AI prompt tracking: how to build a buyer prompt panel
AI prompt tracking starts with a fixed list of the questions your buyers type into ChatGPT, Gemini and Perplexity, run on a schedule, with a record of whether your business is named. Twenty-five prompts is enough to begin. Write them in buyer language, split them by intent, and freeze the wording so this month's result can be compared with next month's.
Start with what buyers already ask
Open your last 20 sales emails, call notes or support tickets. The questions in them are your raw material. Add the Search Console queries that contain a question word, and the questions your competitors answer on their FAQ pages.
Then rewrite each one the way a person talks to an assistant: a full sentence with a constraint in it. "Best CRM" is a search query. "Which CRM works for a 15-person recruiting firm that lives in Gmail?" is a prompt, and the constraint is what forces the assistant to produce a shortlist. Buyers do this already. Locus reported in Q2 2026 that 45% of US online shoppers use an AI tool somewhere in the purchase journey.
Split the panel by intent
A panel of 25 that is all category searches will miss most of the ways people find you. This split works for most businesses. Adjust the counts if you have no local component.
| Prompt type | Example | Count |
|---|---|---|
| Category discovery | Best project management software for a 20-person agency | 8 |
| Comparison | Tool A vs tool B for a team with client approvals | 5 |
| Problem | How do I cut size-related returns in an online apparel store | 5 |
| Local or specific | Pediatric dentist in San Jose that takes weekend appointments | 4 |
| Branded | Is [your brand] any good for small teams | 3 |
Keep the three branded prompts out of your citation score, since you will usually be named there. Read them for accuracy: wrong prices, old plans and missing locations show up fast.
When we build a panel for a client, we favor prompts the brand is not cited on today. A panel full of existing wins cannot show progress.
Run the panel the same way every time
Answers change from run to run, so a single check tells you little. These rules keep the data usable:
- Use a fresh session with memory and personalization off.
- Run every prompt three times in each engine.
- Record whether you were named, linked or absent, plus the competitors named and the sources linked.
- Run the panel on the same day each month.
We count a brand as cited when it is named or linked in at least 2 of 3 runs on the measurement day. Pick a rule like that and write it down before you look at any results.
Score each engine on its own. SOCi's Local Visibility Index 2026 found ChatGPT recommends 1.2% of multi-location businesses, Gemini 11% and Perplexity 7.4%, so an average across engines hides where the gap sits. The data sources differ as well. Gemini grounds local answers on Google Maps, while ChatGPT and Perplexity use Bing Places and Yelp. A local miss in ChatGPT may come from a listing problem, so check Bing Places and Yelp before you touch the content.
Spreadsheet or AI prompt tracking tool
Twenty-five prompts across three engines and three runs is 225 answers a month. That is tedious by hand but doable, and a spreadsheet with columns for prompt, engine, three runs, cited yes or no, competitors and sources is all you need. A prompt tracking tool earns its fee when you want weekly runs or several panels. If you are weighing options for AI visibility tracking, this comparison of monitoring tools, agencies and apps covers what each one does.
Turn misses into a work list
The sources column is where the panel pays off. For every prompt where you are absent, look at which pages the engine linked and sort the miss into one of three types:
- You have no page that answers the question.
- You have a page, but the answer is buried below a hero banner and three paragraphs of history.
- The engine cited third-party sites, such as review pages or directories, that do not mention you.
The first two are fixed on your own site. The third means getting listed or mentioned elsewhere, which takes longer.
JCS Fashions, a 4,878-product Shopify apparel store and our first customer, is the worked example. Its sessions from AI assistants went from 20 in January to 569 in October 2026, and AI-assistant orders were 5 in October. Five orders is a small number and we report it as small. The panel is how we watch whether the citations behind those sessions are holding. The full customer story has the other figures.
Keep the panel stable, then extend it
Do not reword prompts between runs. If a prompt turns out to be badly written, retire it, add a new one and note the date. Each quarter, add a few prompts from fresh sales questions. Keep the original set so the trend line stays honest. The AI Visibility Audit sets a 25-prompt panel this way if you would rather have it done for you.
To get a first baseline this week, run the free AI visibility scan, then write your first ten prompts from the last ten questions a buyer asked you.