On this page
Where good prompts come from
The best prompts are copied from customers, not invented. Pull phrases from sales calls, support tickets, onboarding forms, review sites, and the questions you see in Search Console and People Also Ask. People type to AI the way they speak, longer and more specific than a search box.
Include context the way a buyer would: “best accounting software for a 5-person design studio in Canada” will return a different shortlist than “best accounting software”. Write several versions of your key questions with different constraints.
Mix in language from your own market. A founder might say ads manager while a retailer says paid social. Using the terms your buyers use, including slang and abbreviations, makes the prompt set a truer picture of how they actually ask.
Real examples beat made-up ones. Sit with your sales or support lead for 30 minutes and ask for the last ten questions prospects asked before buying. Turn each into a prompt as close to the original wording as you can, and you will have a stronger set than any keyword tool provides.
Organize them by intent
Four groups cover most needs. Category prompts ask for the best option. Comparison prompts name two or more brands. Problem prompts describe a pain with no brand in mind. Brand prompts ask directly about you. Each tells you something different, so score them separately.
Thirty to fifty prompts is a manageable set. Tag each with its intent, product line, and funnel stage, so a drop can be traced. The AI share of voice guide shows how to turn the log into one number.
Remember that prompts age. A question about a product category in 2024 may be phrased differently now, and new competitors and features change the answers. Review the set each quarter, retire prompts that no longer match how people ask, and note the change in the log.
Does ChatGPT mention you yet?
Checks whether ChatGPT, Gemini and Perplexity mention your brand. Tracks it every week.
Run them in a way you can compare
Keep the settings constant: same engines, web search on, same location where it can be set, and a clean session. Run each prompt two or three times, since models vary. Log the date, the engine, whether you were named, the position, the accuracy of what was said, and the cited sources.
Version the prompt list. If you add or reword a prompt, note the date, so a change in the numbers is not mistaken for a change in the answers. Review a sample of full answers monthly, because a score can hide a tone problem.
Be careful about how many things you change at once. If you rewrite a page and add a prompt in the same week, you cannot tell what moved the numbers. Make one change, note the date, and give the results a few weeks to settle before you decide.
Examples of prompts by intent
A category prompt might be: “What are the best CRM tools for a ten-person recruiting agency?” A comparison prompt: “How does Tool A compare with Tool B for small teams?” A problem prompt: “How do I stop duplicate leads from syncing into my CRM?” A brand prompt: “Is Tool A good for agencies?”
Notice the specifics: team size, industry, and constraint. Buyers add them because they want a tailored answer, and engines respond by narrowing the list. Test the same question with two or three different constraints, because you may be strong in one segment and invisible in another.
Also include a handful of follow-up style prompts, such as “what about pricing” or “which is easiest to set up”, since people continue the conversation after the first answer and the follow-up often decides the shortlist.
Whatever you choose, store the raw answers. When a stakeholder asks why the number moved, the evidence is in the text of the answers, and you will want to quote it.
Spreadsheet, script, or tool
A spreadsheet is the right place to start. One tab holds the prompts with their intent tags, another holds the runs with date, engine, mention, position, accuracy, and sources. It forces you to read the answers, which is where the insight is.
A script helps when you outgrow it, but takes care. Use official APIs where they exist, mind terms of service, and remember that API answers can differ from what a person sees in the app, especially with web search and personalization.
A dedicated tool is worth the money when you need weekly runs across several engines and prompt groups, with history and exports. Check that it shows the full answer text, not just a score.
Turn what you learn into work
Misses are a to-do list. For each prompt where you are absent or wrong, ask which page should have answered it. If none, write it. If one exists, check that it opens with the answer and that its facts match your other pages. Get cited by ChatGPT covers the page shapes.
A weekly run by hand is tedious, so this is the part to hand off: Sprites runs your buying prompts through ChatGPT, Gemini and Perplexity every week and tracks which brands they name. The companion piece is track your brand in ChatGPT.
Questions
What is AI prompt tracking?
It is running a fixed set of buyer questions through AI tools on a schedule and recording which brands and sources the answers name, so you can see visibility change over time.
How many prompts should I track?
Thirty to fifty covers most businesses if they are spread across category, comparison, problem, and brand questions.
Why do I get different answers each time?
Language models generate answers rather than retrieving one fixed result, and web search can change what is pulled. Run each prompt several times and look at patterns.
Is this the same as managing my own prompts?
No. Managing your own prompts is about versioning the instructions your team sends to AI tools. This guide is about tracking what AI search answers say about your brand.