How to Find the AI Prompts Your Buyers Use and Which Ones Mention You

1. Why prompts are harder to find than keywords

Keyword research rests on published demand. Google shows search volumes, and every SEO tool is built on that data. None of the AI engines offers an equivalent public tool. Google's Search Console counts AI Mode prompts in its Performance report, but mixes them with ordinary searches and shows only the ones where your site appeared, and ChatGPT shows site owners nothing at all. Every prompt list, including the ones tools sell, is an estimate built from other evidence.

Prompts also look different. A buyer may type three words into Google and twenty into ChatGPT, adding their role, their market, their budget and the tools they already use. One keyword such as "payroll software" breaks into dozens of prompts that each carry a different buying context.

The good news is that the evidence exists, and some of it already sits in your own accounts.

2. Six places to find the prompts your buyers use

1. Search Console long queries

People now type into Google the way they type into ChatGPT, and some paste a full prompt straight into the search box. On our own site, 1,366 of the 3,354 queries that earned impressions between 29 July and 27 October 2026 had eight or more words: 41% of queries and 17% of impressions. One of them read: "how do i know what prompts i should track for our company in an ai search analytics tool?" It went on to ask for "the ones that actually matter for our category and buyer journey" instead of 500 generic questions. That is a prompt, and it arrived through Google.

How to pull it: in the Search Console Performance report, add a Query filter, choose Custom (regex), and use ^(\S+\s+){7,}\S+$ to keep queries of eight words or more. Run a second pass with ^(how|what|which|why|who|is|are|can|does|should)\b for question phrasing. Export both, covering the last three months. Search Console hides very rare queries, so treat the export as a sample of how buyers phrase things.

2. Sales and support calls

For a company with a sales team, this is the richest source, because buyers say out loud what they would type. Read or search call recordings and notes for the questions prospects ask before they commit: how long implementation takes, whether you integrate with their system, how you compare with the supplier they use now, what it costs for a team their size.

How to pull it: add one question to your discovery call script, "What did you ask ChatGPT or Google before you contacted us?", and log the answers word for word. From support, export the twenty most common ticket subjects for the last quarter.

3. People Also Ask, Reddit and Quora threads

These show the follow up questions real people ask around your category, in their own phrasing, including the doubts and complaints a buyer would not raise with a salesperson.

How to pull it: search your main category terms in Google from your market and record every People Also Ask question, expanding each one to reveal more. Then search Reddit and Quora for the category plus words such as "recommend", "alternative", "worth it" and "vs". Our guides to Reddit SEO and AI citations and Quora SEO show which threads AI engines quote.

4. The questions AI engines ask themselves

AI engines do not answer a prompt with one search. Google's own documentation says AI Overviews and AI Mode "may use a 'query fan-out' technique", issuing "multiple related searches across subtopics and data sources" to develop a response. Those sub questions show you what the engine thinks a buyer needs answered next.

How to pull it: run your strongest prompts in Google AI Mode, Perplexity and ChatGPT, then record the follow up questions each one suggests and the subtopics its answer is organised around. Perplexity usually lists related questions under its answers. Each of those is a candidate prompt in the engine's own words.

5. Comparison and review pages in your category

Buyers who reach the shortlist stage ask comparison questions, and the pages AI engines quote for those questions show the criteria buyers care about.

How to pull it: open the "best of" lists, "X vs Y" articles and review platform category pages for your market. Note the criteria each one compares on, such as price, integrations, support or local presence, and the competitor pairs they set side by side. Each criterion becomes a prompt: "Which payroll software for Singapore SMEs has the best local support?"

6. Bing Webmaster Tools grounding queries

Since February 2026, the AI Performance report in Bing Webmaster Tools shows how your content appears across Microsoft Copilot, AI generated summaries in Bing and selected partner integrations. Its grounding queries are, in Bing's words, "the key phrases the AI used when retrieving content that was referenced in AI-generated answers".

How to pull it: verify your site in Bing Webmaster Tools (you can import it from Search Console), open AI Performance and export the grounding queries. Read them as topic evidence: they are the engine's own retrieval phrases, shorter than what users type, and Bing describes them as a sample of citation activity.

3. Turn the raw questions into your prompt set

The six sources give you a long, messy list: fragments from Search Console, half sentences from call notes, forum titles, engine phrases. Tracking that list as it is would measure noise. Five steps turn it into the prompts that matter for your business.

1. Rewrite every item as a full question a buyer would type. Keep the buyer's own words and add the context they would give: role, company size, market, constraint. "payroll sg" becomes "What payroll software works best for a 40 person company in Singapore that already uses Xero?" The keywords you already rank for belong here too: rewrite each one as the question a buyer would ask behind it.
2. Write it in the language your buyers search in. A Dutch restaurant owner asks in Dutch and an Indonesian shopper asks in Bahasa Indonesia. Translating an English list afterwards produces questions nobody asks, so build each market's set from that market's sources.
3. Merge near duplicates. Ten phrasings of the same cost question become one prompt, with the clearest wording kept.
4. Tag each prompt with its buyer stage and its source. The stages are in the next section. The source tag tells you later which evidence produced your best prompts.
5. Score and cut. Rank each prompt on three questions: how close it is to a purchase, whether the engines already name suppliers when they answer it, and whether you have, or could build, the page that deserves to be cited. A prompt where no engine ever names a supplier cannot show you a recommendation, so it ranks lower. Keep the top 20 to 50 and freeze them.

4. Sort your prompts by buyer stage

A balanced set covers the whole journey, because each stage shows you something different. Five groups work well:

Keep branded prompts in their own group and report them separately. They show whether the engines describe you correctly, while the other four show whether the engines recommend you to buyers who have not heard of you yet.

5. How many prompts to track

Start with 20 to 50 per market and language. Fewer than 20 leaves whole stages empty; more than 50 is hard to read by hand, and a list you cannot read is a list nobody acts on. If you sell in three markets, that is three sets of 20 to 50, each built from its own sources.

Freeze the set for at least a quarter. If you keep editing the prompts, you can no longer tell whether a change in results came from the engines or from your edits. Review the list once a quarter, retire prompts that no longer match your offer, and add the ones new sales calls surface.

6. Find which prompts mention you

Run every prompt in each engine your buyers use: ChatGPT, Gemini, Perplexity and Google AI Overview are the usual four. For each answer, record three things: whether your brand is named in the text, whether one of your pages is cited as a source, and which competitors are named instead. Answers change between runs, so run each prompt more than once and read the result as a range.

The results rarely spread evenly, and the pattern is the useful part. In our August 2026 study, which put 89 buyer questions to ChatGPT and Google AI Overview every day for a month, and 39 of them to Perplexity as well, one vendor's naming rate depended almost entirely on the kind of question:

A second vendor in the same study shows why every engine needs its own line. On its own 50 category questions, asked of both engines every day, it was named in 74 of 1,484 Google AI Overview answers (5.0%) and in 1 of 1,550 ChatGPT answers. A single blended score would have hidden that ChatGPT almost never names it. Our study of AI search as three separate channels found the same split at page level: 94.2% of cited URLs appeared in exactly one engine.

The stage labels come from how each question was tagged when the study began, and the figures describe these two vendors in their own categories. Run the same count on your own set; it takes an afternoon for 30 prompts across four engines.

7. What the results tell you

Once the sheet is filled in, three patterns cover most of what you will see, and each points to a different fix.

• Named on branded prompts, missing on shortlist prompts. The engines know you but do not recommend you to new buyers. The usual causes are no page built for the category question and too few other websites listing you as an option; our guide to why ChatGPT and AI keep recommending your competitors covers all nine causes.
• Cited as a source, but not named. Your page feeds the answer while your name stays out of it. We measured these ghost citations across 6,601 answers and found they hit some brands in three out of four citing answers.
• A competitor owns the shortlist prompts. Note which sources the engine cites for those answers. They are the comparison pages, review platforms and threads you need to appear on.

When you are ready to track the set every month, our guide to AEO tracking covers scoring, cadence and when to move from manual runs to a tracker.

8. A prompt sheet you can copy

One row per prompt per engine per run keeps everything comparable. These columns are enough to start:

9. Still not sure which prompts your buyers use?

If the six sources feel like a lot of work for a first look, start with our free AI prompt discovery tool. It analyses your domain and sends you 30 to 50 buyer prompts across five funnel stages within 48 hours, which gives you a first set to test and edit. If you would rather talk the list through, book a 30 minute call with our team and bring the questions you are unsure about.

When a company works with us on answer engine optimisation, the prompt set is part of the Decision layer of The Novastacks Methodology. Prompt Radar Pro selects the buyer questions to track from the signals in the Data layer: Search Console, sales calls, customer reviews, the product catalogue and competitor pages. The set is configured per market and per language, so each market gets its own buyer questions. Each tracked question is checked several times and reported as a range, and we report per engine when an engine cites the brand and when it recommends it by name. The set is agreed with you before tracking starts.

Frequently asked questions