GEO
Recommended by AI: how to measure it honestly
· 3 min read · Oddbird
Asking ChatGPT once whether you get named tells you little. Answers differ per attempt, per assistant and per phrasing. This is how to measure AI visibility in a way you can act on.
What you take away
- An AI answer is not a fixed ranking. The same question gives a slightly different answer on every attempt.
- Measure with a fixed set of questions, several times, in several assistants. Then it becomes a number instead of an impression.
- Look not only at whether you get named, but at who does and which source they are based on.
Almost everyone who hears about AI visibility does the same thing first: open ChatGPT, type a question and check whether their business shows up. That is a good start, but as a measurement it is unreliable. Ask the same question an hour later and you get a different answer. Sometimes you are in it, sometimes you are not.
Why one question proves nothing
A language model picks its words with a touch of chance. On top of that, for many questions it searches the web live, and which pages it finds differs each time. An answer is therefore more of a sample than a fixed result.
- Per attempt: the same question, asked again, regularly produces different names.
- Per assistant: ChatGPT, Gemini, Perplexity and Claude gather information in different ways, so they do not name the same businesses.
- Per phrasing: "best dentist in The Hague" and "dentist in The Hague with an appointment soon" are two different questions to a model.
- Per location: many assistants take into account where the person asking is, just like Google Maps.
Relying on one answer means drawing conclusions from chance. It may turn out better or worse, but you do not know.
How to measure it properly
It only becomes reliable when you repeat and count. That sounds laborious, but the setup is simple.
- Write a fixed list of questions. Not your services in your words, but the questions a customer asks just before choosing someone. Ten to twenty is enough to start.
- Ask every question several times, and in more than one assistant. Only then do you see what is structural and what is chance.
- Count per question how often you get named. That share is your visibility. It is rough, but you can compare it.
- Note who does get named, and which source sits below the answer. That is almost always where the explanation is.
- Repeat it every month with exactly the same questions. A measurement without comparison is a snapshot.
What the sources teach you
The most interesting part of the measurement is not the number but the list of sources. If a competitor keeps being recommended, there is usually a specific page behind it: a comparison site, an industry list, an article in a local outlet, or reviews with real substance. It is no secret; it sits right below the answer.
Check per source what it says about your competitor and not about you. The gap is often smaller than you think: a missing listing, a profile that was never filled in, or a page on your own site that does not answer the customer's question.
| What you see | What it often means | Where to start |
|---|---|---|
| You are never named | A model does not know you or cannot read your site well | Access for AI crawlers and structured data on your site |
| You are named sometimes | You exist, but there is little outside confirmation | Listings and reviews in the places your sector looks at |
| A competitor is named every time | They appear in a source the model trusts | Find that source and see whether you belong there too |
Why we have a separate branch for this
Tracking this by hand is possible, and we recommend it to everyone for a first impression. But dozens of questions, several times, in different assistants, every month again, is work nobody keeps up next to running a business. That is why we built Oddrank: our own branch that runs exactly this measurement, keeps track of it and shows what needs to happen to get recommended more often.
Whether you do it yourself or have it done, the principle stays the same: fixed questions, repeat, count and look at the sources. Then AI visibility becomes something you can improve, rather than something you occasionally hope for.