Add as preferred source on GoogleBuild an AI Search prompt set from sales calls, support tickets and a stakeholder review, so your visibility numbers describe the buyers you actually sell to.
In client work, the useful question about a draft prompt set is the buyer one: are our buyers far enough along to type these category terms, or are they a step behind, asking in plainer words? Most teams cannot answer it in the room.
Building a prompt set forces that question, because a prompt set is a model of your buyer. Most teams build it from the wrong material: keyword lists, tool-generated suggestions, the titles of pages they already published. Then they read visibility numbers that describe a buyer who does not exist.
This guide covers how we design prompt sets: where the prompts come from, what each one needs to contain, how close to the buying decision they should sit, and who should review them before tracking starts. The mechanics of setting up tracking, topics and tags live in our prompt tracking guide, so we link there instead of repeating it.
Key takeaways
- Build prompts from what buyers say in sales calls and support tickets. Buying decisions happen in prompts with zero recorded search volume, so a keyword-derived prompt set measures a market that does not exist.
- Ask for a recommendation, or the answer rarely names a brand. A generic "what is email marketing" prompt returns a category explainer; a recommendation ask returns names, and those prompts form the core of the set.
- Prompts model buying situations. Nobody types your 180-character prompt word for word, and the tracking tool runs it as written: the persona, the pain point and the recommendation ask have to be right.
- Track plain-language and expert phrasings side by side. When experts disagree on buyer maturity, track both registers: visibility shows where you stand in each, and sales calls show which one buyers actually use.
- Have the people closest to buyers review the set before launch. Sales and unit leads catch prompts no real prospect would ask and settle the competitor set; their sign-off is a design step.
Why keyword-derived prompt sets measure the wrong market
Most of the advice on prompt tracking starts from your keyword list: turn your SEO keywords into prompt variants, or let a tool generate prompts from your brand name, your keyword set or your URL. It fails for a structural reason. Prompts are conversations with context, a persona, a company, and a problem inside them. Keywords are context-free strings.
Where the prompts should come from
Keyword volume is a broken gate on an editorial calendar, and the same holds for a prompt set. Buying decisions happen in prompts with zero recorded search volume. If you filter your prompt set through a keyword tool first, you cut the exact prompts where your buyers decide.
The fix is a different starting material. We build prompt sets from the audience and the CRM: sales calls, won and lost deals, support tickets, customer interviews, and the long-tail queries buyers already type into Google. The full build method, including where each input comes from, is in our prompt tracking guide.
What a prompt that reflects a real buyer contains
A prompt that produces usable visibility data has four parts: a job title, a company size or type, a specific pain point, and a question that asks for a recommendation. Across our client setups we write prompts between 180 and 210 characters, which is our own practice rather than a study result. That length is enough room for all four parts without turning into an essay.
What each part does
The persona and company context steer AI Search interfaces like ChatGPT toward situation-specific sources instead of generic encyclopedic ones. The pain point, written the way a buyer would describe it out loud, is what matches the fan-out queries an engine runs on its own. The recommendation ask is what produces brand mentions: ChatGPT or Perplexity asked to recommend will name tools; asked to explain a category, they rarely do. How to earn those mentions is its own topic, covered in our guide to getting cited in AI Overviews and ChatGPT.
The difference is easiest to see in a pair:
- Keyword-style: "email marketing automation tools"
- Buyer-modeled: "I run marketing at a 40-person B2B SaaS company and our nurture emails barely get opened. Which email marketing tools should I look at for a team without a dedicated email person?"
The first is a keyword wearing a prompt costume. The second is a buying situation, and it is the kind of prompt where an answer either names you or names a competitor.
Keyword-derived vs buyer-modeled prompts
| Keyword-derived prompt | Buyer-modeled prompt | |
|---|---|---|
| What it contains | A topic, sometimes with a modifier | Persona, company context, specific pain, recommendation ask |
| What the AI returns | A category explainer, often sourceless | A recommendation with named tools and cited sources |
| What you can measure | Whether you exist for a topic | Whether you get recommended to your actual buyer, and who wins when you do not |
Prompts written around content titles
A subtler failure mode than keyword recycling: prompts written around existing content titles. A customer does not ask for a "go-to-market orchestration platform". They ask how to stop sales and marketing working in silos. If your prompt set reads like your sitemap, you are tracking your own vocabulary, and the answers will tell you how well you rank for words nobody types.
How close to the money should prompts sit?
Start with category prompts: "best X for Y", "alternatives to X", pricing comparisons, head-to-heads. In our client work, generic informational prompts almost never produce brand mentions, so a set made of them produces a visibility number that cannot move. The closer a prompt sits to a buying decision, the more the answer is a list of names, which is exactly the surface you want to measure. When you lose those head-to-heads, why ChatGPT recommends a smaller competitor is the question to read next.
The harder question is register. In client work we keep meeting teams who cannot say whether their buyers search in expert terms or in plainer ones. A B2B software buyer might ask about a "revenue operations orchestration platform", or they might ask how to stop sales and marketing working in silos. Both describe the same buyer at different stages of maturity.
Track both. Split the set across plain-language and expert phrasings and tag them separately: the visibility data shows where you stand in each register. Which register your buyers actually use is a different question, and it gets answered where buyer language lives: sales calls, support tickets, and the long-tail queries in Search Console. When your own experts disagree on buyer maturity, tracking both registers is the honest alternative to guessing.
But nobody types that prompt word for word
Stakeholders raise this objection often, and it misreads what tracking measures. Your 180-character prompt will rarely appear in a real user's chat history verbatim, and that's fine: the tracking tool runs your prompt as written. The prompt stands in for a buying situation, so what has to be right is the persona, the pain point, and the recommendation ask. The prompt stands in for every buyer in that situation, the way a persona stands in for every customer who fits it.
Spend your effort on the situation rather than the phrasing.
What else stays in the set
Recommendation prompts are the core, and the mix we run across client setups is 70 to 80% long context-rich prompts, 10 to 20% short-tail, and 10 to 20% brand. Short-tail prompts answer a different question: whether you count as a category player at all. "Best CRM for small business" with no context attached is worth tracking, because a brand that never appears in the answer to the category question has a problem no long-tail prompt will surface.
Brand prompts get their own treatment. In our own Peec project, brand prompts mentioned us in 99.7% of tracked responses between 12 July and 11 August 2026, and every other prompt in 13.3%. Averaging the two produces a headline number that flatters you and moves for the wrong reasons. Give brand prompts their own topic, keep them out of the headline visibility figure, and read them for sentiment and positioning instead.
Why is the stakeholder review part of the design?
Before a prompt set goes into tracking, walk it past the people who talk to buyers every week: sales leads, unit heads, support. Their reactions are design input. "No prospect would ask that" kills a prompt. "Our buyers are two steps behind that wording" rewrites one. Both are exactly the corrections a keyword tool cannot give you, because they come from memory of actual conversations.
A lightweight loop from our client work: send the draft set to each unit, give a deadline, collect the corrections, then lock the set and start the baseline. Without a deadline the review drifts, and tracking never starts.
The review surfaces more than wording. It settles category boundaries, which "best X" bucket the buyer actually shops in, and it settles the competitor set, which names belong in the comparison prompts. In our client work it usually surfaces a pain point sales hears constantly and marketing has never written down.
How the prompt set and the content roadmap feed each other
The prompt set and the content plan develop together, in both directions. Prompts come from onboarding documents, expert interviews, sales-call language, Search Console long-tail and competitor gaps, and from the roadmap's planned topics. The roadmap then gets validated by which prompts lose, the loop we describe in our guide to building a content plan connected to pipeline.
Losing prompts become the work list. For each prompt you lose, read who wins it and which sources the engine cites, then decide where the fix lives: a page on your site, a platform like YouTube or a review site, or a third-party list you should be on. The off-site half of that decision has its own guide in AI visibility off-page SEO, and the YouTube route is mapped in our guide to getting cited from YouTube. The reading method, prompt by prompt, is in our prompt tracking guide.
The sequence around launch
- Set up tracking 7 to 14 days before your first report, so a baseline exists before anyone draws conclusions.
- Put a kick-off review in the calendar two to four weeks after go-live.
- In that review, cut prompts that come back generic or sourceless even though they tested fine, fix the wording where the intent misses, and confirm the rest. Prompts that tested well by hand can still underperform in live tracking, so the set keeps moving after launch.
If you take one thing from this: a prompt set is a model of your buyer, and the material you build it from decides whether your visibility numbers mean anything. Build it from buyer language and the numbers describe a buyer you recognize.
What Radyant does with prompt sets
Radyant builds every client's prompt set from buyer language: sales calls, customer interviews, reviews, support tickets, and Search Console, expanded into the fan-out queries AI Search interfaces like ChatGPT and Perplexity run on their own. Every prompt gets tested by hand, two to three times across ChatGPT, Perplexity, and Google AI Mode, before it enters tracking. Then we read the results next to the leads they produce, through self-reported attribution, AI-referral sessions, and branded-search lift, the method we describe in our AI Search attribution guide.
The method works with any tracking tool. We use Peec AI and are a Trusted Partner in their directory; that link is a referral link. The quality of the prompt set matters more than the tool that tracks it.
If you want to know where you stand in AI Search, run a free audit.
FAQ
How many prompts should a B2B SaaS team track for AI Search visibility?
Our reusable setup is 30 long, context-rich prompts plus 5 brand prompts. That covers the main topics and personas without making manual testing impractical, and every prompt gets tested by hand before it goes in. Short-tail prompts come on top of these 35 when a category check is needed.
How long should a tracked AI Search prompt be?
Between 180 and 210 characters across our client setups, which is our own practice rather than a study result. That's room for a job title, company context, a specific problem, and a question that asks for a recommendation.
Do buyers really type long prompts into ChatGPT?
The tracking does not wait for someone to type your prompt. The tool runs your prompt as written, so it measures whether ChatGPT recommends you for that buying situation, whatever words a real buyer would have used. The persona, the pain point, and the recommendation ask have to be right; the exact phrasing doesn't.
Should marketing or sales own the prompt set?
Marketing builds it, and the people closest to buyers review it. Sales leads and unit heads catch prompts no real prospect would ask and settle the competitor set, so their sign-off is a design step. When they disagree on buyer maturity, track both plain and expert phrasings.
Why keep brand prompts out of the headline visibility number?
They saturate. In our own Peec project, brand prompts mentioned us in 99.7% of tracked responses between 12 July and 11 August 2026, and every other prompt in 13.3%. Blending them inflates the headline figure, so brand prompts get their own topic and are read for sentiment and positioning instead.
