nothing on this list requires access that an outside party does not have, which is exactly why it can be measured honestly.

// 03 · the €900 audit

How does the €900 audit measure it?

20 prompts, 5 engines, one dated run — the scope, on paper.

A fixed set of 20 buyer-intent prompts, run once across five engines — ChatGPT, Perplexity, Google AI Overviews and AI Mode, Copilot, Gemini — with every citation recorded verbatim. Results are compared against two or three named competitors you choose.

A written report and a written Q&A round land within 5–10 business days, with every question answered within one business day. That single run becomes a dated baseline, not a verdict — the next section explains why the distinction matters.

// 04 · one run against two

Why does a single run prove nothing on its own?

One reading is a baseline. Two readings are evidence.

Because AI answers vary by the hour and by the visitor, one run of one prompt set tells you what happened once, not what happens typically. The audit reports it as exactly that: a dated baseline plus a fix list, never a score to defend.

The number that actually carries signal is the delta between two runs of the same prompt set, taken at least a month apart, against the same named competitors — that is the entire point of the monthly retainer's tracking: the same fixed method run twice, so a change in the number reflects a change in the world, not a change in the method.

// 05 · the limits

What cannot be measured — even by me?

What cannot be observed from outside does not enter the report.

A global "ranking" in ChatGPT or any other engine does not exist to be measured — the phrase describes something these systems simply do not have.

Causal attribution of a single change is not provable from one delta either: your content, the engine's underlying model and your competitors' content all move between two readings, so a shift after a fix is evidence that it worked, not proof.

And anything that requires access these engines do not expose — why a model chose one citation over another, its internal retrieval logic, real click-through on a citation — stays out of the report, because it cannot be honestly observed from outside. If a deliverable cannot be verified by you, independently, it does not go in.

// 06 · GEO and SEO

Does GEO replace SEO?

Most of AI visibility is technical SEO done properly.

A large share of AI visibility is downstream of ordinary technical SEO done well — clean crawlability, correct structured data, a coherent entity, content that actually answers the question being asked.

Anyone selling GEO as a replacement for SEO is selling a rebrand, not a new discipline. The audit says plainly which of your gaps are classic SEO work and which are genuinely AI-specific, because blurring the two is how an invoice gets padded.

// 07 · your checklist

How should you use this page before you buy?

Five questions. Ask them before you sign anything — mine included.

Take these questions to any vendor before you sign anything: which engines do they sample, how many prompts, how often, against which named competitors, and what exactly do they refuse to promise.

A vague answer to a measurement question is the tell — hold on to your budget.

1. Which engines do they sample?
2. How many prompts, and how were they chosen?
3. How often do they re-run the same set?
4. Against which named competitors?
5. What do they refuse to promise?