Why is a website rated A in GEO, online for years, actually invisible in AI answers?

The problem is not necessarily optimization itself. It lies elsewhere, and it is simpler than it looks: optimizing is not measuring.

An optimized website, and no presence at all

The experiment is built on an ordinary question, the kind a real buyer asks: choosing a specialized provider in a given region. Not a laboratory formulation, but an everyday commercial search.

The company studied has had a website online for several years. It is rated A by a tool specialized in GEO assessment.

LirenPrism then measured something other than the website's characteristics. Not what the site contains, but what the AI does when it answers that search. Forty queries were run:

• 20 answers produced without web search;
• 20 answers produced with web search.

Same question, same wording. Only the informational context changed. The result is unambiguous, and it is a fact established by the measurement: 0% presence in both configurations. The company is not cited in a single one of the measured answers.

This finding must be stated precisely. For the query and protocol tested, the excellent level of optimization reported by the GEO tool coexists with a total absence from the observed answers. That does not prove GEO is ineffective. It proves something different, and more unsettling: an optimization score is not enough to establish that a company is actually visible in AI answers.

Yet the AI had no shortage of players to cite

One might assume the query was too narrow, or that the AI had too little information to suggest players in this market. That is the reassuring explanation, and the measurements do not support it. Across the whole experiment, 57 different players were cited. Their presence in the answers is highly uneven:

• the two most present players: 50% of answers;
• then 45%, 42.5%, and 32.5%;
• the company studied: 0%.

So the AI does build a competitive landscape. It even builds it with a degree of stability, since certain names come back from one query to the next. The company studied is simply not part of it.

This is where the distinction between theoretical optimization and observed visibility stops being a matter of vocabulary. A tool can analyze a website and establish that it has the characteristics considered favorable to AI engines. It does not thereby establish that the site will actually be drawn upon when an AI has to answer a question.

Web access changed nothing

The experiment yields a second result, less expected than the first. Without web search: 0%. With web search: 0% as well. Allowing the AI to browse the internet produced no additional appearance of the company in this measurement.

This is an essential distinction, because it contradicts a widespread intuition. We readily assume that a company missing from a model's internal knowledge will naturally regain its place once the model can search the web. The experiment shows that this catch-up is not automatic.

The measurement, however, does not allow us to determine why. It does not allow us to claim that the AI encountered the site and set it aside, that it never encountered it, that it judged it less relevant than others, or that some other mechanism came into play. Those readings would be hypotheses, and it would be dishonest to present them otherwise.

The observable result nonetheless stands. Having a website accessible on the internet, and letting the AI search the internet, does not guarantee a place in its answer.

The AI appears to build its own exploratory field

This is probably the most interesting lesson of the experiment, and the most open-ended. Faced with a question, an AI does not return the web. It arrives at a subset of players, sources, pieces of information and arguments that form the substance of its answer. We can call that observable space its exploratory field.

In this experiment, that field brings dozens of players into view. Some recur regularly, others far less, and the company studied is entirely absent from it. Visibility in an AI answer therefore depends on more than the mere existence of a website or its apparent level of optimization.

What this measurement cannot yet determine are the mechanisms that define the boundaries of that field: why one player enters it, why another stays outside, what shifts those boundaries over time. At this stage, the question remains a working hypothesis. But this is probably where a large part of GEO's coming challenge lies — and it is striking that it is precisely the area optimization scores do not look at.

Does GEO measure an optimization or a result?

Two questions, often conflated, must then be separated. The first: is my website optimized according to criteria considered favorable to AI? The second: when users actually ask AI about my market, does my company appear?

These are not the same measurements, and they are not obtained with the same instruments. A technical or semantic audit answers the first. To answer the second, you have to observe AI engines directly: presence rate, citation frequency, rank, competitors actually proposed, arguments associated with the company, differences between models, change over time. These are result indicators, not compliance indicators.

There is nothing novel about this distinction. It already exists in many fields, where it is even considered elementary: compliance with best practices is not a measure of performance. GEO is probably no exception to that rule.

What this measurement proves, and what it does not

Rigor requires setting out the scope of the experiment. It does not allow generalization to every GEO tool, every company or every AI. It covers one precise wording, one provider and 40 queries. The report explicitly states that its results are relative to the query tested. It passes no judgment on the quality of the analysis tool that issued the A grade, and it does not say that grade is wrong within its own frame of reference.

It proves that one case exists, and that case is enough to shift the question. A website can achieve an excellent level of GEO optimization while obtaining 0% visibility when AI answers are measured directly.

A single counter-example does not refute a practice. But it does forbid treating a score as proof.

Optimizing is an action, visibility is a result

GEO is a young field. Methods evolve, engines evolve, and the way they search for or select information evolves too. It would be premature to conclude from this single experiment that a technique does or does not work. One methodological rule, however, emerges clearly.

A GEO score measures what a tool considers optimized. It does not measure the visibility actually obtained in AI answers.

Between the two lies an area we still understand imperfectly: the exploratory field specific to AI engines. As long as that area goes unmeasured, a company can improve its website, earn excellent grades and apply every available recommendation without knowing whether it has gained a place in the answers its future customers receive.

So GEO's real challenge may no longer be only to optimize for AI. It is to verify whether AI has actually changed its behavior after that optimization. What if the question to ask were no longer "is my website well rated?", but "what has changed, in the answers, since I optimized it?"