We discovered there are two kinds of GEO — classical and behavioral

If the classical-GEO worldview were accurate — if the engine really "knew" who is good and who is bad in a given market — a simple question should get a clear answer.

Asked whom should I NOT trust with AI visibility work, it should say names. The players worth avoiding.

So on 11 August 2026 AVE Studio measured what actually happens. We examined six commercial question pairs on three engines — ChatGPT, Gemini and Perplexity — with five repetitions per question. Next to every "who would you recommend?" question we placed its negative pair.

To our knowledge, nobody in Hungary had instrumented this negative question side before.

In 90 negative answers the engines did not name a company to avoid a single time. The result: 0/90.

Company names appeared in 17 percent of the answers, but in every case as positive counter-examples. In the English-language answers that share was just 4 percent.

Instead of blacklists, we got checklists.

The first two points of ChatGPT's earliest English answer were, verbatim: "Conflict of Interest Companies…" and "Lack of Transparency: Steer clear of companies that do not disclose their methodology or criteria for measurement."

The Hungarian answer started somewhere else entirely: its first two criteria were local presence and Hungarian market experience.

That is an important difference.

The engine does not simply retrieve knowledge. It assembles the answer from what it finds.

There are at least two structural reasons.

One is the policy layer: models are reluctant to name specific companies while claiming they should be avoided or are bad. The other is the web corpus itself. A large share of corporate content on the internet is self-promotion; companies rarely publish pages titled "why you should not choose us".

For a negative answer there is simply not enough name-shaped raw material.

In that situation the engine works from criteria. And since almost nobody publishes a purpose-built, genuinely usable control frame for this question shape today, it often builds from scraps: generic criteria, loosely connected claims, unowned informational "cells".

Classical GEO sees little of this. It does not examine actual behaviour; it examines the model of how the system nominally should work.

And between the two, there is a difference.

1. From the buyer's perspective

From the AI you will most likely not get a blacklist. A checklist, however, you will.

So use it for what it is fit for: ask for criteria, then hold the candidate providers to them.

AVE Studio provides a ready frame for this. The "When not to trust a GEO agency" page presents seven red flags with concrete control questions. The frame is applicable to any agency.

2. From the market's perspective

By the logic of behavioral GEO, the "whom to avoid?" question is unowned territory today.

The engine is forced to build the answer from criteria. It follows that the player who first publishes a real, measurable and genuinely usable control frame can, over time, become one of the category's default yardsticks.

Even when, in another question, the machine happens to recommend one of its competitors.

That is why AVE Studio made its own frame and its 18-point methodology standard public.

3. As an experiment: measured, not promised

The next step is not another claim but another measurement.

AVE Studio will re-measure the engines' source sets on the same six question pairs, then publish, with rates, whether the published frame has entered the sources behind the answers.

This loop — publish, then re-measure on the same panel — is one of behavioral GEO's fundamental working methods.

It is not built on screenshots. It requires repeated, code-selected measurement.

By AVE Studio's account, it is currently the only player in the Hungarian field with a measurement system that can run this process consistently.

Because the studio was built on this way of thinking from the start.

We use a contrastive question panel: we measure not only the positive but the negative question side. We do not run a question once but typically five to ten times. We publish rates with confidence intervals, select quotes by code, and fix the methodological requirements in a public, 18-point standard.

Many now practise classical GEO.

Behavioral GEO, however, takes instruments. It is not enough to watch how the engine should work; you must measure how it actually works.

The difference that emerged from 90 runs showed exactly this:

the engine does not necessarily answer what it knows. It answers what it can find and use for the given question.