A useful AI share of voice report tells you which buyer questions competitors own and where to work next. Most reports make their first mistake before collecting an answer: they measure a shifting pile of prompts and call the result a trend.

Five ad-hoc questions run once. Branded searches mixed with discovery searches. ChatGPT and Perplexity averaged together as though they were the same place. The arithmetic can be perfect and the number still be useless.

AI share of voice only means something when you fix the prompt set, repeat the sampling and write down what counts as a mention. Change any of those and the score can move without your visibility improving at all.

I would rather have a small, repeatable measurement than an impressive dashboard built on moving sand. Here is the setup: 30 to 50 buyer prompts, at least five runs per prompt per assistant, and separate counts for mentions, citations and recommendations. Then I’ll show you how to ask for the results inside ChatGPT or Claude through groas, and turn the missing share into work.

What does AI share of voice actually measure?

AI share of voice measures your brand’s share of mentions across a defined group of brands in answers to a fixed prompt set. Report it per assistant before considering any combined view.

The basic calculation is:

AI share of voice = your brand mentions / all tracked-brand mentions × 100

Suppose ChatGPT answers 40 prompts five times each. Across those 200 answers, it produces 480 mentions of tracked brands, including 84 of yours. Your share of voice is 17.5%.

That is not the same as mention rate:

  • Mention rate: How often did you appear at all?
  • Share of voice: How much of the tracked brands’ presence belonged to you?

In that example, appearing in 84 of 200 answers gives you a 42% mention rate but only 17.5% share of voice. You show up fairly often. You also share the room with plenty of competitors.

Keep both numbers. Mention rate tells you whether you are present; share of voice tells you how crowded that presence is.

Step 1: Which buyer questions belong in the prompt set?

Start with buyer language, not your tagline. “Best scheduling software for dentists” belongs in a discovery test. Your positioning statement pasted into a question does not.

A useful starting mix for 30 to 50 prompts is roughly:

  • 35% problem-aware: “How do I reduce no-shows for a dental clinic?”
  • 40% solution-aware: “Best scheduling software for dentists.”
  • 25% vendor-aware: “YourBrand vs Competitor X.”

There is an important boundary here: keep branded prompts out of headline discovery SoV. Maintain the vendor-aware questions as a separate comparison and accuracy set. If someone types your name, appearing in the answer does not prove that an unfamiliar buyer would discover you.

For a smaller pilot, start with 5 to 8 category prompts, 3 to 5 problem prompts and 2 to 4 comparison prompts. Keep the branded comparisons separate, then expand the unbranded set toward 30 to 50 prompts.

Choose three to five competitors and freeze that list too. Your denominator is mentions of those tracked brands, including yours, not whichever companies happen to be convenient this month.

Same prompts. Same rivals. Same assistants. If you change the set, label the change rather than presenting the new score as a clean continuation.

Step 2: How many times should each prompt run?

Five runs per prompt per assistant is the floor for this method. Ten is better for open-ended prompts where the answers wander.

The reason is variance. Across 968 repeated runs, 42.9% of prompt-engine pairs switched between mentioning and not mentioning a brand at least once. A single answer is a sample, not a verdict.

For each measurement period:

  1. Run the fixed prompt set in each selected assistant.
  2. Use the same number of runs for every prompt.
  3. Keep assistant and mode separate in the results.
  4. Record the period and run count beside the score.

Do not quietly give some prompts ten chances and others one, then treat the pooled result as an equal-weight comparison.

Report ChatGPT separately from Perplexity, Gemini or any other assistant you track. Their answers and sources differ. A combined number can hide a gain in one place and a loss in another.

Skip the blended chart until you can explain every line underneath it.

Step 3: What counts as a mention, citation or recommendation?

These are three different signals. Counting them together makes a report look fuller while making it less useful.

SignalWhat to countWhat it tells you
MentionThe brand name in the answer text, once per brand per answerWhether the brand appears
CitationA linked source used in the answerWhich sources support the answer
RecommendationThe assistant directly suggests the brand as a pickWhether it endorses the brand rather than merely naming it

My counting rule is simple: one mention per brand per answer, however many times the name repeats. That keeps a long paragraph about one company from overpowering a shorter recommendation of another.

Citations need their own denominator. For citation share among competitors, divide links to your domain by links to all tracked brands’ own domains. Keep third-party sources, such as Reddit, YouTube and news sites, in a separate source analysis. They matter, but they answer a different question.

An answer naming no tracked brand contributes no mentions to the SoV denominator. Keep that answer in the dataset for mention rate and prompt coverage, though. “Nobody got named” is useful information, not a reason to delete the row.

Write these rules down before counting. Otherwise, the definition has a habit of improving when the score does not.

Step 4: Should you weight prompts or list positions?

Start with the basic formula. Then add weighting only when you can explain what it changes.

Position-weighted SoV gives a mention a weight of 1 / position: first place counts as 1.0, second as 0.5, third as roughly 0.33. Divide your weighted mentions by all tracked brands’ weighted mentions.

That distinguishes appearing first from appearing fifth. It does not replace the separate question of whether the assistant actually recommends you.

Prompt weighting addresses a different problem: not every question has equal value. A win on an obscure query should not conceal a loss on the question buyers keep asking.

Ahrefs uses AI-adjusted search volume to weight prompts. Whatever weighting you use, apply it to both your mentions and the denominator, and keep the rule fixed between periods.

Show unweighted SoV beside weighted SoV. One describes your presence across the test set; the other reflects the value assumptions you put into that set. Neither should masquerade as the other.

Skip weighting if it adds a layer you cannot explain. A transparent baseline beats mysterious precision.

Step 5: How do you read the results inside ChatGPT or Claude?

This is where I would rather ask a question than build another pivot table.

groas collects AI visibility data alongside competitor ads, funnels, Reddit discussions and SEO data. Your own AI can query that information in plain English. Connect groas to ChatGPT or Claude, then ask for the report in the shape your method requires.

Start with these requests:

Through groas, show our AI share of voice against our three main competitors for the available period, separately for each assistant. State the prompt set, run count and mention-counting rule. Flag anything missing.

Then:

Which prompts lost us the most share this month, and what do the winning answers cite? Separate mentions from direct recommendations.

The first request checks the measurement. The second finds the work. Do not accept a polished answer that cannot state its denominator. Asking an assistant for a method is not proof that the underlying data follows it.

For a closer look at recommendations, use the 3-pass procedure for separating recommendations from mentions on the same fixed prompts.

groas includes AI visibility and competitor data in Unlimited at $19 a month flat, with unlimited teammates, ad accounts and websites. The point is not another place to admire a chart. It is bringing the evidence into the assistant you already use.

Step 6: How do you separate a trend from noise?

Print the method beside the number, every time. At minimum, include:

  • Prompt-set size and whether it excludes branded queries.
  • Tracked competitors.
  • Assistant and mode.
  • Runs per prompt.
  • Measurement period.
  • Counting and weighting rules.

A score without those details is difficult to reproduce and easy to oversell.

Run the prompts weekly or daily and report a four-week rolling average. Keep the underlying per-assistant lines visible. If ChatGPT rises while Perplexity falls, a flat overall score describes neither story well.

I treat visibility as a sample, not a ranking. Look across eight weeks before treating movement as a reason for a major rewrite; a clearer trend may take about three months. That does not mean leaving an obvious factual error untouched while the calendar catches up.

The practical distinction is between fixing a known problem and rebuilding because a snapshot looked ugly. Do the first. Be slow about the second.

Step 7: Which lost prompts should you fix first?

An overall score tells you very little about the next edit. Split the result before touching a page.

Look at four things:

  • Mentions versus direct recommendations.
  • Position when the brand appears.
  • Citation share against tracked rivals.
  • Whether the answer describes the brand accurately.

Then split by prompt bucket. A report showing 12% overall share, 24% on best-tool prompts and 5% on automation prompts gives you a much clearer starting point than “visibility needs work.”

My order of fixes runs from cheapest to hardest:

  1. Improve the page that should answer the lost prompts. Put a clear answer in the first 100 words. State pricing or limits plainly. Use a comparison table with real numbers where it helps.
  2. Inspect the sources behind winning answers. If competitors appear because of a Reddit thread, a YouTube review and two listicles, your page edit alone does not address the whole gap.
  3. Check branded answers for accuracy. Keep this separate from discovery SoV, but do not ignore a wrong description just because your name appears.

Through groas, your AI can draft and ship website changes, with each change logged and reversible. That connects the report to implementation instead of ending with a PDF nobody opens again.

Ask for three to five fixes tied to specific lost prompts. “Create more content” is not a useful deliverable.

What should you pay for, and what should you skip?

A package worth paying for makes its coverage explicit. I want the engine list, prompt count, run count and counting rule in writing before I care about the dashboard.

I also want the report to connect:

  • SoV against named competitors.
  • Citation trends and the sources behind them.
  • Recommendation and accuracy checks.
  • Prompt-coverage gaps.
  • Specific content fixes tied to those gaps.

Skip packages that sell a score without showing the method. Skip the blended headline that cannot be broken down. And skip a second login if all it adds is another chart to check twice and forget.

That is why groas is my starting point for small teams: tracking, competitor context and the ability to act through your own assistant, for $19 a month flat. A standalone tracker has a harder case to make if it stops at telling you what is wrong.

Whatever you buy, ask the same question: how does a lost prompt become a page change?

What can no AI share of voice score tell you?

There is no industry-standard AI visibility score. Different prompt sets, sampling schedules, assistants and weighting rules can produce different numbers without an arithmetic error.

That is why 34% in one tool and 51% in another is not a useful comparison by itself. First ask what each tool measured.

Mode matters too. An answer can mention your brand without linking to it, particularly when browsing is off. A citation score cannot stand in for all visibility, and a mention cannot stand in for a recommendation.

Compare against named competitors and your own consistent baseline, not an unrelated vendor score. No tool can promise a particular share and make the uncertainty disappear.

The method does not eliminate variation. It makes variation visible enough that you can stop mistaking it for progress.

Which terms should stay on the report?

Keep these definitions beside the score so nobody quietly changes the question halfway through the quarter.

TermMeaning in this method
MentionA brand name in the answer text, counted once per brand per answer
CitationA linked source used in the answer
Mention rateShare of all sampled answers in which your brand appears
Share of voiceYour mentions divided by all tracked-brand mentions, per assistant
Position-weighted SoVThe same division after weighting each mention by 1 / position
RecommendationThe assistant directly suggests the brand, rather than merely naming it

What should you do today?

Write down 30 unbranded buyer prompts and name three competitors. That is the first job, before buying a chart or arguing about a percentage.

Run each prompt five times in ChatGPT and five times in Perplexity. Count each brand once per answer. Keep branded accuracy checks separate, and put the method on the same page as the score.

For the longer walkthrough of what to ask and what to fix first, read that next. Then connect groas to the assistant you already use, use it to follow your visibility, and work through the prompts you lost. The number earns its place only when it changes what you do.