Guide

How to choose prompts to track for AI visibility: a method and a starter set

Your AI visibility numbers are only as good as the questions behind them. Here is how to build a prompt set from what customers actually ask, keep it balanced, and keep it stable enough to show a real trend.

Locus Intelligence team
Updated 8 min readVisibility · Engage

Key takeaways

  • Choose prompts from what real customers ask: calls, reviews, search queries and front-desk questions, not from what you wish they asked.
  • Balance five types: category, need plus place, comparison, problem and brand. Most of the set should not contain your name.
  • Start with 20 to 50 prompts. For locations, run a core set per location with the place swapped in.
  • Write prompts the way people type into an assistant: full sentences, with the detail that matters to them.
  • Freeze the set for a quarter. Changing prompts every week makes the trend meaningless.

How to choose prompts to track: the short version

To choose prompts to track, collect the questions your customers actually ask (from calls, reviews, Search Console and your staff), sort them into five types, pick 20 to 50 that match what you sell and where, and write each one the way a person would type it into ChatGPT or Gemini. Keep most prompts unbranded, and keep the set unchanged for at least a quarter.

A prompt is the AI-era version of a tracked keyword. The difference is that people ask assistants longer, more specific questions ("quiet hotel near the airport with a shuttle and late check-in") and the answer names a few businesses instead of listing ten pages.

From raw questions to a tracked prompt set
  1. 01CollectCalls, reviews, search, staff
  2. 02SortFive prompt types
  3. 03RewriteAs a person would ask
  4. 04PrioritizeBy revenue and fit
  5. 05LocalizePer location
  6. 06FreezeOne quarter

Step 1: collect the questions customers really ask

Spend an hour gathering raw questions before you write a single prompt. The best sources are the ones in your own business.

  • Phone calls and chats. The first question people ask when they call: "do you take walk-ins?", "how much is a brake job?". Ask front-desk staff for their top ten.
  • Reviews. What reviewers say they came for: "great for a birthday dinner", "fixed my AC the same day". Those are needs people search for.
  • Google Search Console. Filter your queries for question words (how, what, best, near, which) to see long questions people already type. Our local keyword research guide shows how to pull them.
  • Google Business Profile searches. The search terms that found your profile, in its performance view.
  • Competitor comparisons. What customers say when they mention a competitor: "we tried X but they were booked".

Step 2: sort them into five prompt types

A balanced set covers the whole path from "I have a need" to "is this business any good". Use these five types.

The five prompt types

Category

"Best {category} in {city}". The broad question. Hardest to win, most valuable.

Need plus place

"{Category} open Sunday near {neighborhood}", "{service} with {feature} in {city}". Specific and local; often the easiest place to start winning.

Problem

"My {problem}, who can fix it today in {city}?". The customer describes the problem, not the service.

Comparison

"{You} vs {competitor}", "Is {competitor} or {you} better for {need}?". Shows how you are framed against rivals; see competitor tracking.

Brand

"Is {your business} good?", "What do people say about {your business} {location}?". Checks accuracy and sentiment, not discovery.

Prompt typeShare of the setWhat it measures
CategoryAbout 25%Discovery against the whole market
Need plus placeAbout 35%Discovery for specific, high-intent needs
ProblemAbout 15%Whether you are linked to the problems you solve
ComparisonAbout 10%How you are framed against named rivals
BrandAbout 15%Accuracy and sentiment when people already know you

A suggested starting mix, not a rule. Shift it toward the types that bring you the most revenue.

Step 3: write each prompt the way people ask

Rewrite each raw question as a complete prompt a person would type into an assistant. Include the detail that changes the answer: the place, the need, the constraint.

Weak promptBetter promptWhy
dentist AustinWhich dentists in north Austin are good with nervous adult patients?Full question, a real need, a specific area
best pizzaBest pizza place near Wicker Park that does gluten-free crustsA place and a constraint that splits the field
Smith Auto reviewsIs Smith Auto in Tempe reliable for brake repairs?Brand check tied to the service you care about
plumberMy water heater is leaking, which plumber in Plano can come today?Problem-first, the way urgent buyers ask

Smith Auto is an invented example.

Should this prompt be in your set?
  1. Would a real customer ask this, in roughly these words?
    No
    Drop it or rewrite it in their words.
    Yes
  2. Do you sell this service in this place?
    No
    Drop it. Tracking prompts you cannot win only dilutes the numbers.
    Yes
  3. Would being named here lead to revenue?
    No
    Keep it only if it is a brand accuracy check.
    Yes
  4. Is it different enough from prompts you already have?
    No
    Merge it with the closest one.
    Yes
    Add it to the set.

Step 4: decide how many prompts to track

Start with 20 to 50. Fewer than 20 and a single changed answer swings your numbers. More than 50 at the start and nobody reads the results. Grow the set once the weekly review is a habit.

If you have locations, split the set into a core set run for every location (with the city or neighborhood swapped in) and a brand set run once. Ten core prompts across 30 locations is 300 local prompts per engine per run, which is why multi-location brands use a tool such as Visibility or one of the trackers in our roundup.

TemplateStarter prompt set (copy and fill in)
CATEGORY
1. What is the best {category} in {city}?
2. Top-rated {category} near {neighborhood}
3. Which {category} in {city} do locals recommend?

NEED PLUS PLACE
4. {Category} open {day or late} near {neighborhood}
5. {Service} with {feature, e.g. free parking} in {city}
6. {Category} in {city} that accepts {insurance or payment}
7. Affordable {service} in {city} for {audience, e.g. families}

PROBLEM
8. My {problem}, who can fix it today in {city}?
9. Where can I get {urgent need} in {city} without an appointment?

COMPARISON
10. {Your business} vs {competitor}: which is better for {need}?
11. Is {competitor} or {your business} better in {city}?

BRAND
12. Is {your business} {location} good?
13. What do people say about {your business} {location}?
14. What are the hours and services at {your business} {location}?

Replace the braces. For each location, copy prompts 1 to 9 and swap in that location's city or neighborhood.

Step 5: freeze, review and refresh the set

Keep the set unchanged for at least a quarter so week-on-week changes reflect the engines, not your edits. Review it each quarter: drop prompts that never matter, add new services and locations, and mark the change date on your charts.

Quarterly prompt set review

Keep
Change
Add

Mistakes to avoid

  • Tracking only branded prompts

    Instead: Keep brand prompts to about 15% of the set. New customers do not know your name yet.

  • Writing prompts like keywords

    Instead: Use full questions with the place and the need, the way people type into an assistant.

  • Tracking prompts you cannot win

    Instead: Drop services and places you do not serve. They drag the averages down and teach you nothing.

  • Changing the set every week

    Instead: Freeze it for a quarter and change it on a set date.

  • One set for all locations

    Instead: Run a core set per location with the place swapped in.

What to do with your prompt set

Run the set weekly in ChatGPT, Gemini, Perplexity, Claude, Copilot and Google, and record who was named, where, and which sources were used. Turn the results into mention rate, share of voice and gap rate with our AI search visibility metrics guide. For Perplexity's cited sources, see Perplexity visibility tracking.

Then work the gaps with how to improve brand visibility in AI answer engines, and if you outgrow a spreadsheet, compare trackers in best AI visibility tools. The generative engine optimization guide explains what each engine reads.

Free tool
Test your first prompt for free

The free AI visibility checker runs a question for your business across ChatGPT, Gemini, Perplexity and Claude and shows who gets named.

Check my AI visibility
Visibility by Locus IntelligenceRun your prompt set per location, every week

Visibility by Locus Intelligence runs your prompts per location in ChatGPT, Gemini, Perplexity, Claude, Copilot and Google AI Overviews, and records who was named and which sources each answer used.

Book a demo
AI visibility checker

Free, no account. You get whether ChatGPT, Gemini, Perplexity and Claude name you, how they describe you, and who they pick instead.

Your next steps

Today

  1. Ask your front-desk or sales staff for the ten questions they hear most.
  2. Run one category prompt in the free AI visibility checker.

This week

  1. Fill in the starter prompt set and cut it to 20 to 50 prompts.
  2. Run the set once in each engine and save the answers as your baseline.

This month

  1. Run the set weekly and calculate mention rate and gap rate.
  2. Set a date next quarter to review the set.

Frequently asked questions

How many prompts should I track for AI visibility?

Start with 20 to 50. For locations, run a core set of about 10 per location with the place swapped in, plus a brand set run once.

Should prompts include my brand name?

Some should, to check accuracy and sentiment, but keep them to about 15% of the set. Most new customers ask unbranded questions.

Where do I find the questions people ask AI assistants?

Most AI companies do not share the questions people ask about you. Bing Webmaster Tools shows sample grounding queries for Copilot citations; beyond that, use proxies: calls, chats, reviews, Search Console question queries and what your staff hear.

How often should I change my prompt set?

Once a quarter. Changing it more often breaks the trend.

Do prompts need to be long?

They should be as specific as real questions are, which usually means a full sentence with a place and a need.

Related

All guides
Perplexity visibility tracking

See if Perplexity recommends you, and which sources it trusts.

Guide · Visibility · 8 minGenerative engine optimization: the complete guide

What GEO is, what the engines read, what works, and the weekly routine that keeps a location named.

Guide · Visibility · 20 minShare of voice (AI)

The share of questions in a set for which an engine names a given business, compared with its competitors.

Glossary · Visibility · 2 minAI visibility checker

Whether ChatGPT, Gemini, Perplexity and Claude name you, and who they pick instead.

Free tool · Visibility · 1 minuteLocal keyword research

Six steps to find, sort and place the terms nearby customers search.

Guide · Reach · 10 min

Never miss another customer.

Start with a free audit. See what each location is losing, and where.

Profile score, Maps rank against the top 3, and the fixes that matter most.

Or book a demo
  • Role-based access by region and brand
  • Every action logged and reversible
  • Nothing customer-facing goes live without approval