Prompt Research: How to Choose AI Prompts Worth Tracking

Prompt research is the work of choosing which buyer questions you will measure in AI answers. It tells you whether your brand is named for those questions, which competitors are named instead and which sources each answer draws on. It cannot tell you how many people type a given question. It is also a different job from prompt engineering, which is wording instructions to get better output from a model.

A tracked prompt is one fixed question, run on a schedule on a named engine, with the answer recorded. Three mistakes ruin most lists:

• Choosing prompts no buyer would type.
• Trusting a single run.
• Ranking prompts by a volume figure that no engine publishes.

The sections below fix each in turn.

Start from words real buyers already used, collect 40 to 60 candidates, then cut to about 30. A good first set is 10 category questions, 15 comparison or alternative questions and 5 brand questions. If you have only 20 prompts, keep the same proportions (about 7, 10 and 3). The split is our recommendation, not a benchmark.

Rank your sources by how close they sit to a real buyer:

For the Search Console route, Google reports sites in AI Overviews and AI Mode within the "Web" search type of the Performance report. Add a Query filter set to Custom (regex), where a regular expression (regex) is a pattern that matches text, and enter:

([^\s]+\s){9,}[^\s]+

The pattern needs ten or more space-separated words, so a 9-word query is excluded and a 10-word one is kept. Long queries read like prompts, and they carry the specifics (team size, industry, budget) that make a prompt worth tracking. Search Console leaves out rare anonymized queries, so expect a short list. Treat it as a source of examples, not a complete set. Bing's report is covered by Search Engine Journal's report on its February 2026 public preview.

Forum wording matters most for some engines. In our Study 007, Google AI Mode cited a community platform (Reddit, YouTube or Quora) in 117 of 150 answers, Perplexity in 73, Gemini in 29 and ChatGPT in 3. Those were 150 questions on ten Indian topics, asked once per engine on October 5, 2026, so treat the pattern as a direction for US buyers, not a US figure.

Score each candidate 0, 1 or 2 on four tests and keep prompts scoring 6 or more out of 8. This is our scoring rule, built to drop prompts you cannot win and prompts that never touch revenue.

To run the influence test, ask the question once on each engine and open the cited sources. Add specifics until the prompt describes one buyer, not a whole market.

Illustrative example: "best project management software for a 20-person marketing agency" scores 2, 2, 1, 2 (total 7) and stays. "What is project management" scores 0, 1, 0, 1 (total 2) and goes. These scores show the method; they are not results from a real brand.

Keep the sheet as one row per prompt, with these columns:

Prompt | Source | Stage | Appear | Influence | Revenue | Total | Named in __/3 runs | Verdict

When results come in, compute mention rate as runs naming your brand ÷ runs that returned an answer.

Because ChatGPT, Perplexity and Gemini do not give site owners a count of how often a prompt is typed, and any volume figure is a modeled estimate. Rank by buyer stage and revenue instead, and treat volume as a tiebreaker at most.

The best public view of what people ask comes from the National Bureau of Economic Research study of ChatGPT usage, which classifies a sample of conversations by type rather than listing prompts. Search Console is the closest thing to observed demand on Google's surfaces. Bing Webmaster Tools adds grounding queries for Copilot citations, which are reformulations by the AI, not typed prompts.

If a tool gives you a prompt volume, ask how the sample was drawn, from which engines and over what dates. If you cannot get answers, use the number as a rough direction and nothing more.

Run each prompt three times per engine before acting on it. If your brand is named in all three runs or in none, the result is stable. If it is named in one or two, report it as a rate ("named in 2 of 3 runs"), never as a yes or no. This is our recommendation.

The reason is that identical questions return different answers. In Study 009, a plain re-ask minutes later left out 27% of the brands from the first answer (188 of 698 brands). That baseline was 20 questions on ChatGPT and Google AI Mode only, asked from India on October 5, 2026. Switching from English to Hinglish (Hindi written in Latin letters, with English nouns kept) removed 42% of the 245 brands named in the English answers. Treat both as a direction, not a US figure. If you sell to buyers who search in another language, test in that language too.

Study 009 did not measure rewording within one language. We still recommend testing two English phrasings of your top five prompts, because it separates wording from chance.

On Growth and Enterprise, which run daily, three consecutive days of scheduled runs give you the three samples. Launch runs each prompt once a week, so three runs take three weeks. For a faster check, run your prompts by hand three times.

A first set of about 30 prompts fits inside the 100 prompt slots on our Launch and Growth plans and the 200 on Enterprise. As of October 2026, a tracked prompt costs $2.99 to about $7.50 a month for buyers in the United States. The working is below, and it is our model, not an official figure.

The formulas are price ÷ prompts for the first cost column and price ÷ answers a month for the second. Answers a month assume 52 ÷ 12 weeks for Launch (300 × 4.33 = 1,300) and 30 days for the daily plans (15,000 on Growth, 42,000 on Enterprise). Answers per prompt equal the engine count: 300 ÷ 100 = 3, 500 ÷ 100 = 5 and 1,400 ÷ 200 = 7. That means each prompt gets one answer per engine per run, so repeat samples come from scheduled runs, not extra slots. Agency pricing is custom and annual billing saves 8% across plans, as of October 2026. See our pricing for current plans.

Run them on a schedule, read the answers for who is named and which sources they cite, and fix the gaps. Three follow-ups cover most of the work:

• Turn the prompts where competitors are named instead of you into pages, using a content gap analysis.
• Watch the same set over weeks with AI citation tracking over time, and check where you stand today with a free AI visibility audit.
• Connect the visits that follow with AI traffic tracking in GA4.

To set up a prompt list and plan with us, Book a demo.