How to Find Out Whether AI Recommends Your Business

There is a question that used to have a clear answer and now does not: where do we rank?

For twenty years that meant a position on a results page. You could look it up, argue about it, and watch it move. Increasingly, though, the buyer never sees a results page. They ask a model — "who should I use for X in Y?" — and get two or three names with a sentence of justification each.

If your name is not among them, you have not ranked eleventh. You are absent from the conversation entirely, and no amount of scrolling will find you.

This article is about how to measure that, honestly, including what the measurement cannot tell you.

Why "what's my AI ranking" is the wrong question

There is no ranking. There is no index you can be at position 4 in, no score being kept, nothing to look up. A model produces an answer by generating it, and it generates a slightly different one every time.

Ask ChatGPT twice which dental clinics it would recommend in a given district and you will often get two overlapping but non-identical lists. Ask from a different account, in a different language, or a week later, and it shifts again. This is not a defect to be engineered around. It is what these systems are.

So the honest framing is not what is our rank but how often are we the answer — and the only way to know that is to ask, repeatedly, and count.

The method

You can do this by hand, for free, in about twenty minutes. Here is the procedure:

1. Write the question a customer would actually ask.

Not the question you wish they asked. If you run a clinic, the question is not "what is the best hair transplant methodology" — it is "where should I go for a hair transplant in Istanbul?" Phrase it the way a person types it, in the language they type it in.

2. Ask each engine separately.

ChatGPT, Gemini, Claude, Perplexity and Grok are different systems trained on different data with different retrieval behaviour. A business that is well known to one can be invisible to another, and the gap is often large. An average across all of them hides more than it shows, so record each separately.

3. Ask from a clean session.

Logged out, or in a temporary chat. A model that already knows who you are because you have been talking about your own company for an hour is not telling you about the world; it is telling you about the conversation.

4. Write down three things per answer.

• Were you named at all?
• If so, in what position among the businesses listed?
• Who else was named?

5. Repeat, at least three times per engine.

One run is an anecdote. Being named in four runs out of five is a position. Being named once in five is noise, and treating it as a win is how people convince themselves things are working when they are not.

That third column — who else was named — is usually the most valuable output and the one people skip. It is a list of the businesses that are winning the query you want, assembled by the system doing the recommending.

If you would rather not do this by hand, the AI Visibility Checker on this site runs the same procedure across every engine and reports what each one said, including which engines it could not reach.

Reading the result

There are four outcomes, and they call for entirely different work. Conflating them is the most common mistake.

Not mentioned by anything

The default state for most businesses, and it is not a judgement on quality. Models name what has been written about. If the open web contains no description of what you specialise in, where you operate, and no third party saying you are good at it, there is nothing for a model to retrieve. You are not losing a competition — you are not in it.

The work here is foundational and slow: a clear, specific, machine-readable statement of what you do, and then the much harder task of getting described somewhere other than your own website.

Named in the prose, but not recommended

The most useful result you can get, and the most commonly misread.

The model knows you exist. It has your facts. Asked who to use, it put someone else forward. That is not an awareness problem — it is a judgement problem, and judgement is downstream of evidence. Usually one of three things is true: your positioning is too broad to be the obvious answer to any specific question, the corroboration is thin (no directory listings, no press, no reviews the model can see), or a competitor's evidence is simply better.

This is the outcome with the shortest path to improvement, because the hard part — being known at all — is already done.

Recommended, but fourth or later

You are in the consideration set and losing on evidence. Go and look at who came first. Not to copy them, but to find out what exists about them that does not exist about you. It is rarely a secret: usually more third-party coverage, a more specific claim, or a longer record in the same place.

Recommended first, consistently

Worth knowing, worth monitoring, and worth not over-reading. It is a position in a system that changes. Measure it monthly rather than celebrating it once.

What actually moves this

Three things, in the order they matter.

Be findable by machines at all. If your site blocks AI crawlers, renders its content only after JavaScript runs, or states none of its facts in structured data, you have made yourself expensive to retrieve. This is the prerequisite, not the strategy — and it is checkable in a minute with the AI & LLM Visibility Checker, which asks whether models can read you rather than whether they recommend you, or by reading the llms.txt guide.

Make a narrower claim. "Digital agency" is a category with thousands of members and no way for a model to separate them. "Multilingual e-commerce builds for health tourism clinics" is a category with very few, and a model asked that question has an easy answer. The narrower the claim you can truthfully make, the more often you are the answer — and the fewer questions you are the answer to. That trade is usually worth making.

Get corroborated off your own site. A model weighing two businesses that both claim excellence leans on everything that is not their own marketing: industry directories, local press, forum threads, review platforms, the websites of people you have worked with. This is slow, unglamorous, and the part most people skip in favour of rewriting their homepage for the fourth time.

What does not work, in our experience: stuffing "AI-optimised" language into pages, publishing thin content at volume hoping some of it gets retrieved, or any service that promises a guaranteed position in an AI answer. There is no position to guarantee, and the people selling it know that.

The limits of this measurement

Worth stating plainly, because the number looks more precise than it is.

It is a sample, not a reading. Every run is one draw from a probabilistic system. Three runs across five engines is fifteen data points, which is enough to see a pattern and not enough to detect a small change.

It is not personalised the way a real buyer's session is. Real users have history, location signals, and sometimes a paid tier with different retrieval. Your clean-session result is the population average, not any individual's experience.

Models change underneath you. A model update can reshuffle answers overnight for reasons that have nothing to do with anything you did. This is why the trend over months matters and a single month-on-month move usually does not.

Being recommended is not the same as being chosen. It gets you into the consideration set. Everything after that is the same work it has always been.

Where this sits next to SEO

It does not replace it. The overlap is larger than the GEO industry likes to admit — the structured, specific, well-corroborated site that models recommend is broadly the same site that ranks. We have written about what genuinely differs and what carries over rather than repeating it here.

The practical position: keep doing the search work, add the measurement above so you can see the other channel, and be suspicious of anyone who tells you the first one is finished.

The short version

Ask the question your customer asks, from a clean session, on each engine, three times. Record whether you were named, where, and who else was. Do it monthly.

If nothing names you, the work is being described somewhere other than your own site. If something names you but recommends a competitor, the work is a narrower claim and better evidence. Those are different jobs, and knowing which one you have is most of the value of measuring at all.