How to report AI visibility to your leadership team

The hardest part of this work is not collecting the data. It is presenting a non-deterministic number to people who are used to deterministic ones, without either overclaiming or hedging it into meaninglessness.

The sentence that fails

“We rank third in ChatGPT.”

It fails on the second check, because the answer will have changed. It also invites the right question, which you will not be able to answer: third out of what, measured how, over what period.

The sentence that works

Across 25 buyer questions and five AI answer surfaces, measured daily since 14 August, we were named in 38% of answers, up from 31% in the previous four-week period. The increase is concentrated in comparison-stage questions. Here are three answers where we appeared and two where a competitor did.

It states the question set, the surfaces, the period, the denominator, the direction, where the movement sits, and the evidence. Every predictable challenge is pre-answered, and the last clause turns the meeting from a debate about the number into a conversation about the answers.

Four rules

Report per surface, never blended. An average across surfaces implies they matter equally to your buyers, which is unmeasured, and it hides the mechanism: a brand at 60% on two surfaces and 20% on three needs completely different work from one at 40% everywhere. If leadership wants one number, give them one surface’s number and say why that surface.

Always carry the denominator. Percentages without a base invite exactly the challenge you cannot answer. “38% of 125 answered prompt and surface cells” is harder to argue with than “38%.”

Lead with the change, not the level. Nobody has a benchmark for what a good mention rate is, because the category is too young. What people can act on is direction: what moved, on which questions, since when.

Bring the raw answers. One screenshot of an answer naming a competitor instead of you does more to fund the work than any chart. It also demonstrates that the number is checkable, which is what makes the rest of the deck credible.

Say the limits before you are asked

Stating limits reads as competence, not weakness, and it protects you when somebody finds them later.

• These are the questions we chose. They represent the market; they do not enumerate it.
• Answers vary between runs. We compare completed runs over time, not single answers, and we do not report movement below a threshold.
• We measure visibility inside AI answers. We are not claiming attributed pipeline from it, and here is what we would need to do that.
• Where results are missing, that is recorded. An absent Google AI Overview is a real observation, not a failure.

Handling the three questions you will get

“Can we just get this to 100%?” No, and you would not want the cost. Some questions are ones a competitor should win. The goal is movement on the questions tied to revenue, not saturation.

“Why did it drop last month?” Check the shared basis first. If the two runs did not cover the same prompt and surface combinations, or the tracked competitor set changed, the drop may be an artifact rather than an event. Set-relative metrics like share of voice move whenever the competitor set changes, with nothing about your visibility changing at all.

“What is this worth?” Be honest: not yet directly attributable. What you can show is the count of buyer questions where a competitor is recommended and you are absent, which is a concrete list of moments where a purchase decision is being shaped without you. That framing has funded more of this work than any traffic model.

Cadence

Monthly. Weekly reporting on a non-deterministic metric guarantees you will narrate noise, and doing that once costs more credibility than a month of silence.