A single AI answer feels like evidence, especially when someone sends it as a screenshot with your name in it, or without it. It is real, and it proves very little. AI answers are generated fresh each time a question is asked, so honest measurement means asking like a stranger, asking more than once, and reporting what the set of answers shows.
Ask like a stranger
When owners check their own visibility, they usually start with a branded question: the business name, followed by a prompt about quality or hours. That question hands the assistant its subject. Of course the answer talks about you. You put yourself in the question.
A branded answer tells you whether the machine can describe you when it is pointed at you. It tells you nothing about the moment that decides whether a new customer finds you: the moment the assistant builds a shortlist and nobody has said your name.
The unbranded question is the real market. A place near the station. Somewhere quiet for an anniversary. Somewhere that can handle a particular dietary need, a budget, or a late arrival. No names, no hints, just a need described the way a person describes it shortly before they start looking. The assistant answers from everything it has read, and your business is in the running only if it can be identified, described consistently, and supported by other sources.
Ask unbranded first, and branded second if at all. Ask in the words guests use, not in industry terms. Ask in every language your guests use, because the answers differ by language. And write down the exact questions, word for word. The wording is part of the result, and a result you cannot repeat is a story rather than a measurement.
Wording changes the outcome in ways that surprise most owners. In Nixuma’s first published study, the same tracked business was named in 1 of 29 general dinner answers, 3 of 30 anniversary answers, and 12 of 30 answers to a question about a guest who does not eat meat or fish. Same business, same day. The question changed, and the answer changed with it. A test that asks one question measures one doorway into your visibility, not your visibility as a whole.
One answer proves nothing
An assistant does not look your business up in a table and read out a row. It composes each answer on the spot, from sources it has read. Ask the same question twice and the machinery runs twice. Names move in and out, order shifts, and a fact stated firmly in one answer can soften in the next. The answer also shifts with wording, language, city and context.
None of this makes the answers useless. It makes them a distribution: a range of possible answers to the same question. A distribution cannot be photographed. It can only be sampled, several times, and described honestly.
That is why a screenshot, good or bad, should not drive a decision. It is one draw from a machine that would answer differently on the next draw. Treat it as proof, and you will make confident decisions on the strength of a coin toss.
Repeat the question
The basic discipline is to ask the same question, word for word, three separate times.
One answer is an anecdote. Two answers that agree might still be coincidence. Three begins to behave like a rate: named every time, named sometimes, named never. Those three verdicts mean very different things for what you should do next. More repetitions are better when the stakes justify the time. Three is the smallest number that separates a pattern from an anecdote.
Repetition also protects you from the opposite error: celebrating too early. A business named in three of three answers should still watch the rate over time rather than frame one good screenshot. The rate is the result. A single day’s answer is the weather on the day you looked.
What we measured. Research Note 001, measured November 1, 2026 and published November 2, 2026, asked AI assistants for recommendations in Ginza, Tokyo. The design used three scenarios, a general dinner, an anniversary dinner, and a dinner for a guest who eats neither meat nor fish, each written in Japanese and in English, each put to three AI systems five times. That gave 90 planned tests and 89 answers, of which 85 completed in full. One tracked business appeared in 1 of 29 general answers (3.4 percent), 3 of 30 anniversary answers (10 percent), and 12 of 30 vegetarian answers (40 percent). Across systems, it appeared in 10 of 45 Japanese answers and 6 of 44 English answers. Read the full note.
Counting needs care as well as repetition. In that same study, a first automated count put one business at 10 appearances. A person then re-read every answer by hand and found 8 real recommendations the first count had missed. The corrected total was 16, not 10. The misses hid in name variants, in unstructured lists where a recommendation does not announce itself, and in cautions and comparisons where a business is recommended in the middle of a warning about something else. An automated count looks for a pattern. A reader understands a sentence. Any visibility number produced without human checking should be read with that gap in mind.
Report ranges, not single scores
When repeated answers become a summary, report a range.
Named in two of three answers is an honest sentence. It admits the machine moves, it says how much, and it can be compared with next month’s two of three, or three of three. A range is a stronger claim than a single score, because it is the claim the evidence can carry.
Be wary of any score that pretends to be exact for a moving system. A figure stated to a decimal point claims a precision the underlying answers do not have. False precision does real damage: it makes a lucky day look like a strategy, a bad day look like a crisis, and it invites everyone to chase the decimal instead of fixing the business. When repeated scores swing by more than eight points, the honest statement is the range, with the number of runs beside it.
Reading movement then has strict rules. Same questions, same wording, same languages, compared across time. A move from named in none of three answers to named in two of three is a genuine change in odds, even though no single answer on any single day could prove it.
Show the denominator
Share of voice is the cleanest summary of presence, and the easiest to distort. The definition, with nothing hidden, is this.
Choose the unbranded questions that matter. Ask each one three times. Read every answer and count every business recommended, yours and everyone else’s. That total is the denominator. Count how many of those recommendations are you. That is the numerator. Share of voice is the second number divided by the first, and both numbers should travel with it, always.
A small example shows why. One question, asked three times, produces three answers containing 11 recommendations between them. Your business is named twice. Your share for that question is 2 of 11, about 18 percent. Stated as 2 of 11, anyone can check it, and you can recompute it next month with the same questions. Stated as 18 percent visibility on its own, it sounds like a grade and carries almost no information.
The denominator also guards against a subtler error. If assistants start answering with fewer names overall, your share can rise while your actual mentions stay flat. You did not become more visible. The room got smaller. Only the pair of numbers can tell those two stories apart. Answers that name nobody add nothing to the denominator, because no recommendation was made, and that rule should be stated when you report the metric.
Three corruptions are common. Drop the denominator and 2 of 11 becomes a bare percentage. Shrink the question set to questions you already win and the share climbs without a single new guest hearing your name. Count your mentions carefully and estimate everyone else’s, and the arithmetic becomes autobiography. In each case the numerator is real and the frame around it is rigged.
What honest measurement looks like
A report worth trusting shows three things.
Repetitions: the question was asked several times, and the count says how many. Ranges: where results moved, the movement is shown rather than averaged into one confident number. A human in the loop: a person read the answers and stands behind the count.
It also states its limits beside its numbers: the date, the area, the systems tested, the languages used, and whether answers came through an app or an API. Small studies can still be useful. They stop being useful when their limits are hidden.
That standard applies to Nixuma’s own work first. Our studies are small, we say so in them, and the method is published beside the numbers so anyone can run the same questions and check the arithmetic.
Key takeaways
- Ask unbranded questions first, in guest wording, in each language your guests use, and record the exact wording.
- The same business measured 3.4 percent in general answers and 40 percent in vegetarian answers in Research Note 001. One question cannot describe visibility.
- One answer is one sample from a changing distribution. Ask each question three times before drawing a conclusion.
- Report ranges with the number of runs beside them. Avoid single scores that claim precision the answers do not have.
- Share of voice means your recommendations divided by all recommendations counted. Always publish both numbers.
- Counting needs a human reader. In Research Note 001, hand checking lifted one count from 10 appearances to 16.
Get your baseline
Start with the free Quick Scan to check public search results. Use the Full Audit to record a baseline for what AI assistants say about your business. For the full method and limits behind this guide, read Research Note 001.
Check your business
See what public sources say about your business for free.
Run Quick ScanAbout Nixuma
NIXUMA is a GEO (Generative Engine Optimization) and AI-accuracy company. It measures what AI assistants — ChatGPT, Claude, Gemini, Perplexity and Google's AI — say about a business, checks those answers against the business's own confirmed facts, and publishes the measurements.