You do not need a tool to get a first reading on this. You need an afternoon, a spreadsheet, and the discipline to run the test in a way that can actually produce bad news. Most self-checks fail on that last point.
Here is the method, and then the four ways people accidentally rig it.
The method
1. Write down the questions a customer would ask. Not questions about your business — questions about their problem. A customer who has never heard of you asks things like who does emergency boiler repair in Sheffield, is it safe to get dental implants in Mexico, what's the best boutique hotel in Tulum for couples, how much should I expect to pay for X. Aim for at least fifteen, and include the early, cautious ones asked before anyone has settled on a shortlist.
2. Ask each one in all four assistants. ChatGPT, Claude, Gemini, and Perplexity. They draw on different data and behave differently. Being absent from one means little; being absent from all four is a finding.
3. Repeat each question several times, in fresh chats. Answers vary between runs. One response is a sample, not a measurement.
4. Record four things per run. Were you named? Who else was named, and in what order? Which sources were cited, where citations were shown? And was anything said about you factually wrong?
5. Count. For each question, the rate at which you appeared. The competitors who appeared most often. And the sources that recurred across the whole set — that last list is the most valuable thing the exercise produces.
The four mistakes
Asking about yourself. By far the most common, and it invalidates everything. Tell me about Cedar & Stone Roofing will produce a competent summary of your business and a warm feeling. It tests whether the assistant can read your website. Your customers are not typing your name; that is precisely the problem you are trying to measure.
Testing in a chat that knows you. Assistants with memory or personalisation carry context between conversations. If you have discussed your business, asked it to write your marketing copy, or mentioned where you work, it is more likely to volunteer your name — and you will read that as visibility. Turn memory off, start fresh, and do not preface the question with anything.
Asking once and stopping. Owners tend to stop at the first result that matches what they already suspected. If you appear in the first run, you stop and feel fine; if you do not, you stop and feel bad. Both are conclusions drawn from one sample of a system that varies substantially between runs.
Only asking the questions you'd like to win. There is a strong pull toward the phrasing that flatters you — the niche specialism, the exact service you are proudest of. Customers do not know your specialism yet. The questions worth testing are the broad, early ones where you would rather not discover the answer.
Reading the result honestly
Three findings tend to come out of this, and each has a different meaning.
You are absent everywhere. Deflating, and the most actionable outcome of the three, because the cause is nearly always the same: you are not present in the sources the assistants lean on. That is a finite list, and you now have it, because you recorded the citations.
You appear sometimes. Better than it feels. You are in the material but not the strongest candidate. Look specifically at which competitor takes your place on the runs you lose, and which source is carrying them.
You appear, but something is wrong. An old price, a service you no longer offer, a closed location, a credential attributed to you that is not yours, or a confident claim that you do not do the thing you primarily do. This is more urgent than absence. A customer reading an incorrect answer does not check; they cross you off.
What a manual test cannot give you
Two things, and both matter.
Coverage. Fifteen questions × four assistants × five runs is three hundred conversations, and the manual version of that is a full week that most owners will abandon on day two. What actually happens is a much smaller sample and a conclusion drawn too early.
A record. An answer read in a browser window and half-remembered is not evidence. It cannot be compared against next quarter's answer, shown to anyone, or used to tell whether something you changed made any difference. Screenshot as you go, with dates. The value of this exercise multiplies when you can run it again in three months and see what moved.
If the manual version tells you nothing else, it will tell you whether you have a problem worth taking seriously. That is a reasonable afternoon's work.