What GEO actually means in 2026: measuring visibility in AI answers

For twenty years, being visible online worked through a simple interface. A customer searched, a search engine returned a list, and companies competed for a position on it. That competition gave rise to one of the core disciplines of digital marketing: SEO.

That interface is changing. Today a potential customer can ask an AI assistant "what software should a 50-person company use to manage its sales team?" and, more and more often, they don't get ten links. They get an answer.

It looks like a small difference, but it changes the unit of visibility. In a search engine, companies compete to be clicked. In an AI assistant, they compete to be found, understood and recommended inside the answer.

That is the problem Generative Engine Optimization (GEO) tries to address. This article explains how we think it should be understood and measured, with our own data.

GEO is not SEO for ChatGPT

The term was introduced in a 2023 research paper (Aggarwal et al.) that studied how to improve the visibility of content inside answers produced by generative engines. Since then, a whole industry has formed around GEO, AEO, "LLM optimization" and AI visibility.

Treating GEO as keyword optimization misses the real change. A traditional search looks roughly like this:

Query → ranking → click → website

A conversation with an AI assistant looks more like this:

Question → retrieval → comparison → synthesis → answer

Your website still takes part, but it is no longer necessarily the destination. The assistant may read your site, your competitors' sites, media, review platforms, forums and directories, and then decide what goes into the answer.

So the question for a company changes too. It is no longer only "where do we rank on Google?", but:

When customers ask AI about the problems we solve, are we part of the answer?

Three layers of AI visibility

At Yovara we think about it in three layers.

1. Being found. Does AI bring your company up when the customer doesn't mention you? Asking "what is Yovara?" tests whether AI knows the brand. Asking "what tools tell me whether AI recommends my company?" tests commercial discovery. They are different tests, and mixing them inflates the result: a company can be perfectly well described when asked about by name and be invisible when customers describe the problem it solves.

2. Being understood. When AI does talk about you, does it get right what you do, who you serve, which problems you solve and what sets you apart? That information comes from your website, but also from everything around it. AI visibility is partly a website problem, not only a website problem.

3. Being recommended. When someone asks "what should I use?", does the AI include you? Who does it include instead? And why? This is where GEO stops being a technical exercise and becomes a mix of content, authority, third-party evidence, technical accessibility and competitive positioning.

It is not a ranking

There is no global list where a company is "number 3 in ChatGPT". Answers depend on the model, the search behind it, the country, the language and the exact wording of the question.

The research points the same way. A 2026 survey of 45 studies (Martinez; a preprint, not yet peer-reviewed) argues that GEO is not a single ranking task but a stochastic, partially observable pipeline: whether the engine searches at all, what it crawls and retrieves, what it cites, how prominently, and what the user finally reads. An industry study by Semrush and Kevin Indig of more than 50,000 brands across 1,094 US categories in ChatGPT found that only 15.2 % of categories had a clear "owner" brand.

The practical consequence: you can't measure GEO by asking one question once. If an assistant mentions you once, you haven't "won GEO"; if it leaves you out once, you aren't necessarily invisible. Measuring requires repeating, covering several commercially relevant questions and comparing several assistants. If you want to do it yourself, we explain it step by step in how to check if ChatGPT recommends your company.

What we see in our own measurements

Theory is useful, but we prefer data. These are the aggregated, anonymous numbers from our first 14 analyses: 7 companies (including our own) and 2,196 answers from ChatGPT and Perplexity to questions that don't name the company. It is a small, early sample, not a market study, but it already shows some patterns:

• The company was named in 13 % of the answers.
• In 19 %, the AI named at least one competitor and not the company.
• In 41 answers, the AI cited the company's own website as a source without naming the company. The website had become a source for the answer, but the company was not part of it as a brand. A citation and a recommendation are not the same thing.
• The two assistants disagree a lot. Of the 89 questions where at least one of them named the company, only 40 (45 %) were named by both. Measuring only one assistant can give a very incomplete picture.
• Repeating the same question on the same assistant is fairly stable: when a company was named, 93 % of the time it was named in all three repetitions.

The last two points matter for anyone measuring this. The biggest variation we see isn't in asking the same thing again: it is between different assistants.

A real case: fixing is not the same as improving

Futboleras is a women's football platform in Spain, and one of our own projects. We measure it with 22 questions that don't name it, each answered 6 times (132 answers per measurement).

• 23 September. Two measurements on the same day: 61 % both times (the share of answers that name Futboleras). Its own website was a cited source in 54 % of the answers. The technical review found 93 pages whose content only appeared after running JavaScript, 93 pages without a main heading (H1) and 135 pages with very little text.
• The team fixed them. In the next measurement, on 26 September, the JavaScript and heading problems were no longer detected, and pages with little text dropped from 135 to 19.
• The result: 63 % on 26 September and 64 % on 29 September. Its website as a cited source: 58 %.

A move from 61 % to 64 % is within the normal variation we see between measurements. So we can't say the fixes improved Futboleras' visibility in AI answers, and we don't. What we can say is that the site is now easier for AI to read, and that a real effect, if there is one, will need more time and more measurements to show.

That is the honest way to do GEO: first fix what you can observe, then wait, then measure again, and only then judge whether visibility changed.

The question matters more than the keyword

SEO taught us to think in keywords. GEO forces us to think in questions and intent.

A company that sells accounting software can be asked about as "best accounting software for small businesses", but also as "how do I automate my company's invoices?", "which tools integrate with Stripe?" or "what should a five-person startup use?". Each question is a different moment of the purchase.

The goal isn't to "rank for a prompt", nor to measure hundreds of arbitrary prompts to produce a big number. It is to know whether AI consistently connects your company with the problems your customers are trying to solve.

Citations tell only part of the story

Several quite different things can happen:

• the AI doesn't consult your website at all;
• it consults it but doesn't cite it;
• it cites your website but doesn't name your company;
• it names you based on another source;
• it recommends you and cites evidence that supports it.

Each state calls for a different intervention. When the assistant consulted your page but relied on other sources (some services report which pages they consulted besides the ones they cite), it did find you: the problem isn't discovery, it is that your content wasn't clear or authoritative enough to become part of the answer. When it cites your website without naming you, the problem is different again. A useful measurement has to tell these cases apart, not just count mentions.

Measure, understand, act, measure again

We believe serious GEO needs four layers:

1. Measurement: which questions matter, where you appear, on which assistants and how consistently.
2. Evidence: which sources the AI uses, whether it consults your website, which external domains weigh in.
3. Diagnosis: why competitors appear where you don't. Is the page missing, can it not be read, is it weak, is your positioning unclear, do others have more outside evidence? Always separating what was observed from what is a hypothesis: generative systems are too opaque to present every correlation as a cause.
4. Action: what exactly to change.

And since AI visibility isn't static, the process is a loop, not a checklist: measure → explain → recommend → act → measure again, with the same questions, to separate real change from normal variation. That is far more defensible than promising a secret formula to "rank first in ChatGPT". Today no one can honestly promise that.

What we don't know yet

GEO is still a young discipline. We can measure what AI assistants consult, cite and mention, compare those observations over time and check whether changes coincide with different outcomes. What we can't do yet is turn every correlation into a cause.

There is no universal recipe that guarantees that changing a page today will make every assistant more likely to recommend you tomorrow. Building that evidence takes controlled, repeated measurements over time. It is one of the questions we want to investigate with Yovara.

How we measure it at Yovara

Yovara asks ChatGPT and Perplexity the questions your customers would ask, about 30 per company and most of them without naming you. Each is asked three times on each assistant. We record where you appear, who appears instead and which sources each answer relies on, and turn it into recommendations linked to that evidence. The full detail is in our methodology.

Frequently asked questions

What is GEO (Generative Engine Optimization)?

It is the discipline of improving how AI assistants such as ChatGPT or Perplexity find, understand and recommend a company when they answer users' questions. Unlike SEO, the goal isn't a position in a list of links but being part of the answer.

Does GEO replace SEO?

No. Much of what makes a website understandable, crawlable and authoritative is still valid. GEO adds a new layer on top: how systems that read, compare and synthesize information represent you.

Can you guarantee that ChatGPT will recommend my company?

No. Answers depend on the model, the search behind it, the country, the language and the wording of the question. What can be done is to measure consistently, understand why you don't appear and check, by measuring again, whether the changes work.

How do you measure AI visibility?

By asking the questions your customers ask, without naming your company, several times and on several assistants, and recording mentions, competitors and cited sources. Measuring a single question once isn't enough.

SEO isn't disappearing, and neither are websites. But a new discovery layer is being built on top of them. For twenty years, companies learned to make themselves visible to search engines; now they also need to understand how they are represented by systems that read, compare and recommend.

On Google you competed for a position. In AI you compete to become part of the answer.