How to Check if AI Recommends Your Business

Claire Taylor

Almost every business owner we speak to has done a version of this. They have opened ChatGPT, typed their own company name, read something broadly flattering and concluded that things are fine.

That test tells you almost nothing. Asking an assistant about a business by name is like asking a librarian whether a book exists. Of course it can find you if you point at yourself. The question that matters is whether it would ever name you when somebody describes a problem and asks who could solve it.

This article gives you a proper method. It takes about an hour the first time and twenty minutes a month afterwards, and it produces something you can actually act on.

Why the Obvious Test Gives You a False Positive

Three reasons the “type your own name” check misleads people.

First, it measures knowledge rather than recommendation. An assistant knowing your business exists is a very low bar and tells you nothing about whether it would put you on a shortlist.

Second, it invites the assistant to be agreeable. These systems are built to be helpful about whatever you raise. Ask about your company and you will usually get a reasonable summary assembled from your own website, which is your marketing copy read back to you.

Third, it ignores the competitive question entirely. What you need to know is not whether you can be found. It is who gets named when you are not mentioned, and why them.

What Should I Actually Ask?

Ask the questions your buyers ask, in the language they use, without your company name anywhere in them. Four types are worth running every time.

The direct recommendation question

“Who are the best [what you do] in [where you serve]?” or “Which companies should I talk to about [the problem you solve]?” This is the closest thing to a buyer asking for a shortlist, and it is the one that matters most.

The problem-first question

Describe the situation rather than the service. “My company needs to replace a supplier who has gone into administration and we are based in the Midlands, who should I approach?” Buyers rarely know the name of what they need, and problem-framed questions retrieve differently from service-framed ones.

The comparison question

“What is the difference between [competitor A] and [competitor B]?” Then run it with your own name substituted in. If an assistant can compare two competitors fluently and becomes vague about you, that is a specificity problem and you have just located it.

The qualifying question

“What should I look for when choosing a [what you are]?” You may not be named at all, and that is fine. What you want is the criteria it lists, because that tells you what these systems consider important in your category, and it is often not what your website leads with.

Where Should I Run the Tests?

Four places, and the differences between them are informative rather than annoying.

Run each question with web search or browsing switched on, then again with it off where the tool allows. Answers with retrieval switched off draw on training data, which is a slower-moving picture of how established your business is. Answers with it on show the live picture. A business that appears with browsing on but never without it is newly visible rather than established, which is useful to know.

How Do I Stop the Results Being Noise?

This is where most self-testing falls apart. People run one question, get an odd answer, conclude the whole thing is random and stop.

Answers do vary. These are probabilistic systems retrieving live information, and they personalise. But the variation sits on top of a pattern, and the pattern is what you are looking for. Five rules make the difference between data and noise.

1. Use a fixed set of questions. Six to eight, written down, unchanged from month to month. The moment you improvise new questions you are measuring something different and calling it progress.
2. Use a logged-out or private session. Personalisation and chat history will skew results towards whatever you have discussed before, which for a business owner is their own company.
3. Run each question at least three times. Across different days if you can. Consistency across repeats is the actual signal.
4. Record the sources, not just the answer. Which domains were cited matters more than the prose. It tells you where authority in your category lives.
5. Never ask leading questions. “Is [your company] the best choice for X?” will produce a yes that means nothing.

What Should I Write Down?

A simple spreadsheet beats any tool for the first six months, because you are looking for movement over time rather than a score. One row per question per assistant per month, with these columns:

• Date and assistant. So you can see change and compare platforms.
• The question, verbatim. Unchanged each month.
• Were you named? Yes, no, or mentioned in passing. Keep it crude; nuance comes later.
• Who else was named, and in what order. Your real competitive set, which is frequently not who you thought.
• Which sources were cited. The single most useful column in the sheet.
• How you were described. Accurate, vague, or wrong. Being described incorrectly is a different problem from being absent and needs a different fix.

After three months that sheet will tell you more than any dashboard. After six, it becomes the evidence for whether the work is paying and where to push next.

How Do I Read the Results?

Five patterns and what each one means.

The last row is the one people undervalue. If an assistant struggles to answer a question in your field at all, that is an opening rather than a disappointment, and it will not stay open.

What Does Good Look Like?

There is no position one, so “good” needs defining differently. Reasonable targets for a small UK business after six months of deliberate work:

• Named in at least half of your direct recommendation questions, across at least two assistants.
• Described accurately when named, including the service you most want to sell.
• Your own pages appearing in the cited sources for at least some questions.
• AI referral traffic identifiable in your analytics, however small, and converting better than the site average.

That last point is the one to hold on to. Volume from this channel is usually modest. Quality is usually remarkable, and we cover why in why has my website traffic dropped in 2026.

Should I Use a Tool Instead?

Eventually, probably. Tools tracking AI visibility are appearing quickly and some are good. Two cautions.

First, do the manual version at least once before buying anything. Reading the actual answers, seeing which sources are cited, watching an assistant describe a competitor better than it describes you, is worth more than a score. You cannot get that from a number on a dashboard.

Second, be careful what a tool is measuring. Some track mentions of your brand name, which is the easy question rather than the important one. What you need is whether you are recommended in response to a buyer’s question, which is harder to automate and more expensive to run.

How We Do This at TU

We run a fixed question set for each client, across the four assistants, with and without live retrieval, repeated monthly. The output is not a score. It is a written record of who is being named, how the client is described when they appear, and which sources the answers lean on.

That source list is consistently the most valuable part. It shows where the conversation about an industry actually happens, and it is rarely where clients expect. We have had businesses discover their category is shaped by two trade publications and a forum they had never read, which changes where the next six months of effort should go.

Then we re-test the same questions every month and report movement, exactly as we would report rankings. Answers shift with model updates and with whatever was published recently, so a single snapshot tells you where you stand today and nothing about direction.

On whether any of it converts: one of our clients is a UK law firm in a specialist field. Around 30 percent of AI-referred visitors become enquiries, against 3.9 percent across their site as a whole. Roughly eight times. Seventeen enquiries from about fifty-six visits in a month. The caveat, stated plainly, is that analytics groups the assistants into one channel, so that covers AI referrals as a whole rather than a single tool.

If you would rather not run this yourself, we will do it for you. Our AI visibility snapshot covers what the assistants say about your business today, who they recommend instead and the specific signals behind the difference. You keep the findings whether or not you work with us.

Frequently Asked Questions

How often should I check?

Monthly is enough for most businesses. Weekly produces noise rather than insight, because the variation between individual answers is larger than the month-to-month trend you are trying to see.

Why does ChatGPT give my colleague a different answer?

Personalisation, chat history, live retrieval that changes, and the probabilistic nature of the models themselves. Use private sessions and repeat each question several times, and the underlying pattern becomes clear.

Should I ask the assistant why it did not recommend me?

You can, and the answer is worth reading, but treat it as a hypothesis rather than a fact. These systems generate plausible explanations for their own outputs rather than reporting their actual reasoning. Useful for ideas, unreliable as diagnosis.

Does it matter if I am named but listed last?

Less than it would in search results. Being in the answer at all puts you on a shortlist of three or four, which is a far shorter list than page one of Google. Order matters, inclusion matters more.

What if the assistant gets my business badly wrong?

Treat it as an entity problem. Wrong descriptions almost always trace back to outdated or contradictory information somewhere public. Find it, correct it at source and be patient, because these systems take time to reflect changes.

Can I see how much traffic AI is actually sending me?

Partly. Analytics platforms have started grouping AI referrals, though they currently bundle several assistants together, so you can see the channel but not always which tool. It is enough to judge whether the traffic converts, which is the question that matters.

Is there any point checking if I have done no optimisation work?

Yes, and that is the best time. You need the baseline before you change anything, otherwise you will never be able to prove that what you did afterwards worked.