How to check whether AI assistants recommend your business.

First, the thing nobody selling this will tell you

There is no rank to track. A model generates its answer fresh each time, so the same question asked twice can return different businesses, different wording and different sources. Ask it tomorrow and the model may have been updated underneath you.

That means every number you produce here is a sample, not a measurement. You are estimating a probability by repetition, the way you would estimate a coin's bias by flipping it, and a single run tells you almost nothing. Anyone presenting you with a precise AI visibility score from one pass is describing something the technology cannot give them.

It is still worth doing, for one reason above all: the common failure is not being absent. It is being described wrongly, confidently, to somebody deciding whether to contact you.

Step one: write the questions a customer would actually ask

Not your keywords. Nobody types "technical seo services london" into ChatGPT. They type the sentence they would say to a knowledgeable friend. Write ten to fifteen of them, spread across three types:

• Recommendation: "Who should I hire to fix a WordPress site that got slow after a redesign?"
• Comparison: "What is the difference between hiring an SEO agency and a freelance consultant in the UK?"
• Verification: "Is [your business name] any good?" and "What does [your business name] do?"

That last category is the one people skip and the one that most often produces something alarming. It is also the category where a wrong answer does the most damage, because the person asking already knows your name.

Step two: run each question at least five times

Across the assistants that matter for your market: ChatGPT, Gemini, Perplexity, Copilot, and Google's AI Overviews via a normal search. Five runs per question per assistant is the minimum that distinguishes a real pattern from a fluke.

Two practical rules. Start a new conversation each time, because a model will lean on what it said earlier in the same thread and you will measure your own influence. And log out or use a private window, since personalisation and location will otherwise shape the answer in ways your prospect will not share.

Step three: record four things, not one

Presence is the least interesting result. For every run, note:

• Were you mentioned at all, in a question where you plausibly should be.
• How were you described, in the model's own words. Copy the sentence rather than summarising it.
• Who appeared alongside you, which tells you who the model considers your peer set, and is frequently not who you think.
• What was cited, where the assistant shows sources. This is the most actionable field on the sheet.

Step four: read the citations, because that is where the work is

When an assistant cites its sources, you are being shown which pages it trusts on that subject. Almost always, most of them are not your website. They are directories, industry publications, community threads and comparison articles.

That reframes the whole problem. If a model describes you using an out of date sentence from a directory listing, no amount of rewriting your own homepage fixes it. You have to correct the directory. This is the single most useful output of the exercise and it is why we treat this as adjacent to digital PR rather than to on site SEO.

Step five: decide what is actually wrong

Sort what you found into three buckets, because they need different responses:

• Wrong. Incorrect services, a closed location, an old price, a person who left. Fix at the source, which usually means a third party page rather than yours.
• Vague. You are mentioned but described in terms that would not persuade anybody. Usually means no page of yours states plainly what you do in a passage a model can lift.
• Absent. You do not appear for questions where you plausibly should. The slowest to fix and the least urgent of the three, despite feeling like the most urgent.

Fix wrong before vague, and vague before absent. Being described inaccurately costs you customers who were already interested. Being absent costs you customers who never knew.

What this method cannot tell you

It cannot tell you volume. There is no impressions figure for how many people asked an assistant about your category, and no provider currently publishes one, so anybody converting this into a traffic estimate is guessing.

It cannot tell you causation. If your description improves after you fixed a directory listing, that is suggestive, not proof. The model may simply have been retrained.

And it goes stale. A snapshot is worth taking quarterly rather than continuously, because the noise between two runs a week apart will usually exceed the signal.

Do it once, properly, before buying a tool

There is a fast growing market in AI visibility platforms, and some are genuinely useful at scale. What they mostly automate is the sampling above. Running it by hand once tells you whether the answer you get back is worth paying to monitor, and for a lot of businesses the honest finding is that the assistants rarely surface their category at all yet, which is a useful thing to learn for the cost of an hour rather than an annual licence.