Can I just ask ChatGPT whether it recommends my business?

You can, but the answer is close to worthless. Putting your name in the question tells the model which answer you want, and it will usually oblige. The only meaningful test is a question that does not contain your name, phrased the way a customer with no knowledge of you would phrase it.

Why do I get a different answer every time I ask?

Language models are non-deterministic, so the same prompt can produce different wording and sometimes a different shortlist. Browsing adds more variation, because the live results change. That is why a single run proves nothing and a count across several runs does.

How many times should I run each question?

Five runs per question is enough to tell a reliable name from a lucky one, and ten is better if the question matters commercially. Record how many runs named you rather than whether any run did. Being named in eight of ten is a very different position from one of ten.

How often should I repeat the whole check?

Monthly is a sensible rhythm for most small businesses. It is frequent enough to catch a change after you have published or fixed something, and infrequent enough that you are looking at a trend rather than at the normal run-to-run noise.

Does checking in ChatGPT tell me about the other assistants?

Only loosely. Claude, Gemini, Perplexity and Copilot draw on different sources and search partners, and they frequently name different businesses for the same question. If an assistant matters to your customers, test it separately rather than assuming the result carries over.

How to check if ChatGPT recommends your business

Checking whether ChatGPT recommends your business takes four steps: ask the buying questions your customers actually ask, in a signed-out or temporary chat so your own history does not colour the result; repeat each question several times, because answers vary from run to run; write down every business named and in what order; then read the sources cited underneath the answer, not only the prose.

What you are actually testing

An assistant does not keep a ranking table you can look up. Each time somebody asks it who to use, it builds an answer from what it learned in training and, where it browses, from what it finds on the web at that moment. There is no record of your position because there is no position.

So the only honest way to find out whether ChatGPT recommends your business is to ask it the way a customer would, and write down what comes back. That makes this a sampling exercise, not a lookup. One answer tells you almost nothing. Twenty answers to five questions tell you whether you appear at all, how often, and who is being named instead.

Step one: start from a clean chat

Your own logged-in account is the worst place to run this. ChatGPT may draw on saved memory, on custom instructions, and on the fact that you have spent months asking it about your own company. Of course it names you. You trained the conversation to.

Use a temporary chat — the option that keeps a conversation out of history and memory — or sign out entirely and use the free tier. If you want a second reading, run the same questions on a different device on a different network. You are trying to produce an answer that owes nothing to you.

Location matters as much as identity. Assistants infer where you are from your connection and from the words you use, and the shortlist changes accordingly. Town names collide across countries — Mansfield, Boston, Richmond, Perth and Hamilton all exist in several — so an unqualified town name regularly produces an answer about the wrong continent. Name the country every time. The same questions also behave differently market to market, which is covered in more detail in our guide to how AI answers differ by country.

Step two: ask the questions a customer would ask

Write five to ten questions that a person with a problem and no knowledge of you would type. Keep your business name out of every single one. These templates work for most service businesses:

Who are the best <trade> in <town, country>?
Which <trade> would you recommend near <town, country>?
I need <specific job> in <town, country>. Who should I contact?
Who are the most reliable <trade> for <specific need> in <region, country>?
What should I look for when choosing a <trade> in <town, country>, and who is good?

Filled in, that looks like: "Who are the best joiners in Mansfield, England?", "I need a bathroom fitted in Leeds, UK — which firms should I ask for quotes?", or "Who are the best commercial law firms in Austin, Texas for a small software company?" A retailer would ask "Where can I buy handmade ceramic gifts online that ship to Ireland?"

Include at least one question with no location at all if you sell nationally or online, and at least one comparison question naming two competitors but not you. Comparison questions are revealing: being absent from a list of two rivals is a sharper signal than being absent from an open question.

Step three: run each question several times

Open a fresh temporary chat for each run — not a follow-up message in the same thread, which carries context forward and contaminates the result. Run each question at least five times.

The repetition is not padding. Models are non-deterministic, and the shortlist genuinely moves between runs, sometimes substantially. A business named once in ten runs and a business named in nine of ten look identical if you only ask once. We go into the mechanics of that variation in why answers differ between runs.

Step four: record what you see

A spreadsheet with one row per run is enough. Columns: date, question, assistant, whether you were named, your position in the list, every other business named and in what order, the sources cited, and a free-text note for anything odd.

Position matters more than it looks. The first name in an answer carries most of the weight, because a person asking an assistant for a recommendation is usually looking to stop searching, not to compile a shortlist. Third of three is much closer to absent than to first.

The competitor column is the most useful thing on the sheet. After twenty runs you will have a tally of who the model reaches for in your category, which is a far better list of who you are actually competing with than the one in your head.

Step five: read the citations, not just the prose

When ChatGPT browses, it shows its sources. Open them. Most people read the paragraph and stop, and the paragraph is the least informative part of the answer.

What you usually find is that the answer was assembled from two or three directories, review platforms and "best X in Y" roundups rather than from the businesses' own websites. If the same roundup is cited in most runs and you are not on it, you have found a concrete, fixable reason you are absent. Which sources tend to carry this weight is the subject of what AI assistants actually read.

Note the date on each cited page too. A roundup from four years ago that still gets cited tells you something about how thin the current material in your category is.

What good looks like

A healthy result: you are named in a clear majority of runs across several differently worded questions, you appear in the first two positions more often than not, the description of you is accurate on the things that matter — services, area, speciality — and the cited sources include at least one page you do not control. A weak result is being named only when the question is phrased one particular way. An absent result is being named in none of them while the same three competitors appear in nearly all.

Also check what the answer says about you, not just that it says something. We regularly see businesses named accurately on trade but wrongly on area, or described using a service they dropped years ago. That is a content problem on readable pages, and it is usually quicker to fix than absence is.

The mistakes that make the exercise worthless

Asking a leading question. "Is <your business> a good joiner in Mansfield?" will nearly always get a polite yes. It proves nothing except that the model is agreeable. The same goes for "what do you know about <your business>?" — recall is not recommendation.
Running once and drawing a conclusion. One good answer is a coin landing heads.
Testing in your own account. Memory and history quietly rig the result in your favour.
Leaving the country off. You will get a confident answer about a town with the same name on another continent.
Counting any mention as a win. Being listed in a table of local firms at the end of an answer is not the same as being the business named in the first sentence.
Retesting an hour after a change. Pages have to be crawled before they can be cited, and training-data recall moves over months. Give a change weeks, not minutes.

Turning a snapshot into a trend

A single session tells you where you stand today. The value comes from repeating the same questions, worded identically, on a schedule — monthly for most businesses — so you can see movement. Put it in the calendar, because the discipline is the hard part, not the method.

How you repeat it is a matter of time and scale. A spreadsheet and a recurring reminder cost nothing and work perfectly well for one location and a handful of questions; the free approaches are genuinely usable. Writing your own script against the assistant APIs gives you consistency for the price of maintaining it. Products in this space, Nusan among others, exist mainly to remove the manual labour and keep the questions identical over time. The method does not change either way.

Whichever route you take, be clear about what the result is. Nobody — no tool, no agency, no platform — can see the conversations real customers had with ChatGPT. What you are measuring is how the model answers a consistent set of questions when you ask them. That is a directional index: it shows whether you are absent, whether you are improving, and who is named instead of you. It is not a transcript, and anyone selling it as one is overselling.