No assistant publishes the questions people ask it, so finding them means combining sources that approximate real questions: your own customer conversations, Search Console queries, community threads, and the follow up questions AI itself suggests. The bottom line: start with what customers say to you, add long queries from Search Console, check the prompt databases some tools maintain, and confirm each question by running it on the platforms you care about.
Related: our guide to answer engine optimization defines the term and lists eight steps.
Why this is hard, and what is known
Semrush found that between 65 and 85 percent of ChatGPT prompts do not match any traditional search keyword and that the average prompt is 23 words. Ahrefs says its Brand Radar index covers hundreds of millions of AI prompts a month, most of them Google AI Overviews and AI Mode, and Semrush maintains a database of more than 26 million prompts and responses, so commercial databases exist, but they sample broad topics and may not contain your customers' specific questions. That is why your own conversations come first.
Seven sources, in order of value
| Source | How to use it | What it gives you |
|---|---|---|
| Sales calls and email | Read 20 recent inquiries and list every question | The exact words buyers use |
| Support tickets and chat transcripts | Group by topic | Objections and post purchase questions |
| Search Console queries | Filter for long, question style queries with a regular expression | Real searches that already show your pages |
| Bing Webmaster Tools AI Performance | Read grounding queries, Topics, and Intents | The searches Microsoft's AI ran, which are not the user's prompt |
| Community threads | Search Reddit, forums, and Q and A sites for your category | Questions in natural language |
| AI follow ups | Ask an assistant the main question and note the suggested next questions | How systems expand the topic |
| Competitor pages | Read their FAQ and comparison pages | Questions they have decided matter |
Steps
- Collect 50 raw questions. From the first three sources.
- Add long queries from Search Console. In the Performance report, add a query filter of type custom regex and use the expression below, then sort by impressions.
- Add the place and situation. Turn "best roofer" into "who installs a metal roof on a coastal home in Sarasota."
- Remove duplicates and group by intent. Compare, price, how to, who to call, and whether it is worth it.
- Run each question five times on ChatGPT, Gemini, and Perplexity. Record whether you are named and which sources are cited.
- Rank by gap and value. Prioritize questions where you are missing and a customer would pay.
- Refresh the list quarterly. Questions change as products, prices, and regulations do.
^(who|what|when|where|why|how|which|can|does|do|is|are|should|will|best)\bA caution about grounding queries
Bing's AI Performance report shows grounding queries, the phrases Microsoft's AI used internally to retrieve sources. They are not the user's prompt. Use them to see how the system searches, and not as a list of what people asked.
Use the list, then measure with it
The same question set becomes your baseline. SparkToro and Gumshoe found less than a 1 in 100 chance that two AI responses returned the same list of brands, so run each question many times and track the share of runs that name you.
Questions people ask
How do I find out what questions people ask AI about my industry?
Combine your own sales and support conversations, long question style queries from Search Console, community threads, competitor FAQs, and the follow up questions assistants suggest, then test each on the platforms you care about.
Are AI prompts published anywhere?
No assistant publishes them. Tools such as Ahrefs Brand Radar and Semrush maintain prompt databases, but they sample broad topics and may not include your customers' specific questions.
What are Bing grounding queries?
The phrases Microsoft's AI used internally to retrieve sources for an answer. They are not the user's prompt.
How many questions should I track?
Start with 20 to 50 real questions, run each five times on several platforms, and refresh the list quarterly.