Your AI share of voice can rise from 18% to 26% without a single new recommendation. Change the questions, competitors, or platforms in the report, and the percentage can improve while the answers buyers see stay exactly the same.

That is the problem I want settled before anyone puts the chart in a budget presentation. AI share of voice measures a selected sample, not a census of buyer conversations. The selection is part of the result.

To see why, start with a business tracking 40 buyer questions. I’ll use that same panel throughout: first to build the fraction, then to separate an honest improvement from a better-looking report, and finally to decide what work deserves funding.

Build the fraction before trusting the chart

Say you track those 40 questions across ChatGPT, Gemini, Perplexity, and AI Overviews, sampled once each. That gives you 160 answers.

Now suppose 48 answers mention your business. There are at least two ways to turn those appearances into a headline:

  • Mention rate: 48 ÷ 160 = 30%. This measures the proportion of sampled answers that mention you.
  • Mention share: 48 ÷ 190 = about 25%. If the answers contain 190 tracked-brand mentions in total, this measures your portion of those mentions, counting each brand once per answer.

Same appearances. Different denominators. Different headlines.

An answer can name several vendors, which is why a high visibility rate can coexist with a much lower share of mentions. Neither calculation is inherently dishonest. Calling both “AI share of voice” without showing the formula is where the trouble starts.

There is another denominator to watch: answers that name at least one tracked brand, rather than all answers sampled.

In one test of 24 buyer prompts across four engines, 52 of 96 answers named no tracked vendor. Exclude those 52 answers and the denominator falls from 96 to 44. Without earning another appearance, a brand’s percentage becomes about 2.18 times larger.

The narrower calculation describes your presence among answers that already name a tracked vendor. The all-answer calculation describes how often your name appears across the whole sample. Those are different questions, and both need labels.

Before comparing scores, ask what sits below the division sign.

A fixed question still produces a variable answer

Freezing the formula does not freeze the answers. The same prompt asked again can produce a different brand list.

In the SparkToro test, 600 volunteers ran 12 prompts 2,961 times and found less than a one-in-100 chance that two answers to the same prompt named the same brand list. That does not make measurement pointless. It makes a single answer a poor substitute for repeated observation.

Return to the 40-question panel. If you check it once in June and once in July, some of the movement may reflect sampling variation rather than anything your business changed. Switching engines, locations, or wording adds more moving parts.

One screenshot is weather. Repeated runs are climate. The analogy stops there: repeated sampling does not turn a handpicked question list into the whole market.

My practical starting point is to run each prompt three times per check and report each engine separately. Three runs do not magically establish statistical certainty. They do stop one favorable answer from being the entire story.

Small panels also limit what you can conclude about small changes. A 30- or 40-question tracker can help identify gaps and follow a consistent sample. It should not automatically referee a monthly retainer because the headline moved a few points.

Keep the conditions consistent, repeat the sampling, and show the variation. A precise-looking percentage is not the same thing as a precise measurement.

The question list decides what the score means

The next layer is selection. Even a perfectly calculated, repeatedly sampled score answers only the questions you chose to ask.

Start with branded prompts. Asking an AI about your business and then counting its mention of your business as evidence of discovery mixes two different jobs. One checks what the model says about you. The other checks whether it introduces you when the buyer has not supplied your name.

Both matter. They do not belong in one unlabeled average.

In our 40-question panel, I would keep branded questions separate and make the unbranded buyer-intent result the headline. Otherwise, a healthier chart can conceal unchanged discovery.

Question type matters, too. In that same 24-prompt test, vendor-seeking prompts were 9% free of tracked-brand mentions, problem-aware prompts 63%, and conceptual prompts 91%. Remove the conceptual questions and you remove the category where answers are least likely to name a tracked vendor.

The score gets easier. The business has not necessarily become more visible.

For a plumbing business, compare these questions:

  • “Best emergency plumber in Austin” asks for a provider.
  • “Why does my water heater rumble?” asks for an explanation.

The second answer can do its job without naming a plumber. Its absence from that answer is not automatically a failure. But deleting the question also means you stop observing how AI explains a problem that could lead someone toward a purchase.

Nor does appearing in a conceptual answer prove you will make a later shortlist. The panel measures the answers you sampled, not a buyer journey you never observed.

Keep discovery, problem education, and branded questions distinct. Do not let an easier mix masquerade as better performance.

Make providers show the panel, not just the percentage

Two providers can give you different scores without either making an arithmetic error.

One divides your appearances by all answers. Another divides your mentions by all tracked-brand mentions. One tracks ChatGPT and Perplexity. Another includes Gemini and AI Overviews, where your coverage may be thinner. One uses unbranded buyer questions. Another includes questions built around your name.

The headline hides all of that.

Competitor selection matters as well. If mention share uses tracked-brand mentions as its denominator, removing a frequently mentioned competitor can raise your share without changing your own appearances. Changing prompts, competitors, or platforms can change the reported share without improving the underlying answers.

You are buying someone’s sample. I want to see the sample before treating it as evidence.

Before signing, ask for these four items in writing:

  1. The full question list, its categories, and the date it was frozen.
  2. The exact formula, including what counts as an answer, mention, and denominator.
  3. The sampling conditions: engines, locations, and runs per prompt.
  4. The treatment of branded questions, including whether they affect the headline.

The gap between mention rate and mention share is the first thing to check. Then check what the provider leaves out.

A provider does not need to measure every possible question. It does need to explain the boundaries of the measurement it sells you.

Freeze the baseline; put new questions in a separate panel

Once the measurement is explicit, the rule is simple: freeze the panel, then let the answers move.

For our example, keep the same 40 questions for a quarter. Record the engines, locations, denominator, and sampling procedure. Run each prompt three times per check, and report results per engine rather than blending everything into one feel-good average.

Then keep two panels:

  • The frozen panel tracks the same questions under the same measurement rules.
  • The expansion panel tracks newly added questions separately.

Suppose new pages go live in July and the team wants to add 15 questions. Fine. Add them to the expansion panel. Do not fold them into June’s baseline and call the resulting change growth.

The frozen panel shows whether measured visibility changed on the original questions. The expansion panel shows your visibility on the new questions. Neither, by itself, establishes that the questions represent market demand.

Next, report the spread rather than selecting the best run. A result of 22% across three runs, ranging from 18% to 26%, tells a different story from a slide displaying only 26%.

Keep two or three competitors in view as context, too. If everyone moves together, that is a reason to investigate a broader engine shift rather than immediately credit your provider. Competitors are useful context, not a perfect experimental control.

Finally, version the panel like code. Date every change and record why someone added or removed a question. If July’s lift comes from deleting questions you lost on, the change log should make that obvious before renewal day.

The baseline can evolve. It just cannot quietly rewrite its own history.

Turn missing mentions into investigations, then shipped work

A stable panel makes the report defensible. It does not make the service useful on its own.

Suppose 14 of the 40 frozen questions never mention your business. That is a place to investigate, not proof of 14 identical content failures. A missing mention might point toward a page that does not answer the question, a crawler-access problem, or weak third-party evidence. It might also be a question whose answer does not need a vendor.

The measurement should lead to a work decision, not just another chart.

For a relevant question such as “how much does commercial dryer vent cleaning cost,” I want the team to inspect what the business actually provides:

  • Is there a page that answers the pricing question directly?
  • Can the model fetch and read it?
  • Does it contain a useful service or specification table?
  • Is there independent evidence supporting the business’s claims?

Before paying for more tracking, make sure the pages are worth citing: accessible HTML, a direct answer in the first 100 words, and a named proof source where the claim needs one.

That is why I want implementation in the same scope as measurement. The team should connect a missing answer to an investigation, connect the investigation to a justified fix, and record what shipped. Publishing a page is not proof that an AI will recommend it next Tuesday. It is work the business can inspect.

This is also the standard I would put in a brief for groas Done For You, the offer I recommend for businesses that want to hand the work over. Keep SEO and AI visibility together: a stable prompt panel alongside page fixes, technical access work, and Reddit and competitor research that inform the implementation.

No placement promise. No claim that a team controls how ChatGPT or Gemini phrases an answer. Demand a stable measurement panel, a record of changes and their reasons, and shipped work tied to the gaps being investigated.

If a package delivers the measurement without the implementation, you bought a thermometer. Buy the work that changes the temperature.

Take the model into the next budget meeting

Now reason through a new report using the same rules.

Suppose the frozen 40-question panel moves from 22% to 24% across three runs per engine, with results ranging from 21% to 27%. The 15 questions added in July score 41%. Competitors move the same two points as your business.

The attractive version blends in the new questions, highlights the best run, and announces momentum.

The defensible version sounds like this:

Our unbranded mention rate on the frozen 40-question panel moved from 22% to 24%. Results across runs ranged from 21% to 27%. The 15 new questions scored 41% and are reported separately. Nothing was removed, the denominator stayed fixed, and competitors moved by the same two points. I would not call this a demonstrated win yet.

That explanation does not bury the work. It separates the work from a claim the measurement cannot support.

The next discussion is concrete: what happened with the 14 questions where the business still does not appear? Which gaps justified fixes? Which pages shipped? Can the models fetch them? Pipeline remains unproven; a visibility percentage does not settle it.

A defensible retainer connects a stable measurement to inspectable implementation. Neither substitutes for the other.

Run the next pitch through that model. An agency offers to manage how you appear across ChatGPT, Gemini, and Perplexity. Ask which questions, frozen when, using which denominator, sampled how often, with branded questions separated or blended in.

Specific answers should lead to a report that is boring in the best way: honest movement on a fixed panel, plus shipped work tied to the questions you lose. That is something you can defend in a budget meeting.

Anything that promises a rising percentage without naming the denominator is selling you a better fraction, not more buyers seeing your name.