If someone has sold you "AI visibility" lately, you may have been shown a chart, a percentage, or a line saying your business appears in AI answers. The first question worth asking is what was actually measured. Usually the honest answer is: somebody typed a few questions into an assistant and wrote down what came back. That is not nothing. It is also not a ranking.
What a sampled observation is
A sampled observation is a record of one conversation with an AI assistant, on one day, about one question. You ask what a customer might ask — who does this near me — and you write down exactly what the assistant said, the date you asked, and the words it used. That is the whole artifact: a quote with a timestamp, kept word for word so nobody can smooth it into a better story later.
It is not an average, not a trend, and not a measurement of your market. It is a photograph of one answer.
Why a fixed question set is the only fair basis
If you ask a different question each time, you cannot compare anything. Two reports built on different questions are two unrelated photographs, and putting them side by side would suggest a change nobody measured.
Fixed wording matters because assistants are sensitive to phrasing. "Best dentist near me" and "who is a good dentist in my area" can produce different answers on the same afternoon. Change the question and you have changed the experiment.
So the fair basis is a set that never moves: the same five questions, in the same wording, about the same business in the same place, at the same point in each cycle, recorded the same way. Only then is a difference between two samples a difference in the same thing. It also protects you from the sales version: a vendor who picks the question after seeing the answer can make any business look visible or invisible.
What a single sample cannot tell you
- It cannot tell you your rank. Assistants do not publish an ordered list of everyone in your category. A sample can say your business was named in this answer, or that it was not. It cannot say you are fourth of twelve, because there is no twelve.
- It cannot tell you a share of anything. A share needs every competitor's name, in every answer, across a stable set, over many samples. One answer does not contain a market.
- It cannot tell you what your customers see. Two people can ask the same question and get different answers, depending on where they are, what they are signed into, and what they asked before. Your sample is one path through that.
- It cannot show a trend. Two samples a month apart are two data points, and assistants change underneath them. A difference might be your business, might be the model.
- It cannot prove cause. If you publish a page and the next sample mentions you, the page is a candidate explanation. Nothing in a sample demonstrates that a campaign worked.
- It cannot be scored honestly. Turning "named once" into a number out of 100 would be inventing precision the data does not have.
What a sample can do is narrower: show you the wording an assistant used about your category, and whether the public information you control is the kind that could be quoted at all.
What we actually do today
BravData's AI-search observation method is built and it is not switched on. Rather than describe that in our own words, here is the note the product itself prints wherever the method is mentioned:
The method is built: five fixed questions about your business and businesses like yours in your area, and when a run samples them it records the question, the date and what came back, quoted word for word. It is not switched on yet — no run asks an AI assistant anything today — so a report says plainly that we did not sample, never a zero and never a guess. When it does run it is a sample of what one assistant said on the days we looked: never a ranking, never a position, and not scored.
That paragraph is the entire claim. A plan that lists AI-search visibility observations is selling a designed method, not a measurement we have taken. A report from us says we did not sample, because we did not. We would rather show a gap than fill it with a number.
Questions worth asking any vendor
- Which questions were asked, word for word?
- Which assistant, on which date, and was it the same set last time?
- Can I read the answer as it was given, or only a score?
- If nothing was sampled, does the report say so plainly?
If the first two cannot be answered, the figure on the slide is not an observation.
What to do while none of this is running
You do not have to wait for sampling to act. The things an assistant would have to quote about you — what you do, where you are, when you are open, what you are called — are the same things you can check today on your own pages, without any tool. Getting those right is worth doing whether or not a sample ever mentions you.
