AI answers vary from one run to the next, differ between tools and change as models and indexes update. That makes one-off checks almost useless: a single screenshot showing your brand mentioned, or not, tells you little. A prompt set solves this. It is a fixed list of questions that represent what buyers ask, run on a schedule across the main AI tools, with results recorded in the same way each time. Over months, it shows whether your visibility is improving, which competitors appear instead, and which sources the tools rely on. This article explains how to build, run and use one.
Choose prompts from real buyer questions
Start from what buyers actually ask, not what you would like to rank for. Sources include sales call notes, questions in demos, support tickets, Search Console queries, People Also Ask results, forum threads and customer interviews. Write prompts the way a buyer would phrase them to an assistant, which is often longer and more conversational than a search query. Cover several intents: category questions (best tools for a given job), comparisons (product A vs product B), costs and pricing, how-to questions where your expertise is relevant, problem questions describing a pain, and brand questions about your own company. Include your location or market where it matters, such as India or a city.
Size and structure the set
A set of 30 to 100 prompts is enough for most companies. Group them by intent and by product or service line, so you can see where visibility is strong or weak. Keep the wording fixed once chosen, because small changes in wording can change answers. Add a few prompts for each main competitor's brand, which shows how tools describe them and which sources they draw on. Review the set twice a year and add new prompts for new products or changing buyer questions, keeping the originals so trends remain comparable.
Running the prompts consistently
Run the set in the tools your buyers use, typically ChatGPT, Perplexity, Gemini and Google with its AI features, and Microsoft Copilot where relevant. Use the same settings each time: logged out or a clean account where possible, the same location, and web search enabled where the tool offers it. Because answers vary between runs, consider running each prompt more than once per cycle and recording whether your brand appears in each run. A monthly cycle suits most companies. Running manually is fine for small sets; for larger sets, a script using the tools' APIs or a monitoring tool saves time, as long as you check that API results resemble what users see.
What to record
Record the same fields for every prompt and tool in every cycle.
| Field | What to capture |
|---|---|
| Mentioned | Yes or no; for multiple runs, the share of runs mentioning you |
| Position | Where you appear in a list, if the answer is a list |
| Cited | Whether one of your pages is linked as a source, and which |
| Accuracy | Whether what is said about you is correct and current |
| Sentiment | Positive, neutral or negative framing |
| Competitors | Which competitors are mentioned |
| Sources cited | The third-party pages the answer relies on |
Scoring and reporting
Turn the records into a few measures. Mention rate: the share of prompts, or prompt runs, that mention your brand, by tool and by intent group. Citation rate: the share citing at least one of your pages. Share of voice: your mentions compared with competitors' mentions across the set. Accuracy issues: a count of answers with wrong information about you. Report these monthly as trends, alongside AI referral traffic from analytics. Avoid reading too much into a single month; look at the direction over a quarter.
Turning results into actions
The most useful output is the list of sources AI tools cite for your category. If a directory, review site, comparison article or publication keeps appearing, make sure you are present and accurately described there. If tools cite a competitor's guide or benchmark for a topic, consider whether you can publish a more specific or more current resource. If answers contain wrong information about you, trace where it comes from, often an outdated listing or old article, and correct it at the source. If a prompt group shows no mentions at all, check whether you have a clear, crawlable page that answers those questions.
Limits to keep in mind
Prompt tracking is a sample, not a census. Real users ask many variations, have personalised settings and conversation history, and tools change frequently. Results from logged-out or API runs may differ from what a logged-in user sees. Treat the measures as indicators of direction and use them alongside other evidence: AI referral traffic, branded search trends and what prospects say when asked how they found you. The value is in consistency over time, not in any single number.
Example prompts by intent
Examples make the intents concrete. For a B2B CRM consultancy in India, a category prompt might be: Which consultants help Indian manufacturers implement Zoho CRM? A comparison prompt: Should a 50-person distribution company in India choose HubSpot or Zoho CRM? A cost prompt: How much does a Zoho CRM implementation cost in India? A how-to prompt: How do I route IndiaMART leads into Zoho CRM automatically? A problem prompt: Our sales team does not update the CRM; what can we do? A brand prompt: What does the company named in your brand do, and is it reputable? Writing ten or so prompts per intent for each main service line gives a set that reflects how buyers actually move from a problem to a shortlist.
Asking buyers directly
Prompt tracking shows what AI tools say; asking buyers shows whether it matters. Add a question to your enquiry form or first sales call asking how the prospect found you or researched options, with AI assistants as one of the choices alongside search, referrals and social media. Record the answer in the CRM. Over a few months, this shows how many prospects used AI tools in their research, which complements the prompt set and the referral traffic in analytics. It also gives sales useful context, because a prospect who arrived via an AI recommendation may have been told specific things about you that are worth confirming or correcting early in the conversation.
Tools versus a spreadsheet
A growing number of tools monitor brand visibility in AI answers automatically. They can save time with large prompt sets, many competitors or several markets, and they usually store history and calculate share of voice. Before buying, check which AI tools and locations they cover, whether they run prompts in a way that resembles what users see, how often they run, and whether you can export the raw answers and cited sources, which are the most useful part. For a small company, a spreadsheet and a monthly manual run of 30 to 50 prompts is often enough to start, and it builds a clear understanding of what the tools say before deciding whether automation is worth paying for. Whichever approach you choose, keep the prompt wording and run conditions fixed, so the history remains comparable.
FAQ
Usually 30 to 100, grouped by intent and product line, with fixed wording so results are comparable over time.
Those your buyers use, typically ChatGPT, Perplexity, Gemini and Google with its AI features, and Microsoft Copilot where relevant.
Monthly suits most companies. Run each prompt more than once per cycle if possible, because answers vary between runs.
Your brand mentions compared with competitors' mentions across the prompt set, by tool and intent group.
Trace the likely source, often an outdated listing or article, correct it there, and make sure your own site states the facts clearly and consistently.
Not at first. A spreadsheet and a monthly manual run of 30 to 50 prompts works for many small companies. Tools help with larger prompt sets, more competitors and several markets.
Yes. A few prompts about each main competitor show how tools describe them and which sources they rely on, which helps you find gaps.
The list of third-party sources AI tools cite for your category, because it shows where you need to be present and accurately described.
Most should not, because buyers usually ask category or problem questions first. Keep a separate small group of brand prompts to check how tools describe you directly.