GEO
GEO: why your brand needs to be found by AI
· 6 min read · Oddbird
More and more people put their questions to ChatGPT and Perplexity instead of Google. That changes not only where you get found, but how. Here is what happens under the hood, and what you can do about it.
What you take away
- An AI answer is not a list of ten options but one summary with two or three sources. Position eleven no longer exists.
- Models cite what they can read, recognise and see confirmed. Technology, entity and repetition are three separate problems.
- You can measure today whether you get named, and it costs you an afternoon.
The shift underway is not that people search less. They search just as much, but increasingly get an answer instead of a list. Someone who used to get ten blue links and picked for themselves now gets a paragraph with two or three sources underneath. For your brand that means something uncomfortable: the room to be found has narrowed.
With classic SEO, position eight was still a position. Someone scrolling would come across you. In an AI answer there is no position eight. You are named or you are not, and that difference is absolute. That is the core of why Generative Engine Optimization, GEO, is something other than SEO with a new name.
What happens between the question and the answer
To understand where you have influence, you need to know what happens in those few seconds. Roughly four steps, and at every step you can drop out.
- The question is pulled apart. One user question becomes several internal searches. "Where do I rent a car on Curaçao" also becomes "car rental Willemstad", "reliable car rental Curaçao reviews" and "car hire Curaçao prices".
- Pages are fetched. Not your whole site, but individual chunks of text resembling those sub-questions. A model rarely reads a page top to bottom; it takes fragments.
- Those fragments are weighed. Does more than one source say the same thing? Is it recent? Does the source appear elsewhere as trustworthy?
- An answer is written with sources. What makes it, makes it. The rest does not exist for that user.
Every step in that chain is a place where you can act. And at most companies it already goes wrong at step two, for a reason that has nothing to do with marketing.
Problem one: is the crawler even allowed in
Before a model can say anything about you, a crawler has to have fetched your site. Those are different crawlers from Googlebot, and they are far from universally allowed. Often that was not even a decision: someone once added a block, or a plugin did it by itself.
| Crawler | Whose | What it decides |
|---|---|---|
| GPTBot | OpenAI | Training data. Does not decide whether you appear in ChatGPT search results. |
| OAI-SearchBot | OpenAI | ChatGPT's search function. This is the crawler that governs citations. |
| ChatGPT-User | OpenAI | Fetches a page live during a conversation, at the user's request. |
| ClaudeBot | Anthropic | What Claude can read from your site. |
| PerplexityBot | Perplexity | Perplexity, which almost always answers with sources. |
| Google-Extended | Gemini and the AI Overviews above the search results. |
Checking takes a minute: fetch your own robots.txt and see whether any of these names gets a Disallow on the whole site. If so, that is your first action point, and every other action is pointless until it is resolved.
Problem two: does a model know who you are
A crawler that fetches your site still only has text. That you are a car rental company on Curaçao, or an AI platform for acquisitions, has to be recorded somewhere unambiguously. Structured data exists for that: a block of machine-readable facts stating who you are, where you sit, what you deliver and where else you appear.
In practice this is where we find the biggest gap at most sites. The site is fine, the copy is good, and there is not a single line of structured data. A model then has to infer from running text what kind of business this is, and that goes wrong more often than not: it confuses companies with the same name, or mistakes a service for a product name.
- Organization or LocalBusiness: who you are, with address, phone number and the places you also appear.
- WebSite: the site as a whole, tied to that organisation.
- Service for every service you genuinely deliver, not for every page you happen to have.
- FAQPage, but only where the questions are also visible on the page. Markup without a visible question is a violation, not a shortcut.
- BreadcrumbList, so the hierarchy of your site is readable.
Have those blocks refer to each other with one fixed id instead of describing the company again on every page. Then a crawler reads one business with several pages, rather than ten separate mentions that happen to share a name.
Problem three: does anyone else confirm it
This is the part that takes longest and that you control least. A model weighs whether several independent sources say the same thing. One page where you claim you are the best weighs almost nothing. Three places outside your own domain naming the same facts do weigh.
SEO was about ranking. GEO is about being named. That does not ask for more pages, but for more confirmation from outside.
In practice: make sure your details are identical everywhere, including how you write your name and address. Make sure you appear in the sources your sector looks at, whether that is a trade register, a trade publication or a comparison site. And write about the questions that come before the purchase, not only about your offer, because those are the questions put to an AI.
Measure it yourself, this afternoon
You do not need an agency to find out where you stand. Do this with a notebook beside you.
- Write down the five questions a customer asks just before they end up with you. Not your services, but their problem in their words.
- Put them to ChatGPT, Perplexity and Gemini. Note per question who gets named and which source sits underneath.
- If you are not named, look at who is. Open that source and see what is there that is missing from yours.
- Fetch your robots.txt and check the crawlers from the table above.
- View your page source and search for application/ld+json. If there is nothing there, you know your biggest action point.
What you find is almost always the same pattern: the site is fine, the technology underneath is half set up, and outside confirmation is missing entirely. That is good news, because the first two are days of work, not months.
What this is not
GEO is not a trick to convince a model. There is no keyword density, no hidden instruction in your HTML that works, and anyone promising otherwise is selling you something that breaks two model versions from now. What does work is dull and durable: making sure you are readable, that who you are is recorded, and that others confirm it.
That is exactly the same movement as ten years ago in SEO, when tricks gave way to work that was worthwhile even without a search engine. The difference is that the reckoning is harder now: with ten blue links you could come eighth. In a three-sentence answer there is no eighth place.