Generative Engine Optimization (GEO): getting read, understood and cited by AI systems

Some of the questions about your market never reach a search engine at all: they are put to an AI assistant, which replies with a piece of text and, sometimes, with a few sources. If you appear in that answer, you are in the running. If you do not, then as far as that user is concerned you do not exist.

It is a channel that works by rules of its own. Ranking first on Google does not guarantee being cited in a generative answer, and the opposite happens more often than people think. They are two distinct jobs that rest in part on the same foundations: this is why they are measured separately.

Plenty of promises circulate on this subject. We prefer to tell you straight away where the data ends and the estimate begins.

What can be measured with certainty concerns your site, and it is verifiable: which bots read it and for what reason, what you allow them to read, how understandable your content is to a machine, and how clearly and consistently the site declares who you are. These are facts that sit in your traffic data and in your code.

What can only be observed concerns the behavior of the models: whether and how they name you, how consistently, and from which sources they retrieve you when they do. We find that out by querying them systematically, across many questions and several times over. It is a sample, not a counter, and we treat it as such.

The first part tells you what to fix. The second tells you whether it is working. Both are needed, but confusing them is the quickest way to fool yourself.

Where we start: who reads you, and why

AI systems read your site for different reasons.

They can crawl it to train their models: this is how your brand enters their memory. And they can also crawl it to index the site and later retrieve information while they are building an answer: this is how you get cited as a source.

Distinct mechanisms, with different bots and different rules. We measure them separately: how much they read you for training, how much they index you, how often they come back to you while building an answer. You only improve what you measure.

Accessibility to AI systems

Before optimizing anything, you have to decide who is allowed to read you and for what purpose. Access rules have to be set separately for the three types of crawling, because blocking training and blocking retrieval have opposite consequences: the first keeps your content out of future models, the second makes it impossible for you to be cited today.

It is a decision with a strategic side and a legal one, and we take it with you case by case, rather than applying the same blanket rule everywhere.

Machine readability

When a generative system retrieves information from the web with real-time search, it does not read a page: it extracts a portion of it, cutting it out of the source content, and uses it to build an answer.

What makes your content usable by an AI system at this stage is the way you have organized the information inside it.

We work so that every important passage in your pages stands up on its own, without depending on what came three paragraphs earlier; so that statements come with the data that supports them; and so that the structure of the page makes it obvious which part answers which question.

Consistency and strength of your entity

Before it can cite you, a model has to establish who you are: what you do, what you are an authority on, which topics and which people you are connected to. It does not work that out from a single page. It reconstructs it by putting together everything it finds about you, wherever that may be.

If that information is consistent with itself and confirmed by independent sources, the entity consolidates and your name is associated with your topics with confidence. If it is contradictory, incomplete or present only on your own site, one of three things happens: the model confuses you with someone else, it describes you inaccurately, or it names whoever has a sharper profile than yours instead.

We work to make that reconstruction unambiguous. Names, roles, areas of competence and connections declared consistently on the site, and aligned with what can be found about you elsewhere.

It is work that produces effects over two different timescales. In the immediate term, on how you are retrieved and described when somebody asks a question today and the AI decides to look for the answer on the web. In the long term, on how you will be represented inside the models that will be trained in the coming years, which is where everything written about you around the web right now ends up.

We cannot guarantee that you will appear in an answer from ChatGPT or from any other LLM, and we are wary of anyone who promises it. Generative systems change behavior from one week to the next, they do not publish the criteria by which they choose their sources, and they give nobody a tool that confirms or denies your position in the AI’s “mind”.

What we can do is this: make your site accessible and understandable to the systems that read it, take your name where they go looking, and measure over time whether things are moving. It is the same logic we apply to organic ranking, applied to a younger and less predictable channel.

If somebody is promising you more, ask them how they plan to verify it.

Market research measures how contested your niche is in generative answers: who gets cited today, on which questions, and how settled the group of sources the models draw on already is. It tells you whether there is room, and what it costs to take it, before you invest.

We have been working on it since before it had a name

We started measuring how the models cite their sources before anybody was selling it as a service, and we have not stopped since.

That means our recommendations do not come from an article read last week, but from measurements taken several times over, on systems that have changed a great deal in the meantime. Having watched them change is what makes it possible to tell a development that shifts results from one that just makes noise for a month. Our GEO·AI Observatory documents what we find, including the method by which we find it: you can read it before you even talk to us.

We keep data and estimates apart

It is the most uncomfortable part of the job and it is the one we insist on the most. Everything we state comes with the evidence that proves it.

On a channel like AI answers, which nobody can really control from the outside, it is the only defense you have against people who tell you what you want to hear. If we have verified something on the basis of objective data, we say so clearly. If instead it is an observation based on a sample of questions, we say that just as clearly. You need it in order to decide how much weight to give each piece of information when you have to allocate a budget.

We have not thrown SEO away

Visibility in the models’ answers also rests in part on foundations that have been in place for years: clear content, readable structure, recognized authority.

These are the same things we have worked on for more than fifteen years on the organic channel, and indeed a good part of the work we do here produces effects there too. Anyone who comes to GEO without that background treats every system as a case of its own and starts again from scratch every time a new one appears. We start from what we already know and add to it what has really changed.