Track the questions a buyer asks before they know your name, phrased the way they would type them, in each market and language you sell in. Cover different intents rather than many wordings of one question: the category, the alternatives to a rival, the constraint that sets you apart, the problem you solve, and the buyer describing their own situation.
Our research note on what seven AI engines cite used the questions Cituna tracks for itself, 22 of them by the end of its window. This guide shows how to choose yours.
What a prompt set is for
A prompt set is a fixed list of questions you ask the engines on a schedule, recording whether each answer mentions or cites you. It is the raw material of prompt tracking and of AI share of voice.
Because it is a sample, the choice of questions decides what the numbers can tell you. A set made only of questions that name your brand will say how engines describe you, but not whether they find you. A set made only of broad category questions will miss the specific needs you actually win on.
The five kinds of prompt worth tracking
| Type | Example from our set | What it tells you |
|---|---|---|
| Category | "best AI visibility tracking tool" | Whether you are on the default shortlist |
| Alternatives | "best Peec AI alternative" | Whether you are offered to a rival's unhappy customers |
| Constraint | "AI visibility tool with an API for developers" | Whether your differentiator registers |
| Problem | "how to get cited by ChatGPT" | Whether your content is read on the problem you solve |
| Persona | "I run marketing at a small B2B SaaS. Which AI visibility tracker covers ChatGPT, Claude and Grok without enterprise pricing?" | Whether engines match you to your actual buyer |
Aim for a few prompts of each type for each product line, and write the persona prompts in full sentences with the context a real buyer would give.
Where to find real questions
- Search Console. Filter your queries for question words or for queries of eight words or more. How to use Google Search Console for AI search gives the regular expressions.
- Bing Webmaster Tools. Its AI Performance report, in public preview since February 2026, shows a sample of grounding queries: the phrases the AI used when retrieving pages it then cited in Copilot, Bing's AI summaries and some partner services.
- Roundups and rival pages. In our data, each cited "best tools" list was tied to specific questions. A list of the best Perplexity rank trackers was cited 103 times for "best Perplexity rank tracker" and "best tool to check Perplexity rankings". The lists in your category show which questions they are built to answer.
- Sales and support. Collect the exact words prospects use on calls and in tickets, and test them as prompts.
Phrase them the way people ask
The engines do not search your exact words. OpenAI says that when ChatGPT search uses outside search providers, it typically rewrites the question into one or more targeted queries for them. Google says AI Overviews and AI Mode may run several related searches on subtopics, which it calls query fan-out. Its example turns a question about fixing a weedy lawn into searches about herbicides, chemical-free weeding and prevention.
Google adds that its AI systems understand synonyms and the general meaning of a question. The engine does its own rephrasing, so spend your prompts on different intents rather than near-identical wordings, and write each one as a person would, not as a keyword. Our glossary entry on search intent covers how to tell intents apart.
Watch for names that can be misread
Some prompts get answered as a different question. In our scans, Perplexity answered "best Profound alternative" with lists of alternative-rock bands, 183 citations in all, and Perplexity and Gemini answered "best Otterly AI alternative" with 96 citations to roundups about Otter.ai, a transcription app.
Keep such prompts, because buyers type them. But read the sources behind each answer before you count it, and add a variant that names the category, such as "best Profound alternative for AI visibility tracking".
Run every prompt on every engine, separately
The same question gets a different reading list on each engine. In our scans, ChatGPT's five most-cited sites and Grok's five had nothing in common, and Perplexity cited 20.2 third-party sources per answer against 3.5 for Gemini. A prompt set tracked on one engine tells you about that engine only.
Expect gaps on Google's side too. Google says AI Overviews appear only when its systems judge them additive to classic search, and often do not trigger. In our scans, Google AI Overviews returned 523 answers and Google AI Mode 248, against 552 for each of the other five engines. Record whether an AI answer appeared at all, not only whether it named you.
Hold the conditions steady
Answers depend on who is asking and from where. OpenAI says ChatGPT may estimate a user's location from their IP address, and may use saved memories when it rewrites a search query. Our methodology notes that answers vary by account, location, language and time, and that the same prompt asked twice can return different text and sources.
- Run every prompt from the same location, in the market's language, from a logged-out session or a clean account.
- Run each prompt repeatedly and report how often you appear, not what one answer said.
- Change the set rarely. When you do, record the date, because a new prompt changes the totals.
A tracker automates the repetition. Cituna, built by the team behind this site, runs this kind of check across seven engines daily.
Do not build a page for every prompt
Tracking many questions is fine. Publishing a page for each variation is not. Google says creating separate content for every way people might search, including fan-out queries, primarily to manipulate rankings or its AI responses violates its scaled content abuse policy. Track widely, and answer the underlying needs with fewer, better pages.
What to do first
- Draft prompts for each of the five types, for each product line.
- Pull question-shaped queries from Search Console and add the best ones.
- Check each prompt with a misreadable name, and add a variant with the category.
- Fix the location, language and account you will run them from.
- Record the date of the first run, and of every change to the set.