Understand

SEO and GEO visibility through AI agents

For two years now, a growing share of searches no longer ends on a list of links, but on a written answer. Someone asks an assistant which software to choose, and reads three names. If yours is not among them, your position on Google does not save you: the question no longer goes through there. Search visibility now has two faces, and the second is measured by none of the tools of the first.

1. Two kinds of visibility, not one

SEO aims for a place in a list of results. Its mechanics are well known: pages that genuinely address an intent, a technically clean site, links pointing to you.

GEO — generative engine optimisation — aims at something else: being cited in an answer written by a model. It is not a ranking, it is a presence or an absence. You are not seventh there: you are in, or you are out.

The two are not obtained through the same gestures. A model does not crawl the web like an indexing robot; it draws on what it has learned, on what it goes to fetch at the moment of answering, and on sources it judges reliable. A page designed to catch a short query is often useless to it, whereas a page that clearly explains a subject — with verifiable claims and an identifiable author — ends up cited.

What does not change: both reward treating a subject honestly. What does change: the second is verified by asking the question, not by consulting a ranking.

2. What genuinely gets measured

Four measurements are enough to know where you stand. Three are classic, the fourth is new.

Rankings — for your keywords, country by country and language by language, with dated history. A position without history says nothing: it is the slope that informs, not the point.
The state of the site — an audit that reviews the technical side, the content, mobile display, structured data and performance. Most serious blockers are there, not in the keywords.
Competitors — who occupies the places you are aiming for, and on what.
Share of voice in AI answers — on the questions your customers actually ask, how often you are cited, against whom, and which sources the assistant used to cite you. This is the measurement that classic search tools do not produce.

That last one is recorded honestly in only one way: by actually asking an assistant the question and reading the answer. An "AI visibility" figure computed without ever querying a model is an estimate disguised as a measurement.

3. Being readable by the models

Before hoping to be cited, you have to be readable. Three points decide almost everything, and none is difficult.

Do not shut the door. Many sites block the models' crawlers in their robots.txt, often without knowing it. A site invisible to those crawlers will never be cited, whatever the quality of its pages.

Say who is speaking and about what. Correct structured data — the organisation, the product, the author — lets the model attach a claim to an identified entity rather than to an anonymous text.

Publish an llms.txt. It is a short file, at the root of the site, summarising what the company is and pointing to the reference pages. It is in no way mandatory, but it gives the model a short and accurate version, rather than leaving it to reconstruct your business from a promotional home page.

4. What an agent does every day

Search visibility rarely fails through ignorance. It fails because nobody looks at it for six weeks, and the drop is discovered once it has settled in.

This is exactly the kind of work that gets delegated: regular, measurable, with no judgement to pass. In nullbot, an agent can record your keyword rankings, redo the site audit, re-ask the GEO questions to see whether you are still cited, compare with the day before, and flag only what has moved. What it observes feeds an action plan ranked by impact and effort — each line of which can be handed to another agent, or handled by hand.

The gain is not measuring faster. It is that a drop in ranking or a lost citation is seen the next day rather than the next quarter.

5. What cannot be automated

Three things remain out of an agent's reach, and it is better to know that before starting.

Authority. Being cited by other sites, by the press, by serious directories: this is earned through the company's real work and through relationships. No line of code manufactures it, and the schemes that claim otherwise get penalised.

Having something to say. A page that brings nothing no longer ranks and does not get cited. An agent writes quickly, which makes it all the more tempting to publish emptiness — and emptiness, today, is detected.

Patience. A young domain takes months to be taken seriously, whatever the quality of what it publishes. A tool that promises results in three weeks is describing a mechanism that does not exist.

6. Frequently asked questions

What exactly is GEO?

Generative engine optimisation: everything that makes a company citable by an AI assistant when it writes an answer. SEO seeks a place in a list; GEO seeks a mention in a paragraph. The two are worked on in parallel, with different measurements.

How do I know whether ChatGPT or Perplexity talks about my company?

By asking the question and reading the answer — there is no other reliable method. That is what nullbot's GEO check does: it puts your questions to an assistant, records whether you are cited, with which excerpt, which sources were used, and which competitors appear beside you. The answer is kept as evidence, so the measurement stays verifiable.

Do I need a paid tool on top?

Rank tracking, the site audit and the GEO checks are included in nullbot. An external search provider can be connected if you have one and want more depth, but it is not necessary to get started.

Where should I start?

With a site audit and five GEO questions — the ones your customers really ask before buying. In an hour you will know whether you are readable by the models and whether you are already being mentioned. Rank tracking comes afterwards: there is no point measuring a slope on a site that still blocks the crawlers.