Increasingly, an assistant decides who gets mentioned.

Generative engine optimization — usually shortened to GEO — is the work that gets an assistant to name your business when someone asks a question you can answer. Not a place in a list, but a sentence in an answer, with a source next to it.

It differs from ordinary findability on three points. First, there is no ranking: there is one answer, and you are either in it or not. Second, the shape of your text weighs more heavily — a model has to be able to extract a usable piece. Third, what others say about you counts towards deciding whether the answer is right.

At the same time it is less new than it sounds. The assistants answering today lean heavily on the same search indexes that already existed. Whoever is technically unfindable is unfindable in both worlds; whoever answers a question better than the rest wins in both. GEO is therefore an addition rather than a replacement.

What is new: the number of clicks can drop while your visibility rises. Someone who already gets the answer in the conversation may not visit your site. That makes it more important to be named well and completely in the answer, with your name and your specialism — because that is what sticks.

When someone asks 'which installer in Utrecht can fit a heat pump', something predictable happens under the hood. The question is turned into one or more searches, sources are fetched, what looks usable is selected, and one answer is written from those pieces.

Some assistants fetch sources live while answering; others lean more on what they stored earlier. In both cases a search index sits in between. Whoever is not in it does not take part — however good the copy.

In selecting, a model weighs whether a piece of text answers precisely this question, whether it is recent and concrete, and whether it comes from a source that repeatedly appears reliable. Vague marketing language yields nothing to quote; a paragraph with a clear fact does.

In composing, corroboration plays a role. If your site says something that appears nowhere else, the chance of it landing in the answer is smaller. If your site, your business profile, a directory and a review page all say the same, it becomes a fact the model dares lean on.

And there is a limit: models make mistakes, confuse businesses with the same name and sometimes fetch outdated details. That cannot be fully prevented — only reduced by being identical and unambiguous everywhere.

Step one is access. AI crawlers have their own names and can be allowed or refused separately in your robots file. If you want to appear in answers you have to allow them — and that is a deliberate choice, because it also means your copy gets read.

Step two is readability without scripts. Many fetchers do not run JavaScript or do not wait for it. If your text only enters the page after something has loaded, they see an empty page. Switch JavaScript off in your browser and look at what remains — that is what they see.

Step three is speed and reliability. A fetcher does not wait long. A server answering slowly or occasionally returning an error drops out while the competitor is done within a second.

Step four is structure that carries meaning: headings that pose the question, structured data saying what a page describes, a visible date on content that changes, and clear sourcing on figures. That helps both selection and correct citation.

Step five is an llms.txt: a short, factual file with who you are, what you offer and where the main pages live. It is a convention rather than an official standard, but it costs little and forces you to write your own facts down unambiguously.

A model cites passages, not pages. The trick is therefore to write so that separate, accurate passages can be lifted out and stand on their own. A paragraph only comprehensible with the three before it ends up nowhere.

The form that works best is simple: a heading posing the question the way people ask it, an answer of two or three sentences below, and only then the nuance. That reads better for people too, so you need not choose between readers and machines.

Be concrete. Numbers, terms, conditions, exceptions and definitions are exactly what gets taken over. 'Sharp prices' and 'years of experience' are not — there is nothing there that improves an answer.

Put facts in ordinary sentences, not only in a table, an image or a PDF. Tables are fine as an addition, but the fact belongs in the running text as well. What exists only in imagery usually does not exist for a model.

And keep it current. Outdated prices or copy from a previous year are either skipped or — more annoyingly — cited anyway. A visible date on content that changes helps both sides.

Assistants reason about things: businesses, places, services, people. To be named as the answer to 'who can do this for me', an unambiguous picture has to exist of who you are, what you do and where you work.

That starts with consistency. Name, address, phone number, service area and specialism should be exactly the same on your site, in your business profile, in directories, on social profiles and in the trade register. Every deviation forces a guess, and guessing often means naming someone else.

Put those details explicitly in your site's structured data too, so no interpretation is needed. For businesses with an address that includes a local business profile, with opening hours and the area you serve.

Also make sure there is a page somewhere that simply explains who you are: what you do, for whom, since when, with which certifications, in which area. That sounds dull and it is exactly what a model needs to name you with confidence.

And have it corroborated. Listings with trade associations, suppliers, local organisations and trade media are not only good for ordinary findability; they are the second voice saying your details are right.

An answer about 'the best option near me' rarely comes from one site. The model leans on reviews, directories, comparison sites, forums and articles. You cannot write that part, but you can influence it.

Reviews weigh most. Not only the score, but what they say: people naming what you did and where supply exactly the details a model can build on. So ask for them the moment someone is satisfied, and reply to what is written.

Directories and trade platforms count as corroboration, provided your listing is correct and complete. An old address in a well-found directory is an active problem: the model sees two versions and picks one.

Trade and local media are the more expensive but strongest route: if a journalist or trade magazine names you as an example, that is a source weighing more than what you write yourself. You earn that with work people talk about.

Forums and Q&A sites deserve attention but not manipulation. Taking part as a professional explaining something is fine; fake accounts dropping your own name are visible and backfire the moment they are noticed.

There is no dashboard telling you how often an assistant names you. What you can do: ask the same questions your customers ask, at fixed intervals and across several assistants, and write down what comes out. That is manual work, but it gives a fairer picture than any tool promising otherwise.

Watch three things: are you named, is what is said about you correct, and is your site cited as a source. The second point is often more important than the first — a wrong opening hour or a service you do not offer costs customers who never get in touch.

In your own numbers you see it indirectly. Referrals from assistants show up as separate sources in your analytics, but the largest part stays invisible: people who remember your name and later come directly or search for it. A rise in direct traffic and in searches for your own name is the best signal of that.

Just ask customers too. 'How did you find us' increasingly produces 'ChatGPT said so' — and that is more measurable than anything else.

And correct what is wrong: fix incorrect details in directories, make a page that sets the record straight, and make sure the right version is corroborated everywhere. Models repeat what they encounter most often.

There is no setting, contract or technique that makes an assistant name you. Anyone promising that is selling something else. What is possible: being the source that is easiest and safest to cite.

Do not expect stability either. Answers differ per assistant, per version, per phrasing of the question and sometimes per attempt. So measure across several attempts and several weeks, not on one screenshot.

What to avoid: hidden text meant only for machines, pages showing something different to a crawler than to a visitor, and instructions in your copy trying to steer a model. That last one — sometimes called prompt injection — is recognised, can lead to exclusion and sits in your visible text anyway.

Also avoid: mass-producing thin pages 'for AI'. The same rule applies as with search engines — pages that answer no question weaken the rest of your site.

And be honest about figures and claims. If a model adopts your claim and it turns out to be wrong, an incorrect fact about your business ends up in a thousand conversations. That is a risk most businesses underestimate.