How to Choose Which Buyer Questions to Track in an AI Visibility Tool (and How Many Prompts Are Enough)

Track the questions a buyer asks before they know your name. That is the whole rule. A good starting set is 20 to 30 questions, weighted roughly half toward unbranded category and use-case queries, a fifth toward comparisons and alternatives, a fifth toward commercial intent like pricing and reviews, and only a handful toward your own brand name. Twenty-five well-chosen questions asked repeatedly over weeks beats two hundred asked once, because the signal you need is a trend line, not a snapshot.

Here is how to build that list from nothing.

Step 1: Write the ten questions your buyers ask before they know you exist

Open a blank doc. Write the ten questions a prospect types into ChatGPT in the two weeks before they ever hear your company name. Not the questions you wish they asked. The ones that end in a shortlist.

Mine them where buyers already use their own words. Discovery calls are the richest seam: the first real question on a discovery call is almost always the question that sent that person searching in the first place. Your support inbox and your churn notes come next, because people describe the problem there before they learn your product vocabulary. Then say your category out loud without naming your company. "Best X for Y" and "how do I fix Z" are the two shapes that reliably make an engine produce a list of vendors.

Keep the buyer's phrasing even when it is clumsy. "Software that stops our onboarding emails from going to spam" is a real prompt. "Deliverability optimization platform" is a category page title nobody types.

Step 2: Split the list across five buckets, not one

Most people fill their slots with brand prompts: "What is Acme?", "Is Acme legit?" It feels like monitoring. It is actually the one part of the funnel where you are guaranteed to get mentioned, so it teaches you nothing.

Allocate roughly like this for a 25-slot set:

• Category and best-of (10 to 12 slots). "Best onboarding software for fintech compliance teams." This is where you are invisible and where it hurts.
• Comparison and alternatives (5 slots). "Acme vs Competitor," "alternatives to Competitor." High purchase intent, and it shows you which companies the engines keep bundling you with.
• Problem and use-case (4 to 5 slots). "How do I cut B2B SaaS churn in the first 90 days." Lower volume, far easier to win with one good post.
• Commercial intent (2 to 3 slots). Pricing, reviews, implementation time, "is X worth it."
• Brand (2 slots). Enough to catch the day an engine starts describing you wrong.

Step 3: Phrase each question the way a real buyer would type it

An engine answers "best project management software" and "I run a 12-person design agency in Austin, what project management tool should we use" very differently. The second names fewer and more specific companies, which is exactly the answer your buyer sees.

So write the buyer into the prompt. Company size, industry, the constraint they actually care about, and the city if you sell locally. One line is enough. Dropping that line is the most common reason a tracked question comes back with a bland list of category giants and never your name, and the team then concludes AI search does not matter in their market. It did. The prompt was wrong.

Step 4: Pick a number, then stop optimizing the number

Twenty-five priority questions covers an early-growth B2B SaaS with one product and two or three buyer personas. Fifty covers a company with multiple product lines or genuinely distinct verticals. Past that you are mostly adding variants of questions you already track, and variants dilute attention instead of adding coverage.

The real answer to variance is not more questions. It is the same questions asked again and again. Generative engines re-rank their answers continuously, so a single run tells you close to nothing. Clarity monitors four engines daily on paid plans, ChatGPT, Claude, Gemini, and Perplexity, which is what turns one noisy answer into a mention rate and an average position you can trust. When the question pool runs low, Clark suggests new ones, and every suggestion has to clear a deterministic pre-screen and then a separate reviewer model that scores how realistically a person would type it and how likely the answer is to name providers. Anything under the bar is rejected with a reason. That screen is the thing that keeps a growing set from filling up with prompts no human would ever type.

Set your alerting thresholds in the same spirit. Clark emails on a first appearance in AI answers, a 15-point week-over-week swing in mention rate, or a 2-place move in average position, at most one alert per business every 6 days. Steal those numbers even if you are tracking in a spreadsheet. They are a workable definition of "a change worth acting on," and they stop you from re-reading a 25-question board every morning looking for movement that is just noise.

Step 5: Read mentions and citations as two different things

Being named in the body of an answer is not the same as being linked as a source. Track both. The source links matter more than most teams realize: when a competitor keeps getting named, the stored source list usually shows the same handful of third-party roundups and review pages sitting behind it. That is a concrete target list, not a vague instruction to go do PR. Watching competitor mentions alongside your own is what turns a prompt set from a scoreboard into a plan.

Step 6: Turn the zeros into your content calendar

A tracked question where no engine ever names you is not a failure. It is a brief. Write the answer to that exact question, publish it, then keep asking the same question word for word so you can measure whether the post changed the answer.

That word-for-word part is the whole trick, and it is why the loop is automated on our side: a gap becomes a priority topic with the original question wording preserved, so monitoring can measure the visibility change that specific post produced. The reverse matters too. Questions you already win, named in at least half the AI answers with a best position in the top 3, get handed to the writer as off-limits, so a publishing slot is never spent re-arguing a point you have already taken.

Everyone shows you rank. The point of a prompt set is to find out whether any of it pays.

If you want a baseline before you commit to a list, the free visibility check runs your domain against the questions your buyers ask and shows you which competitors get named instead of you. That is one snapshot. The list you just built is the trend line.