GEO — "Make your site accessible and optimized for AI so they come to find your information." Do you really believe it?

The Observatory compared several measurements of answers produced with web search enabled. The results reveal a phenomenon far less mechanical than one might assume.

Accessible does not mean consulted

In one of the measurements analyzed, the phenomenon works almost as one would hope. The official website of the company studied is cited 100 times out of 1,153 source citations, about 8.7% of the total. It is even the third most used source by AI in that measurement.

So AI is perfectly capable of going directly to a company's website to find information. But the other measurements tell a very different story.

• on one query: 36 domains used, 172 citations — the studied company's domain does not appear among the sources;
• on another measurement: 16 domains, 227 citations — same observation;
• on yet another: 24 domains, 105 citations, with no consultation of the studied company's domain.

So AI does search the web. It simply does not necessarily search on your site.

2,336 citations: an average that hides the real phenomenon

Across five reports where source citations and the studied company's domain can be matched clearly enough, we count 2,336 source citations. The company's website is cited 100 times, that is 4.28% of citations. But this average percentage is misleading.

The distribution observed across the five measurements is: 0 — 100 — 0 — 0 — 0.

So this is not simply a matter of low consultation frequency. It is a matter of selection. In some situations, the company's website enters the field explored by AI and becomes a major source. In others, hundreds of citations are produced without its domain being used at all.

And yet AI does consult company websites

Another observation prevents any simple conclusion that AI would favor media outlets, Wikipedia, Reddit or editorial sites over official websites. In a measurement where the company studied is absent from the answers, several competing players are directly associated with their own domains among the main sources.

Three domains belonging to featured players each collect 18 citations, and a fourth collects 17. Those four domains alone account for 71 citations out of 172, that is 41.3% of the citations in that measurement.

So AI does not reject commercial websites. On the contrary, it can use them massively. The question then becomes far more interesting: why does it consult some company websites and ignore others?

GEO may come after a step we still rarely measure

A large part of today's GEO recommendations focuses on the content itself: page structure, clarity of information, structured data, explicit answers, citations, authority or ease of use by models. These practices may have their merit. But our measurements highlight a prior question: what is the point of optimizing information for an AI if that AI never selects the page containing it?

We observe here an important distinction between three phenomena that are easily confused:

• being accessible;
• being consulted;
• being restituted in the answer.

These are not necessarily the same thing. A website can exist on the web and be technically accessible without ever appearing among the sources an AI actually uses. Conversely, a company can become a major source once it enters that field of consultation.

A particularly revealing example

A local measurement makes the effect of the web directly observable. Without web search, the company studied is not cited by any of the three providers measured. With web search, its presence rate rises to 100% with Anthropic and 60% with Google, while remaining at zero with OpenAI. In that same measurement, its official website receives 100 citations and becomes the third most cited source.

It would be tempting to conclude that consulting the website directly causes the company to appear. The data does not allow that causality to be demonstrated. It does allow two phenomena to be observed simultaneously: the massive entry of the website among the sources, and a sharp increase in the company's presence once web search is enabled. That deserves to be measured.

The real issue may be the "exploratory field"

These results suggest a useful distinction. There is the available web: every page an AI could theoretically reach. And there is the web actually explored: the domains and pages the system really selects when building its answer. We can call that space the AI's exploratory field.

Our measurements show it to be highly selective. Two companies present on the same market do not appear to benefit automatically from the same exposure to that exploration. One can become a repeated source; the other may never be consulted at all.

This shifts part of the GEO question. It is no longer only "is my content optimized to be understood by an AI?". It also becomes: "does my site even enter the AI's exploratory field?"

What this study demonstrates — and what it does not

The measurements demonstrate that AI can use company websites directly and frequently. They also show that, on other queries, it can mobilize dozens of domains and hundreds of citations without using the studied company's domain at all. They therefore show a strong asymmetry of consultation.

They do not, however, allow the exact mechanisms behind that selection to be identified. Nor do they demonstrate that being cited as a source is enough to be recommended, or that any particular GEO technique allows entry into that exploratory field. That is precisely why measurement matters.

Before optimizing, measure

GEO raises a legitimate question: how can information be made easier for artificial intelligence to use? But a strategy built solely around that question risks forgetting another one, which comes just before it: does AI come looking for that information at all?

Our measurements show that it can. They also show that it can carry out hundreds of consultations elsewhere without ever reaching the studied domain. So before asking "how do I optimize my site for AI?", a company might start by asking: "when AI looks for an answer on my market, does it actually come to my site?"

Accessible does not mean consulted. Optimized does not mean selected. And selected does not necessarily mean recommended.

That may well be where the measurement of GEO really begins.