It’s easy to build an AI visibility dashboard that looks impressive and says very little.
Load a hundred prompts. Run each one once on a few AI platforms. Count the brand mentions, turn them into a percentage and call it “AI share of voice”.
The hard question comes later, when a marketing manager asks: “What should we do differently because of this number?”
Usually the dashboard isn’t the problem. The prompt list doesn’t represent a clear business question, and the measurement doesn’t say what it can and can’t support. Good tracking needs both the right questions and a method that separates a real pattern from a noisy answer.
Take a company selling appointment software to clinics. The obvious thing to track is “best appointment software”. But a clinic manager comparing options is more likely to ask:
These sit at different stages of the same purchase. One names the problem, one tests requirements, one compares vendors. Track only variations of “best appointment software” and you’ll miss the questions where your product is most relevant.
| Intent family | Example | What you’re trying to learn |
|---|---|---|
| Problem | “How do clinics reduce no-shows?” | Is the brand present in category education? |
| Requirements | “Scheduling software with multiple doctors and WhatsApp reminders” | Do answers understand the relevant capabilities? |
| Comparison | “Alternatives to [competitor] for small clinics” | Is the brand in a meaningful shortlist? |
| Branded | “Does [your product] integrate with our clinic system?” | Are brand-specific descriptions accurate? |
You don’t need equal numbers in each group. A young brand may need more discovery questions; an established one may watch comparisons and inaccurate descriptions more closely. What matters is that every prompt has a reason to be there.
Start with what’s already inside the business: sales and demo notes, onboarding objections, support tickets, site search, Search Console, reviews, and the criteria buyers use in RFPs.
Then use prompt discovery tools to find related phrasings and missing questions. Treat their output as candidates to review. Third-party estimates of “AI prompt volume” aren’t verified query logs from the platforms, so if one drives a priority, label it as an estimate.
Give each candidate a score from 1 to 3 in four categories:
| Criterion | Score 1 | Score 3 |
|---|---|---|
| Buyer relevance | Loosely related to the category | Directly tied to a customer problem or decision |
| Business value | General awareness, little clear value | Tied to evaluation, trust or a commercial goal |
| Visibility gap | No meaningful problem found | Competitors appear while you’re absent or misdescribed |
| Fixability | No practical next step | A page, technical issue, fact or legitimate source you can improve |
It’s a planning tool, not a prediction of what an AI will cite. Then ask one more question: if this prompt changed next month, who on the team would act on it? If nobody, it can wait.
AI assistants don’t return the same recommendations every time. The same wording can give different lists, and a reworded question can change them even more.
A May 2026 preprint by Jack and colleagues tested commercial recommendations with thousands of same-prompt and paraphrase runs on selected OpenAI and Anthropic models (arXiv). It measured Jaccard similarity, the overlap between sets of recommended brands, where 1.0 means identical.
| Comparison in the study | Recommendation-set similarity |
|---|---|
| Same prompt, rerun | 0.50–0.61 |
| Cosmetic rewording, similar intent | 0.288 |
| Rewording that adds constraints | 0.135 |
These figures describe that study, not every assistant. The lesson still carries: exact phrasing matters, and one answer is a weak base for a trend claim.
Say your brand appears in 12 of 20 monitored responses. That’s 60% of your sample. It isn’t 60% of AI search in your category, because your prompts are a hand-picked list, not a random sample of everything buyers ask.
What repeated runs can tell you is how likely one prompt, under fixed conditions, is to name you. Rerun the same prompt with the same settings and each run is close to an independent draw, so a standard margin of error applies. Worst case, at 95% confidence:
| Runs of one prompt | Margin of error |
|---|---|
| 10 | ±31 points |
| 20 | ±22 points |
| 50 | ±14 points |
| 100 | ±10 points |
So 12 of 20 means somewhere around 38% to 82% for that prompt. That’s enough to spot a brand that’s missing everywhere. It isn’t enough to claim a 5-point gain since last month. Keep statements like this at the prompt or intent-family level, and say how you sampled.
More runs won’t fix a biased prompt list. Better coverage comes first.
For each important intent family, check:
Position can help when an answer gives an ordered list, but “first brand mentioned” isn’t a standard ranking across platforms. Name the metric for what it measures. And when a number improves after a content edit, check model changes, competitor updates and normal variation before crediting the edit.
The outcome should be a work plan, not just a prettier graph.
The goal isn’t the biggest prompt library. It’s a clear picture of where your brand belongs in a buyer’s decision, how AI assistants describe it today, and which gaps your team can realistically close.
Prompt Intelligence in ApexaRank.ai helps you discover relevant questions, monitor brand and competitor visibility, and connect what you find to actionable work. Start with ten questions from real customer conversations before you expand the list.
EXPLORE PROMPT INTELLIGENCESee where you stand, uncover visibility gaps, and discover what to improve next.
Get your first visibility score in minutes.