What GEO is, and why a site can be invisible despite good SEO

GEO is the optimization of a site for answers generated by artificial intelligence: for AI Overviews in Google, and for ChatGPT, Claude, Perplexity and Copilot. GEO stands for generative engine optimization. The goal is different from classic search optimization: it is not about a high position on the list of results, but about the model naming your company in the answer it shows instead of that list.

The difference has a practical consequence. A site can sit in the top three results and never appear in the answer the person asking actually reads. Visibility is there in the statistics, and no clients come out of it.

How GEO differs from SEO

Put briefly: SEO fights for a place on the list, GEO for a sentence in the answer.

GEO does not replace SEO and is not separate from it. Models draw on the same index, so a site absent from search will not be cited. SEO is the precondition, GEO is a layer on top of it.

Where models get their answers

Models draw on three sources at once. First, on knowledge memorized during training, which is always older than today. Second, on a search engine the model reaches for while answering. Third, on pages it fetches and reads on the spot.

You have influence over the second and third only, and those are the ones that count. The conclusion is simple: the site has to be reachable by crawlers, readable without running scripts, and built so that a finished answer can be lifted out of it.

What really makes a model cite a page

Five things go into it, listed from the most important.

A passage that answers the whole question. The model is looking for a piece of text that stands as an answer on its own. A paragraph starting “as mentioned above” is useless, because taken out of context it means nothing. A heading in the form of a question, with the answer in the first sentence under it, is the simplest way to produce a passage like that.

Specifics that can be verified. Numbers, dates, names, conditions. The sentence “we serve clients all over the country” carries nothing quotable. The sentence “wholesale shipping from twenty cartons, dispatched within two working days” does.

Structured data. Markup in the JSON-LD format states outright what the company is, which service is described and who the author is. It is the one place where you can tell a machine something unambiguously, instead of hoping it reads it out of the text.

A consistent description of the company across the web. The name, the address and the scope of the business should read the same on the site, in the Google Business Profile and in industry directories. Discrepancies lower a model’s confidence, and an unsure model would rather name nobody.

Access for AI crawlers. The crawlers of language models are separate programs from Google’s crawler, and they are allowed in or blocked separately in the robots.txt file. It happens that a company blocks them without realizing, along with everything else.

It is worth knowing that the choice is finer-grained than it looks. Anthropic runs three separate crawlers: ClaudeBot collects content for training models, Claude-User fetches a page when the person asking asks Claude to check it, and Claude-SearchBot indexes content for search. Each of them is allowed in or blocked separately, so you can consent to being cited while refusing to be used for training, or the other way round. OpenAI and Perplexity make a similar split.

That is a business decision rather than a technical one, and it is worth making it deliberately instead of leaving it to chance.

How to check whether artificial intelligence names your company

The check takes fifteen minutes and needs no tools at all.

1. Write down five questions a client asks before buying. Not the name of your company, but the problem: “where to buy dog treats wholesale”, “how much does a company website cost”.
2. Ask each of them in ChatGPT, in Perplexity and in Google search, looking at the answer at the top rather than at the list of results.
3. Note whether your company is named, and in what context, and who appears instead of it.
4. Repeat a month later, with the same questions.

A single reading says little. The value comes from repetition, because only the second and third attempt show a direction. That is why you write the questions once and then stick to them.

What GEO will not solve

A citation is not a sale. The model often answers in full and the person asking never reaches the site, so the statistics show nothing even though the brand was named. That has to be accepted when settling what counts as a result.

GEO does not work quickly or predictably either. The rules change every few months, and the same questions can give different answers a week apart. Anyone promising a permanent place in the answers is promising something nobody controls.

And the basic point: if a site says nothing specific, no amount of optimization will make up for it. A model cites content, not the effort around it.

Where to start in a small company

Start with three steps, in this order, and each of them can be done on your own.

Check whether the site is in the index and who it lets in. Without a presence in the index the rest is pointless. Google Search Console and Bing Webmaster Tools are enough, both free, plus opening yourdomain.com/robots.txt and checking that the crawlers of language models are not blocked there.

Turn one page into a set of answers. Pick the one that matters most to you, and rewrite the headings into the client’s questions, and the first sentence under each heading into a complete answer. That usually gives the most for the least effort.

Do the measurement from the section above. Note the date, the questions asked and the result. In a month you will have something to compare against, instead of relying on an impression that things are better.

When this stops being a job you can do on your own

The three steps above exhaust what can be done without going into the code and without ongoing care. Beyond them begins work that needs access to the site’s templates and regular measurement. Every type of page has to be described with structured data, the company’s description across the web has to be put in order, the right crawlers have to be let in, and the context of the mentions has to be checked every month.

I do this as part of looking after sites, so visibility in search and in AI answers counts from day one for me, not from the moment somebody notices a problem. On the Smaczaki wholesale site the zero point was the day the site opened to search engines. I run the site you are reading the same way. At Spretech the zero point was the day I took over its care, because the site had been running earlier and the whole thing sits on HubSpot rather than on my server. That last part matters: the measurement does not depend on whose technology it is, only on whether anybody is running it at all.

If you would like to check how your site stands today, get in touch and let’s set up a call about scope.