How many prompts do I need for a credible first reading?

Enough to cover the range of questions your buyers actually ask, which in practice means tens rather than a handful. A set of 40 to 100 prompts across category, use case, comparison and objection questions gives you a denominator large enough that one odd answer does not move the figure much.

Should I check one model or several?

At least two, because a finding that only appears on one engine tells you about that engine rather than about your visibility. Seenby queries two providers at the MVP, OpenAI and Perplexity, and is built to keep reporting from the remaining provider if one becomes unavailable.

My score moved five points this week. Was that us?

You cannot know unless the model version behind each answer was recorded, because a provider update can shift phrasing and recommendations on its own. Seenby tags every stored response with its model version so model-update drift can be separated from real change, and flags it in the weekly digest where it applies.

What if we have no obvious competitors to name?

Then a share figure is close to meaningless, because share is a proportion of a competitive set. Name the three brands your buyers actually mention on sales calls, even if you consider them imperfect comparisons, and report the set you chose alongside the number.

What to tell your CMO about AI search visibility | Seenby, Share of AI Voice tracking

Your CMO forwards a screenshot: ChatGPT named three competitors and not you. They want something for Monday. The answer that holds up is one number with its sample attached. Tell them what you asked, how many prompts, on which answer engines, on which dates, what share of those answers named you against three named competitors, and what the number cannot yet tell them. A spot check answers "what did ChatGPT say when I asked on Tuesday afternoon". It does not answer "how does ChatGPT describe us". The distance between those two sentences is where marketers lose the room.

The minimum defensible sample you can assemble by Friday looks like this. Write 40 to 100 prompts a real buyer would actually type, spread across category questions, use case questions, comparison questions and objection questions. Run every prompt on at least two answer engines rather than one. Record the model version next to every answer. Count mentions of your brand and of three named competitors. Report your figure as a share of that set, with the sampling window said out loud in the same sentence. One pass of that set is a defensible baseline. It is not a trend, and week-over-week movement can reflect a model update rather than anything your team did. Say that before someone else asks.

The trigger moment, and the trap in it

The instinct is to open ChatGPT, retype the same question, and see what you get. You will get something different. Answers vary between sessions, between accounts, between model versions, and sometimes for no reason you can observe. If your reply to a screenshot is your own screenshot, you have started a screenshot war with your CMO, and you will lose it the first time someone in the room gets a third result on their phone.

The second instinct is worse: promise a full audit in two weeks. Two weeks is long enough for the model behind the answer to change underneath you, and your CMO already has the board date.

What a single spot check actually proves

It proves more than nothing, and you should say what it proves precisely:

That this answer is possible. One model, on one date, named those competitors in response to that phrasing.
That the phrasing matters. A prompt naming a category behaves differently from a prompt naming a price point or a company size.
That someone in your market is being described in terms you did not write.

Here is what it does not prove. It does not tell you how often you are named across the questions your buyers ask. It does not tell you where you sit relative to competitors when you are named. It does not separate a real shift in how you are described from a model version change. And it does not tell you whether the tone was favourable, dismissive, or simply absent.

The three follow-up questions that break a spot check

Prepare for these, because they arrive in order.

Did you ask more than once? If the answer is no, the finding is an anecdote. If the answer is "sixty prompts, twice, across two engines", it is a baseline.
What did the competitors get? A visibility number with no denominator is not a share. Share only means something against a named competitive set, which is why three named competitors is the floor, not an optional field.
Was that us or the model? This is the question that ends careers in analytics. If you did not record the model version beside each answer, you cannot answer it, and every future movement you report is contestable.

The wording that survives Monday

Here is the shape of the sentence, with illustrative arithmetic rather than any real result. Suppose your set is 60 prompts and you ran it across two engines, so 120 answers. You are named in 18 of them.

"Across 60 buyer-style prompts run on two answer engines on 24 and 25 September, we were named in 18 of 120 answers, about 15 percent, against 34 percent for the competitor in the screenshot. When we were named, we appeared third or later in two thirds of those answers. This is one sampled week on the model versions current this week. It is a baseline, not a trend, and part of any change we see next month will be model updates rather than us."

That paragraph does four things a screenshot cannot. It bounds the claim. It gives a competitor comparison. It names the date and the model context. And it tells your CMO what the number is not, which is the part that buys you credibility for every number after it.

Where the manual version stops working

You can do one week of this by hand. A junior can copy prompts into two engines and tally mentions in a spreadsheet by Friday. What breaks is the second week, and then the fourth, because the value is in the comparison and the comparison needs identical prompts, recorded model versions, and consistent counting rules. Hand tallies drift. Whoever ran week one leaves for a new job in week six.

That is the workflow Seenby is built for, and since we are pre-launch, here is exactly what it is built to do rather than what anyone has done with it. Onboarding takes a brand name, a description and up to three competitor names, and generates 50 to 100 prompt templates, with a curated fallback set if generation returns too few. A scheduled query engine hits two providers at the MVP, OpenAI and Perplexity, behind an adapter that degrades to the remaining provider if one is unavailable. A parser extracts brand and competitor mentions, classifies sentiment as positive, neutral, negative or absent, and records ranking position within each response. Every stored response is tagged with its model version, so real drift can be told apart from model-update drift, and passes a PII-scrubbing step before it is queryable. The result is a single Share of AI Voice figure per brand per week, stored with its methodology caveat attached to the record itself, surfaced on every score display and on a standing methodology page, with 30-day trend lines, a sentiment breakdown, the prompts where you are and are not named, and a competitor benchmarking view. A weekly digest is built to carry the top insights, drift alerts tied to model-version changes where they apply, and the same caveat.

What to promise your CMO, and what not to

Promise a baseline this week, a written prompt set anyone can rerun, and a date for the second reading. Promise that you will report model versions alongside every figure.

Do not promise a number that only goes up, and do not promise that the number is stable. Non-determinism is a property of the systems you are measuring, not a flaw in your method. The teams that keep executive trust on this are the ones that said so in the first meeting, not the fourth.