Which prompts to track for AI visibility, and where to find them

The questions you track decide everything downstream: the number you report, the gaps you see, the work you do. Track the wrong ones and you'll optimize, carefully and measurably, for answers nobody reads.

Start from the buyer, not the keyword list

A keyword is what someone types into Google. A question to an assistant is longer and more specific, because people talk to assistants the way they'd talk to a colleague: "what's a good exam prep course for someone working full time who failed once already?" rather than "exam prep course".

So a keyword export is a poor place to start. The better sources are where buyers already speak in full sentences:

• Sales calls and demos. The first question a prospect asks is usually the question they asked an assistant the night before.
• Support tickets and onboarding chats. These give you the fact questions: does it do X, does it work with Y, what does it cost.
• Search Console's long queries. Filter queries to ones with six or more words. Those are the questions people type in full, and they overlap heavily with what they ask assistants.
• Forums and communities. The subreddits, Slack groups and review-site Q&A your buyers read. Note how they phrase the problem before they know a product's name.
• Your rivals' comparison pages. They show which head-to-head questions the market already asks.

Split them into tiers

Each question belongs to one of three question tiers, and the tiers deserve different weight:

A healthy set leans toward tier 1: roughly half buying questions, a third preparing, the rest facts. Fact questions are easy to win and pleasant to report, which is exactly why a set made mostly of them flatters you.

Write them the way people ask

• Leave your brand out of buying questions. "Is Prepco good?" measures whether the assistant knows you, not whether it recommends you. Brand questions belong in tier 3, where the point is accuracy.
• Keep the buyer's situation in. "For a beginner", "for a team of five", "under a budget", "in Canada". Assistants answer those differently, and so do buyers.
• One question, one intent. "What is X and which tool is best for it" is two questions; split it.
• Don't track your own slogans. If nobody outside your company says "AI-native exam readiness platform", nobody asks an assistant about it.

Some kinds of business need their own mix: shops track product and budget questions (AEO for ecommerce), local firms put the place in every question (AEO for local business), and software companies add an alternatives question for each rival (AEO for B2B SaaS).

Throw out the phrasings nobody asks

Some keyword tools list dozens of long, oddly formal phrasings with healthy-looking search volume, such as "most reliable AI search optimization tool for data accuracy". Many of them look like AI trackers' own prompts being run through Google by software, not typed by people. A tell is impressions with no clicks in your own Search Console. Don't track them, and don't write pages for them.

How many questions, and how many answers

Twenty to thirty questions per product line is enough to see a pattern and few enough to read. Fewer than ten and one quirky question swings the whole number. Each question then needs several answers a week on each assistant, because prompt tracking samples a question rather than checking it once; our guide to measuring AI visibility explains why.

Keep some questions untouched

When you start doing the work, you'll aim each change at a few questions. Keep the rest of the set stable and unworked for a while: they're your control questions, and they're what lets you say a change worked rather than that the market moved. Changing the whole set every month destroys that comparison.

Review the set every quarter, not every week

Add questions when you launch something or buyers start asking something new. Retire questions that turned out to be off-target. Reword only when a phrasing is wrong, and expect a reworded question to start a new baseline. In Keats, an agent can suggest new questions from Search Console, but it asks before it retires or adds any, because the set is the yardstick.

Expand

Once the set is right, the next step is the work itself: our guide to getting named in ChatGPT takes it from the gaps the set reveals.