How to Choose Prompts to Track in AI Search (A Practical Guide)

By Yahav Fuchs •

Short answer: build your prompt list from real buyer language, weight it toward unbranded prompts, spread it across the funnel from problem awareness to purchase, freeze a core set, and review it quarterly. The list you track determines what your visibility number can possibly tell you, so this is the highest-leverage hour in any AI visibility program.

  • Source prompts from sales calls, support tickets and community threads, not from a brainstorm.
  • Keep most prompts unbranded. That is where discovery is won or lost.
  • Cover problem, solution, comparison and decision stages.
  • Freeze a core set so periods are comparable.
  • Add a small experimental set for new ideas.

Why prompt selection matters more than anything else

Every visibility metric, whether mention rate, share of voice or citation rate, is computed from a prompt set. If the set is wrong, the metric is precisely wrong. Teams make three common mistakes: they track only prompts that name their brand, they track prompts they already win, or they change the list every month. Any of them produces a chart that looks informative and is not.

A good prompt set is a sample of how your future customers will actually ask questions. If you were a pollster, this would be your sampling frame. It deserves the same care.

Step 1: Gather real language

Start with where buyers already speak.

Sales calls. Listen for the first thing a prospect says about the problem. “We have six tools and nobody knows which numbers are right” is a better prompt seed than “marketing analytics platform.”

Support tickets and onboarding surveys. Ask new customers what they typed or asked before they found you. A single question in onboarding, “What were you searching for or asking when you found us?”, produces more usable prompts than a keyword export.

Community threads. Reddit, Slack groups, forums and review comments contain unprompted questions in natural phrasing. Thread titles are, in effect, prompts: “Best X for small teams?” or “Is Y worth it?” This is also why those threads get cited. In ReddGrow’s study of 994 Reddit threads that AI engines cited, prompt words in the thread title and opening were the strongest separator of heavily cited threads. Questions people ask in public are near cousins of the questions they ask an AI.

Search data. Search Console queries and keyword research are fine as a cross-check. They tend to be shorter and more keyword-like than AI prompts, so lengthen them into full questions.

The engines themselves. Ask ChatGPT or Perplexity what questions people ask about your category, and look at the “related questions” and follow-ups that engines suggest. Treat these as hypotheses, since a model’s guess at what buyers ask is not the same as what they ask.

Step 2: Write them like people, not like keywords

AI prompts are longer and more specific than search queries. People add context: their team size, budget, tools they already use, constraints. Compare:

  • Keyword style: “CRM software small business”
  • Prompt style: “What’s a good CRM for a 5-person consulting firm that already uses Google Workspace and doesn’t want to pay per seat?”

The second is how the question arrives in an engine, and the answer is different, including which sources get cited. Include both short and long variants of your most important prompts. A few variants of the same intent also show you how stable your visibility is: if a small rewording drops you from the answer, your position is fragile.

Step 3: Balance branded and unbranded

Prompts fall into two families:

Unbranded: no company named. “Best invoicing software for freelancers.” “How do I track brand mentions in ChatGPT?” This is discovery. A buyer who does not know you yet asks, and the engine decides which names to offer.

Branded: your name appears. “Is [Brand] good?” “[Brand] pricing.” “[Brand] vs [Competitor].” This shows what the engine says once a buyer is already considering you.

A reasonable starting split is about 70% unbranded and 30% branded, including competitor-branded prompts. The right ratio depends on your stage: a new brand with low awareness should weight unbranded even more heavily, since branded prompts will often return nothing for a brand the engine barely knows. A well-known brand should still keep a solid unbranded core, because that is where share is taken from competitors.

The longer argument is in branded vs unbranded prompts. Short version: branded prompts are easy to win and unbranded prompts are where the market is decided.

Step 4: Cover the funnel

Different stages produce different kinds of prompts, and different kinds of sources get cited. Spread your list across four stages.

StageWhat the buyer is doingExample prompt
Problem awarenessTrying to understand an issue”Why are our reports showing different revenue numbers in two tools?”
Solution explorationLearning what category of tool exists”What kinds of tools help track customer feedback?”
EvaluationComparing options”Best customer feedback tools for B2B SaaS” and “Alternatives to [Competitor]“
DecisionChecking price, trust, fit”Is [Brand] worth it for a 10-person team?” and “[Brand] pricing”

Most teams over-index on evaluation, because that is where buying intent is obvious. It is also where competition is most intense. Problem and solution prompts are often cheaper to win, because fewer brands are trying, and they put you in front of buyers earlier. In our Reddit study, prompts about best-of recommendations and how-to questions drew different kinds of threads: best-of prompts were answered mostly by recommendation requests and roundups, how-to prompts by how-to questions and guides. If your list contains only one type, you only see one kind of source.

Step 5: Add the dimensions that change answers

Beyond stage, a few attributes change what an engine says. Cover the ones that apply:

  • Segment. “For startups,” “for enterprise,” “for agencies.”
  • Use case. The same product can win for one job and lose for another.
  • Geography or language, if you sell in more than one market. Answers can differ by location and language.
  • Constraints. Budget, integrations, compliance, team size.
  • Competitor-named prompts. “[Competitor] alternatives” is one of the highest-intent prompts there is.

A grid helps. List your three to five key use cases down one side and the four funnel stages across the top, and aim for at least one prompt in most cells.

Step 6: Prune and test the list

Before you commit, read the list as a skeptic.

  1. Delete duplicates. Five phrasings of the same prompt inflate whichever result they share.
  2. Delete prompts nobody would ask. If it reads like marketing copy, no buyer types it.
  3. Check for brand bias. Is more than a third branded? Are most of them prompts you already win?
  4. Run it once. If every prompt returns your brand, the list is too easy. If none do, check a few to confirm the prompts make sense for the category.
  5. Have a colleague who is not in marketing review it. They will spot the phrasing that no customer uses.

Step 7: Freeze, then evolve carefully

Once you have the list, split it:

  • Core set. 25 to 40 prompts, frozen for at least a quarter. This is what you report against and what makes trends comparable.
  • Experimental set. 10 to 20 prompts you can add and remove freely. Use it to probe new use cases, new competitors and new phrasing.

Promote a prompt from experimental to core only at a planned review. At the quarterly review, also retire prompts that no longer reflect how buyers ask, and write down every change so you can explain any break in the trend.

Choosing how many prompts and how many runs

Because answers vary, volume matters. A rough guide:

  • Under 20 prompts: too noisy to read as a trend, though fine for qualitative learning.
  • 30 to 50 prompts: workable for a small team, with three runs each per engine.
  • 100 or more: useful for agencies and larger brands that want segment-level reads, and where automation becomes necessary.

The more you slice the data by engine, segment or stage, the more prompts you need per slice. If you want to read ChatGPT results for enterprise comparison prompts separately, you need enough of those to see a pattern.

Tooling can help with scale. ReddGrow lets you track your own prompt list across engines and see the sources behind the answers, and the options are compared in best LLM SEO rank trackers 2026. A spreadsheet works fine at the start, as laid out in how to track brand mentions in AI search.

A short example

Say you sell scheduling software for dental clinics. A balanced starter list might include:

  • Problem: “How do dental clinics reduce no-shows?” and “Why do patients miss appointments even after reminders?”
  • Solution: “What kind of software do dental practices use to manage appointment reminders?”
  • Evaluation: “Best appointment scheduling software for small dental practices” and “Alternatives to [major competitor] for dental clinics.”
  • Decision: “Is [Brand] worth it for a two-dentist clinic?” and “[Brand] pricing.”
  • Segment and constraints: “Scheduling software for dental clinics that integrates with [practice management system].”

About a third of that list would be evaluation prompts, with the rest spread across stages, and most of it unbranded. That is the shape to aim for.

Mistakes to avoid

Brainstorming in a conference room. The list reflects how the team talks about the product, not how buyers talk about the problem.

Letting a tool generate the whole list. Tool-suggested prompts are a decent first draft, and you should still edit them against real customer language.

Optimizing for the prompts you track. If you only publish for tracked prompts, you will beat your own metric and miss the market. Keep a few untracked questions in mind from sales.

Forgetting to review. Language changes, new competitors appear and your product moves. Put the quarterly review on the calendar.

The takeaway

A prompt list is a model of your market. Build it from real language, keep it mostly unbranded, cover the funnel, freeze a core, and review it on a schedule. Do that, and the rest of your AI visibility reporting becomes much more trustworthy.

Frequently Asked Questions

How many prompts should I track in AI search?
Start with 30 to 50 that cover your main categories, comparisons and problems. Fewer than 20 makes the numbers too noisy to read, and hundreds before you have acted on the first set is usually wasted effort. Grow the list only when you have a reason, such as a new product line or market.
Where do I find prompts that real buyers use?
Sales call recordings, support tickets, onboarding surveys, community threads and search queries. People type longer and messier questions into AI engines than into search boxes, so copy their wording. Keyword tools are a starting point, not the source of truth.
Should I track branded prompts?
Yes, but keep them a minority. Branded prompts show what engines say about you once someone knows you, including pricing, reviews and comparisons. They will not tell you whether new buyers discover you, which is what unbranded prompts measure.
How often should I change my tracked prompts?
Rarely. Freeze a core set for at least a quarter so you can compare periods. Keep a separate experimental set where you add and drop prompts freely, and promote a prompt to the core set only when it has proven useful.
Do prompts need to be the exact words people use?
They need to be close, not identical. Engines respond to intent and wording, so a paraphrase can change the answer, especially in the sources cited. Include a few variants of your most important prompts to see how stable your visibility is.