How to build a prompt set for tracking AI visibility

A prompt set is a fixed list of questions your customers might ask an AI assistant, which you run on a schedule and record the same way each time. It is the simplest way to see whether assistants mention your business, what they say, and who they name instead. You can build one in an afternoon with a spreadsheet.

The method matters more than the tool. Choose questions from real customer language, group them so the results mean something, keep the wording frozen, and write down what you see in the same columns every time. A list that changes each month can't show a trend.

One caution before you start. A prompt set is a sample you designed, not a record of what people really ask. As of October 2026, we have not found an AI vendor that shows site owners the prompts its users type. Treat your results as a consistent test, and be modest about what they prove.

What a prompt set can and can't tell you

It can tell you whether an assistant names your business for a given question on a given day, whether the description is accurate, which sources it links to, and which competitors appear. Over several runs, it can tell you whether those things are moving.

It can't tell you how many people ask that question, how often your business appears for all users, or how much revenue an appearance is worth. The platforms don't expose that. Google's Search Console report for AI features counts impressions and has no list of queries, as we explain in how to read the Generative AI performance report. Microsoft's AI Performance report shows "grounding queries," which it describes as "the key phrases the AI used when retrieving content," and adds: "They are not full user questions or prompts."

So your prompt set is the one measurement where you control the questions. That is its strength and its limit.

Step 1: collect real questions

Start with what customers have already said to you. Good sources:

• Notes and recordings from sales calls, and the first message in inbound emails and chats.
• Questions your front desk or support team answers every week.
• Reviews, where people describe the problem they had in their own words.
• Queries in Search Console's classic Performance report, especially long ones phrased as questions.
• Grounding queries in Bing Webmaster Tools, if you have data there.

Write each one as a person would say it to an assistant: a full sentence with context. A prompt to an assistant can carry more detail than a typed search, such as the person's situation, budget or location. Our guide to how GEO keyword research differs from SEO covers that shift.

Step 2: group by intent and buying stage

A flat list gives you one number. Groups tell you where the problem is. A business that appears for its own name but never for category questions has a different job to do than one that appears nowhere.

The examples in this table are hypothetical, written for an imaginary accounting firm in Denver.

Add a location group if you serve local customers, with the city or neighborhood written into the prompt. Add a product group if you sell goods.

Don't write a separate page for every prompt afterward. Google's guidance warns that creating content for every variation of a search, done mainly to manipulate rankings or AI responses, violates its scaled content abuse policy. The background is in query fan-out explained.

Step 3: decide how many prompts

There is no standard number. Our suggestion is 20 to 50 prompts for a small or mid-sized business checking by hand. Fewer than about 20 and one odd answer swings the whole result. More than about 50 and manual checks across several assistants stop being realistic, and people quietly stop doing them.

Within the set, we suggest keeping brand prompts to a minority. A prompt that names your business is more likely to produce an answer about it, which can flatter the totals. Report them as their own group.

Tracking tools change the math, because they run prompts automatically. Vendors commonly price by the number of prompts and engines tracked, or by credits, so a tight, well-chosen list also keeps the bill down.

Step 4: freeze the wording

1. Give every prompt an ID, a group and the date it was added.
2. Write the final wording once. Don't tidy it later, because a small change can produce a different answer and break the comparison.
3. Keep prompts neutral. "Who are good accountants for restaurants in Denver?" is a test. "Why is [your firm] the best accountant in Denver?" is a leading question that tells you nothing.
4. If a prompt stops being relevant, retire it with a date. Don't delete its history.
5. Add new prompts as new rows with their own start date, and report them separately until they have a few runs behind them.

Step 5: choose assistants and fix the conditions

Test the assistants your customers are likely to use. For most businesses that means a short list, not every product on the market. If you don't know, ask new customers which they use.

Then fix the testing conditions and write them down, because the vendors document several things that change answers.

• Memory and history. OpenAI's help article on ChatGPT search says: "If memory is enabled, ChatGPT may use relevant saved memories when rewriting a search query." An account that has discussed your business for months is not a neutral tester.
• Location. The same article says ChatGPT "may use an approximate location based on your IP address to provide relevant local results." Run local prompts from the area you serve, or name the place in the prompt.
• Personalization. Google's AI Mode help page says personalization is available "if you're 18 or older and have history and personalized recommendations enabled."
• Feature. AI Overviews and AI Mode are different Google features and should be logged separately.

Our suggested routine: start a new conversation for every prompt, use the same account state each time (signed out where the product allows it, or a clean account kept for testing), and note the device and location. You won't remove all variation. You will stop adding your own.

Step 6: record answers the same way every time

Use one row per prompt, per assistant, per run. These columns are enough.

Keep "mentioned" and "linked" apart. An assistant can name you without linking to you, and can link to your page without naming you. They are different outcomes and call for different work.

Save the evidence. A screenshot or pasted answer with a date lets someone else check your reading later. It also gives you a record when an answer is wrong, which is the starting point in what to do when AI says something wrong about your business.

The "sources cited" column can be the most useful. If the same directories, review sites or publications keep appearing for your category prompts, you have a list of places where your information should be present and correct.

Step 7: set a re-run schedule

Again, this is our suggestion and not a standard: re-run the full set once a month, at about the same point in the month, under the same conditions. Monthly fits a monthly reporting cycle and is a pace most teams can keep up by hand.

If you have the time, run each prompt two or three times per session in fresh conversations and record each run. That shows you how stable an answer is. A business named in three of three runs is in a different position from one named in one of three.

Why answers vary

Generative systems don't return a fixed result the way a database does. The same prompt can produce different wording, different sources and different business names on different runs. Several documented factors add to that.

• Different features use different systems. Google's documentation on AI features says "AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary."
• Assistants that search the web can rewrite your question into one or more searches, and the underlying search results change over time.
• Memory, history, personalization and location shape the answer, as noted above.

OpenAI also tells its own users that "Search results and citations can be incomplete, outdated, or incorrect." Expect noise, and design around it with repeat runs and a long view.

Turning the log into numbers

Keep the math simple and label it honestly.

• Mention rate. Runs where you were mentioned, divided by total runs, per group and per assistant.
• Link rate. Runs where your site was linked, divided by total runs.
• Accuracy issues. A count and a list, not a percentage.
• Who else appears. The names that show up most often in your category group.

State the sample every time. To take a hypothetical result, "mentioned in 9 of 30 category runs" is clearer than "30%." Don't call the result market share. It describes your test, not the market. These figures become one section of the report described in what a monthly GEO report should include.

When a tracking tool makes sense

Manual checks work for a small set and a few assistants. Past that, a tool saves hours by running prompts and storing answers for you. Bring your own prompt set to the tool instead of accepting a suggested list, so the questions stay tied to your customers. Remember Google's caution about every such product: "No third-party tool has access to our internal ranking or AI systems." A tool automates the kind of sampling you would do by hand, and it reports what it sampled. GEO tools vs. hiring an agency compares the options, and tracking AI mentions of your brand covers the wider practice.

Conclusion

A prompt set is a modest instrument that earns its keep through consistency. Take the questions from real customers, group them by intent, freeze the wording, fix the testing conditions, and log every run in the same columns. Then read the trend, not the single answer. If you want a starting view of how your site looks to AI search before you build your list, eSEOspace offers a free GEO audit of your website structure and visibility. More guides are in our GEO and AEO Guide.

Frequently asked questions