# How to Build a Prompt Set for AI Visibility Tracking

AI Visibility Tracking · Published 2026-11-10 · Updated 2026-11-10 · RankMonks
Source: https://rankmonks.com/blog/ai-visibility-prompt-set

> Short answer: A prompt set is a fixed list of the questions your buyers ask AI tools, used to measure how often your brand is named. Build it from real buyer language, cover each type of question, keep it to a few dozen prompts and run it the same way every time so that results can be compared.

A prompt set is the fixed list of questions you put to AI tools, on a schedule, to measure how often your brand appears in the answers. It does the job that a keyword list does in rank tracking, and everything you report later depends on how well it was built.

This guide covers where good prompts come from, the types a balanced set needs, how many to start with, how to word them, and how to run the list so that one month can be compared with the next.

## Key takeaways

- A prompt set is a fixed list of buyer questions that you run through AI tools to measure your visibility.

- The list stays fixed because a change in the questions makes this month's result impossible to compare with last month's.

- The best prompts come from real buyer language in sales calls, search data, support tickets and communities.

- A balanced set covers category, comparison, problem, alternative and brand prompts, and a few dozen prompts are enough to start.

- You should run every prompt under the same conditions and record any change to the list as a new version.

## What is a prompt set, and why does it stay fixed?

A prompt set is a written list of questions, phrased the way buyers ask them, that you run through each AI tool at regular intervals. It stays fixed because measurement only means something against a constant: if the questions change, you cannot tell whether your visibility moved or the test did.

A prompt differs from a keyword in two ways. It is longer, usually a full sentence with context such as company size, budget or use case. Buyers ask assistants longer and more specific questions than they type into a search box, a shift described in our guide to [SEO in the AI era](/blog/seo-in-the-ai-era). A prompt also produces a written answer, not a list of pages, so what you record is whether you were named and how. If those measures are new to you, start with our guide to [AI visibility](/blog/ai-visibility).

## Where do good prompts come from?

Good prompts come from the places where your buyers already put their questions into words: sales calls, search data, support tickets and public communities. A list invented in a meeting room tends to sound like marketing copy, and nobody asks an assistant in marketing copy.

![Six sources feed a prompt set, which is a fixed list of buyer questions: sales calls, search data, People Also Ask boxes, support tickets, communities and review sites.](/posts/ai-visibility-prompt-set/where-prompts-come-from.svg)

Work through these sources:

- **Sales calls and demos.** Recordings and notes show the exact questions prospects ask before they buy, including the awkward ones about price and switching.

- **Search data.** Google Search Console lists the queries people already use to find you, and the longer, question-shaped ones convert neatly into prompts. A keyword tool adds what people search around your topic, as covered in our [keyword research](/blog/keyword-research) guide.

- **People Also Ask.** The question boxes in Google results show how a topic branches into follow-up questions. They are discussed in our guide to [featured snippets](/blog/featured-snippets).

- **Support tickets.** Existing customers describe the problems the product is hired to solve, in their own words.

- **Communities.** Forums, social groups and question-and-answer sites show how buyers ask peers for recommendations, which is close to how they ask an assistant.

- **Review sites.** Reviews show which features and frustrations buyers compare suppliers on.

Collect more than you need. Then merge the duplicates, rewrite each question in plain buyer language and cut the list to the questions that matter most to revenue.

## Which types of prompt should a set include?

A balanced prompt set includes five types: category, comparison, problem, alternative and brand prompts. Each one matches a different moment in a buying decision, and your visibility can differ sharply between them.

| Type | What the buyer wants | Example prompt |
| --- | --- | --- |
| Category | A shortlist | "What are the best payroll tools for a company of 50 people?" |
| Comparison | A choice between named options | "How does Brand A compare with Brand B for a small finance team?" |
| Problem | A way to solve something | "How can we stop payroll errors when staff work in two countries?" |
| Alternative | A replacement | "What are good alternatives to Brand A that cost less?" |
| Brand | Facts about one supplier | "Is Brand A suitable for a company with hourly workers?" |

These types follow the same logic as [search intent](/blog/search-intent): the wording tells you what the person is trying to do.

Category and problem prompts show whether you are found by people who do not know you yet. Comparison and alternative prompts show how you fare against named rivals. Brand prompts show whether the model's facts about you are right, and our guide to [fixing wrong information about your brand in AI answers](/blog/fix-wrong-ai-answers-about-your-brand) explains what to do when they are not.

Tag every prompt with its type. When you report, the split by type is often more useful than the overall figure. A brand that is named in most brand prompts and few category prompts has a discovery problem, not a reputation problem.

Keep brand prompts as a minority of the list. They nearly always name you, so too many of them flatter the headline number.

## How many prompts do you need to start?

You need a few dozen prompts to start, typically between 25 and 50 for one product in one market. That is enough to cover the main questions across all five types without making each round unmanageable.

![A funnel in five stages that narrows a long list of raw questions into a prompt set: collect every question, merge the duplicates, rewrite each one in buyers' words, balance the five prompt types, and fix the list at a few dozen prompts.](/posts/ai-visibility-prompt-set/from-questions-to-prompt-set.svg)

The arithmetic matters. Each prompt is run in every tool, and ideally more than once. Forty prompts across five tools, three runs each, is 600 answers to read and record per round. A long list that never gets run is worth less than a short one that is run every month.

A sensible starting split for a set of 40 prompts:

- 12 category prompts

- 8 comparison prompts

- 10 problem prompts

- 6 alternative prompts

- 4 brand prompts

For example, a cyber security training company might start with 40 prompts for its core product in the UK. It adds a second set of 25 only when it launches in a new country, because buyers there ask different questions and name different competitors.

## How should you word a prompt?

Word a prompt the way a buyer would say it out loud: a full sentence, in plain language, with the context a real person would include. The closer the wording is to real questions, the more the results tell you about real answers.

Rules that keep prompts honest:

1. **Use natural phrasing.** "Which CRM is easiest for a five person sales team to set up?" beats "best CRM small business".

2. **Include realistic context.** Company size, industry, budget and location change the answer, and buyers mention them.

3. **Leave your brand out of non-brand prompts.** Naming yourself in the question almost guarantees you appear in the answer, which tells you nothing.

4. **Do not lead.** Avoid wording that points at your strengths, such as a feature only you have. Test the question a neutral buyer would ask.

5. **Ask one thing per prompt.** Compound questions produce long answers that are hard to score.

6. **Avoid internal jargon.** Use the words customers use for the category, even if your team prefers others.

7. **Write each prompt once and keep it.** Small rewording between rounds changes the test.

Read the list aloud to someone in sales. If a prompt does not sound like something a prospect has said, rewrite it or drop it.

## How do you run the set the same way every time?

Run the set the same way every time by fixing the conditions: a fresh session for every prompt, the same settings, memory off and the same schedule. AI answers already vary on their own, so every extra difference in how you ask adds noise that you cannot separate from real change.

![A checklist for running a prompt set the same way every time. Before the run: start a fresh session, turn memory and personalisation off, and use the same model settings. During the run: paste each prompt exactly, use one prompt per conversation and ask no follow-up questions. After the run: save the full answer, note the date and model, and log any change to the list.](/posts/ai-visibility-prompt-set/run-conditions.svg)

A run protocol to write down and follow:

1. Start a new conversation for every prompt, so that earlier questions do not carry over.

2. Turn off memory and personalisation where the tool allows it, or use an account kept only for tracking.

3. Use the same model or mode in each tool, and note which one it was.

4. Keep location and language consistent, since both can change which brands are recommended.

5. Paste the prompt exactly as written, and do not ask follow-up questions.

6. Run each prompt the same number of times per round.

7. Save the full answer, the date, the tool and the sources shown.

For [Google AI Overviews](/blog/google-ai-overviews), run each prompt as a search and record whether an Overview appears at all, since not every query triggers one.

Providers update their models without notice. You cannot hold that constant, so record the date and model name with every answer. When a figure jumps, the log tells you whether a model release is the likely cause.

Finally, treat the list like a document under version control. Give each version a number and a date. When you add, remove or reword a prompt, note what changed and why, and mark the round in which the new version started. Where possible, run the old and new versions together for one round so that the trend line has an overlap.

With the set and the protocol fixed, the results can be turned into a single percentage, as explained in our guide to [AI share of voice](/blog/ai-share-of-voice). The same answers also feed [AI source analysis](/blog/ai-source-analysis), which looks at the pages the models cite.

## Frequently asked questions

### How is a prompt set different from a keyword list?

A prompt set is different from a keyword list because it contains full questions with context, not short search terms. A keyword list is tracked by position in a list of results. A prompt set is tracked by whether your brand is named and cited in a written answer.

### How often should you change the prompts?

Change the prompts as rarely as you can, and review the list about twice a year. Edit it when you launch a product, enter a market or see that buyers have started asking something new. Every change should be logged as a new version.

### Should prompts include your brand name?

Prompts should include your brand name only in the small group of brand prompts that check what models say about you. Category, problem and alternative prompts should leave it out, because the point is to see whether the model brings you up unprompted.

### Do you need a tool to run a prompt set?

You do not need a tool to run a prompt set of a few dozen prompts, since a spreadsheet and a consistent routine are enough. A tracking tool or a service becomes worthwhile when the number of prompts, tools and repeat runs grows beyond what one person can do by hand each month.

A good prompt set sounds like your buyers and stays put long enough to show a trend. Drafting the list with you, then running it on a schedule, is the first step of our [AI visibility tracking](/services/ai-visibility-tracking) service. If you would like help building yours, [start a conversation](/contact).
