Whether ChatGPT can find your business is something you can check yourself in about ten minutes. For nothing. Without buying a tool. Open ChatGPT, run four prompts about your own business, and read what comes back against what you know to be true. One of four things will happen. It has never heard of you. It has confused you with a business of a similar name. It describes you accurately but two years out of date. Or it names you, correctly, in front of a buyer who asked. Each of those has a different fix, and only one of them is an emergency.
This piece is the check itself, written down: the exact prompts, what to look for in each answer, where ChatGPT is getting its information about you, and what to do about a bad result. It is the do-it-yourself version on purpose. We run the same check by hand for free, and that offer is at the foot of the page, but you should not need an agency to find out where you stand.
- Run the check in a temporary chat or logged out. ChatGPT personalises from memory, so your own account gives you the answer you trained rather than the answer a stranger gets.
- Being confused with another business is worse than being absent. A buyer who gets nothing back asks a second question. A buyer who gets a confident, wrong answer stops asking.
- Your own website being correct is necessary and not sufficient. Assistants read across a spread of sources and take the version that appears most consistently, so listings that disagree with you will win.
- Nobody outside OpenAI knows how ChatGPT ranks businesses, and any agency claiming otherwise is guessing. The honest work is influencing the inputs you can see, then running the check again.
- If your website does not convert the traffic it already has, fix that before any of this. Being named in an answer is not a sale.
Run these four prompts in this order
Before you start, open a temporary chat, or log out of your account entirely. ChatGPT personalises answers using memory and your past conversations, so an account that has discussed your business fifty times is the worst possible place to test whether a stranger can find it. You want the answer a customer gets, not the answer you have trained.
Replace the square brackets with your own details. The order matters, because each prompt narrows the question, and a business can pass the first and fail the third.
One. The name test
What is [business name]? What do they do, and who do they serve?
This is the lowest bar. You are not asking to be recommended. You are asking if your name maps to a real business at all. Add your town or your country if the name is not distinctive, because most good business names are shared by dozens of companies, and a model with no way to tell them apart will pick one.
Two. The category test
Who are the best [what you sell] for [type of customer] in [your town or country]?
This is the commercially important one, and it is the one most people skip. Nobody types your business name into an assistant unless they already know you, and a customer who already knows you was never the point. This prompt asks the question a buyer actually asks, which is a question about a category rather than about you. Note three things: whether you appear at all, where in the list you sit, and who is above you.
Three. The specifics test
Does [business name] do [a specific service you sell]? What does it cost, and how do I get in touch?
Buyers do not stop at a name. They ask follow-ups, and this is where a thin presence comes apart. An assistant that knows you exist but cannot say what you charge, what you actually do, or how to reach you will hedge. A hedge reads to the buyer as a reason to go and ask about somebody else. This prompt is also the one that catches the most embarrassing errors, because it forces specifics the model has to source from somewhere.
Four. The reputation test
What do customers say about [business name]? What are the common complaints?
Ask for the complaints explicitly. An assistant given a neutral question will usually produce a neutral summary, which tells you very little. Asked for the negative, it will surface whatever poor review, old forum thread or abandoned directory entry is carrying weight. That is worth knowing, because a cautious buyer asks the same question before spending real money.
Run each prompt twice, in two separate temporary chats, and write the answers down somewhere you will find them again. Two things make that worth the extra five minutes. Answers vary between runs on the same question, so a single run tells you less than you think. And a written record is the only way the re-run in three months means anything.
What each answer actually means
Four outcomes. They are not ranked in the order most people would guess.
Confusion is the one to fix first, and it is the one owners rank lowest, because being described wrongly does not feel as bad as being missing. It is worse than missing. A buyer who gets nothing back asks a second question. A buyer who gets a confident, detailed, wrong answer stops asking, and you never see the enquiry that did not arrive.
Stale answers almost always trace to a page you do not control. A directory entry somebody filled in once, a listing you inherited when you bought the business, an old press mention that is now the most confident-sounding thing about you on the internet. Your own website being correct is necessary and it is not sufficient, because the assistant is reading across a spread of sources and taking the version that turns up most consistently.
There is a fifth result worth naming, which is being mentioned in passing at the bottom of a list of eight. That is not the same as being recommended, and it is closer to absence than to success. Buyers act on the first two or three names in an answer, the same way they act on the first two or three results on a search page.
If an assistant cannot say what you do, the problem is rarely the assistant. It is that nothing on the open web says it plainly, in one place, in words a machine can lift.
Every price we charge, including the AI visibility line, is on the pricing page.Where ChatGPT gets what it says about you
Two different mechanisms produce the answers you just read, and telling them apart explains most of what you saw.
The first is training. The model absorbed a very large amount of public text at some point in the past, and part of what it says about you comes from there. That part is slow to change and you cannot edit it directly.
The second is search, which is the half you can influence. When ChatGPT looks something up while answering, it fetches live pages, and OpenAI names the crawlers that do it in its published bots documentation: OAI-SearchBot, described as used to surface websites in ChatGPT's search features; ChatGPT-User, which visits a page when a user's question prompts it; and GPTBot, which crawls content that may be used to train the models. If your site blocks those in robots.txt, you have opted out of the part of this you could otherwise affect.
Where the search half gets its candidates is less settled than most vendors imply. Damian Rollison, writing in Search Engine Land in May 2025, described ChatGPT running a Bing search for a local query, gathering twenty to thirty of the top results, and then organising them with its own logic, and noted that it does not see the profile data businesses manage inside Bing Places. That was careful reporting when it was published. It is also over a year old, the integrations have moved since, and OpenAI has never published a ranking system for any of it. What it does publish, all two sentences of it, is taken apart in how ChatGPT chooses which businesses to recommend.
So treat an agency that claims to know how ChatGPT ranks businesses the way you would treat one that claims to know Google's algorithm. Nobody outside OpenAI knows, the mechanism keeps changing, and the honest version of this work is to influence the inputs you can actually see, then measure the output by running the check again. If the proposal arrives wrapped in an acronym, GEO, AEO and AIO describe the same job.
What to fix, in the order that matters
Ordered by the cost of being wrong, not by effort.
One. Find out whether you are blocking the crawlers
Type your domain followed by /robots.txt into a browser. If GPTBot, OAI-SearchBot or ChatGPT-User appear under a Disallow rule, somebody made that decision: possibly a plugin default, possibly a developer being cautious a few years ago, possibly you. Whether to allow them is a real choice with a real trade-off, and it is yours to make. It should be a decision rather than an accident, and on most small business sites we look at, it is an accident.
Two. Say the plain thing on your own pages
An assistant summarising you can only work with sentences that exist. If your website never states, in ordinary words, what you sell, who you sell it to, which markets you serve and what it costs, the assistant will either hedge or borrow the answer from a third party who wrote it less carefully than you would have.
Most business websites fail this on price, because most businesses do not publish one. We publish ours for a commercial reason as much as an ethical one: a page with real numbers on it is quotable, and a page that says contact us for a quote is not. Ours is itemised in full on the pricing page. Why so few agencies will say a number out loud, and what the going rates actually are, is its own piece.
Three. Fix the sources that disagree with you
Directories, listings, review platforms, and anything else that carries your name and your details. This is unglamorous work, and it is the work that moves the answer, because consistency across independent sources is one of the few things a machine can actually verify about a business.
It is also not theoretical. A Dubai interiors store moved footfall 33% in 60 days on exactly this: listings corrected across the directories that feed the map, a steady review programme, and no new advertising spend. That was a local search engagement rather than an AI one, and the input list is very nearly the same, which is the point. How that number was counted is on the case study.
If you would rather not do it yourself, we run this as a $250 a month line inside the Grow retainer at $2,400 a month, and every price is published in full on the pricing page. You can also just have the check done properly: the free AI visibility check is a person asking four assistants about your business, recording what they say about you and your competitors, and sending back the three fixes that matter most, inside one working day, with nothing to sign.
What this check cannot tell you
A ten-minute check that oversells itself is worse than no check at all, so here are its four limits.
- It is a snapshot, not a measurement. Ask the same question twice and you can get two different lists. Treat one run as an anecdote and three runs as weak evidence.
- It tells you nothing about volume. You now know what the answer is. You do not know whether one buyer asked it this month or four hundred did, and nobody can currently tell you that with any confidence.
- It does not tell you whether being named produced money. Appearing in an answer is not a sale. The distance between the two is your website's problem rather than the assistant's, and how people actually decide once they arrive is a separate question.
- It covers one assistant. ChatGPT is the largest, at 900 million weekly users as OpenAI reported in February 2026. Perplexity, Copilot and Google's AI Overviews answer the same question differently, and they sometimes name different businesses.
The $250 a month AI search visibility line and the $2,400 a month Grow retainer are our published prices, not estimates, and the full list sits on the pricing page. The 900 million weekly ChatGPT users figure is OpenAI's own, reported by TechCrunch on 27 February 2026. The crawler names and their stated purposes are quoted from OpenAI's published bots documentation. The account of how ChatGPT assembles local results comes from Damian Rollison in Search Engine Land, 2 May 2025, and is flagged in the text as over a year old because the integrations have changed since. The 33% footfall figure is the measured result for a Dubai interiors store over 60 days, counted in store. The four-outcome framework is our own editorial structure, drawn from running this check for clients, not a survey.
When to leave this alone for now
Some businesses should run the check, note the result, and then go and do something else with the money. We would rather say that than sell a line item to somebody who does not need it yet.
- Your website does not convert the traffic it already has. Sending more people to a page that loses them is a more expensive mistake than being absent from an answer. Fix the page first.
- Your work comes from referral and repeat business, and your book is full. Being findable by strangers is worth something. It is worth less than a full book, and it is worth nothing at all if you cannot take the work.
- Your buyers demonstrably do not start with an assistant. Test that rather than assuming it, in both directions. Ask your last ten customers how they found you, and believe the answer even if it is inconvenient.
The reason to run the check anyway is that it costs ten minutes and it tells you which of those three you are. The dishonest version of this article would have told you the answer was urgent for everybody. It is urgent for some businesses, it is a note in the diary for most, and knowing which one you are is the whole value of the exercise.