Prompt Research: How to Choose Which AI Prompts to Track

Prompt research is the work of choosing which questions you will run through ChatGPT, Claude, Gemini, Perplexity and Google's AI features to measure whether they mention or cite your brand. Done well, it starts from evidence of real buyer demand (sales calls, support tickets, Search Console queries), rewrites that demand the way people actually talk to AI, sorts it by funnel stage and persona, keeps branded prompts out of the headline number, and freezes the set so a change over time means something.

This guide gives you a seven-step method, an evidence ladder that ranks prompt sources by how far you can trust them, a side-by-side of how five vendors estimate "prompt volume" and what they admit it cannot do, the arithmetic of branded-prompt bias with our own tracking data, sizing rules, and a worked 65-prompt set you can adapt. Google, Bing and vendor features were checked on October 18, 2026.

What is prompt research?

Prompt research is the AI-search equivalent of keyword research: it decides the sample of questions every AI visibility metric is calculated on. It has two jobs. Discovery finds out what your buyers ask AI assistants about your category. Selection turns that long list into a fixed, tagged set you run on a schedule and report on.

The second job is the one most teams underestimate. Mention rate, citation share and AI share of voice are all computed over the prompt set, so the set defines the number. Swap ten prompts and your "visibility" can move twenty points with nothing changing in the market. That is why this guide treats prompt research as measurement design, not brainstorming.

It is also different from prompt engineering, which is about writing instructions that get better output from a model. If you are looking for ready-made instructions to give an AI marketing assistant, see our list of AI marketing agent prompts instead. This page is about the questions your customers ask.

How is prompt research different from keyword research?

Keywords are short, measured and return a fairly stable list of links. Prompts are longer, unmeasured and return a generated answer that changes from run to run. Keyword research is still the best seed for prompt research, because keywords are the only demand data with real volumes behind them, but you cannot paste a keyword list into a tracker and call it done.

Two findings shape everything below. First, the way people phrase prompts is wildly diverse: in the SparkToro and Gumshoe study (January 2026), 142 human-written prompts for one headphone need had an average pairwise semantic similarity of 0.081. Second, the answers to those different prompts converged on the same brands. Intent is stable; wording is not. So you research intents and track a few phrasings of each, rather than hunting for the one "correct" string. Our query fan-out explainer covers why paraphrases end up retrieving similar sources.

The seven-step prompt research method

The method moves from evidence to a frozen, tagged set. Each step below starts with what to do, then explains why.

1. Decide what the set has to measure - the business questions, engines, markets and competitors.
2. Collect demand evidence from the sources on the evidence ladder, best first.
3. Convert keywords and notes into prompts using the context real people add.
4. Map every prompt to a funnel stage and persona, and tag it.
5. Split branded from unbranded prompts before you calculate anything.
6. Size the set and the runs so changes beat the noise.
7. Freeze, version and maintain it on a fixed review cycle.

Step 1: Decide what the prompt set has to measure

Write down the decisions the data will drive before you write a single prompt. Each decision needs a different kind of prompt:

• "Are we recommended in our category?" needs category prompts such as "best payroll software for a 200-person company".
• "Are we cited as a source on the problems we solve?" needs problem prompts, where the answer cites guides rather than naming vendors.
• "Do we win against the competitors we lose deals to?" needs comparison and alternatives prompts.
• "Is what AI says about us accurate?" needs branded prompts, measured separately.

In the same document, fix the engines, countries, languages and the named competitor list. These are part of the measurement, not settings you change later.

Step 2: Where do you find the real prompts people ask AI assistants?

Start with what buyers actually said, then widen out to measured search data, and only then to modeled or generated prompts. Most guides list prompt sources as if they were equal. They are not: a question a prospect asked on a sales call is direct evidence, while a prompt an LLM brainstormed is a hypothesis. The ladder below ranks the common sources by the strength of evidence they give you.

Two of these sources are newer than most guides on this topic. Google added a branded queries filter to the Search Console Performance report in November 2025, available to all eligible sites since March 11, 2026. It is AI-assisted rather than regex-based, it counts misspellings and unique product names as branded, and it only works on top-level properties with enough query volume. Bing launched AI Performance in Bing Webmaster Tools as a public preview in February 2026, including "grounding queries": the key phrases the AI used when retrieving content it cited. These are the closest thing to first-party AI query data a site owner can get today, but they describe retrieval, not what the user typed.

The rule we apply: every prompt in the final set should trace back to at least one source from rows 1-8. A prompt that only exists because a model suggested it is labeled "hypothesis" and has to earn its place in the next review.

The bottom of the ladder deserves a warning. When Otterly compared survey-sourced prompts with prompts from its own research tool (study updated February 2026), the real prompts averaged 15.1 words against 8.8, used personal pronouns 52.1% of the time against 18.8%, and were problem-oriented 21.1% of the time against 7.1%. Generated prompts most often opened with "best". The study's "real" prompts were what marketing professionals said they would type, from one category in one country, and Otterly sells the tool it tested, so treat the exact figures as indicative. The direction matches the SparkToro finding: humans write longer, stranger, more specific prompts than tools do.

To speed up the generated end of the ladder, AI Rank Lab's brand visibility module includes an AI Prompt Research tool that drafts prompts from up to 20 seed keywords, from a specific page URL, or from a brainstorm on your domain, for a country and language you choose, and labels each with an intent and a category. Its setup wizard analyzes your website first so suggestions reflect what you actually sell. Treat that output like any row-10 source: a starting list to check against rows 1-8, not evidence on its own.

Step 3: How do you turn keywords into prompts?

Keep the keyword's intent and add the context a real person adds when they talk to an assistant: who they are, what situation they are in, and what constraint matters. A reliable pattern is situation + need + constraint + ask.

Take the keyword "payroll software for small business":

• Short, search-like: "best payroll software for a 40-person company"
• Conversational: "what payroll software should a 40-person company use if we pay hourly staff in three states?"
• Persona-rich: "I'm the only HR person at a 40-person restaurant group with hourly staff in three states. Our payroll takes two days a month. What should we switch to?"

Track all three phrasings for your most important intents. The short version mirrors search-enabled behavior, and the long version mirrors how people actually write when they explain a problem. Semrush's clickstream data shows the two converging: search-enabled ChatGPT prompts nearly doubled to 8.7 words while non-search prompts nearly halved to 13.5 words between early 2025 and early 2026.

Rules that keep conversions honest:

• One intent per prompt. "Best payroll software and how much it costs" is two questions; split it.
• Use the buyer's vocabulary, not yours. If you coined a category name, prompts that use it are effectively branded. Otterly saw real users prefer "track" over "monitor"; those small word choices matter.
• Add a year only on purpose. Ahrefs recommends a time marker such as "in 2026" when you want answers that pull in cited URLs rather than relying on training data. Decide per cluster and keep it consistent.
• No leading prompts. "Why is [your brand] the best payroll tool?" measures nothing.

To pick which seed keywords deserve conversion, start with the ones that already trigger AI answers in Google. The AI Rank Lab keyword planner shows Google volume, difficulty and whether a keyword shows an AI Overview, which is a useful filter before you spend time rewriting.

Step 4: How do you organize prompts by funnel stage?

Assign every prompt one stage and one persona, and weight the set toward the stages where you can still win new buyers. The taxonomy below is the one we use; the shares are our recommendation, not an industry standard, and should follow your sales cycle.

Then tag every prompt with at least six fields: stage, persona, topic cluster, branded yes/no, market and language, and evidence source (the ladder row it came from). Tags are what let you report "category prompts for HR directors in the US" instead of one blended number. In AI Rank Lab, saved prompt research results can be filtered by tag, country and intent, which is the same structure in tool form.

Step 5: Should you track branded prompts?

Yes, but never inside the headline visibility number. A prompt that names your brand will mention your brand almost every time; Peec points out that a prompt naming the brand makes visibility effectively 100% by construction. Mixed into a blended metric, those prompts make you look far more visible than you are.

The arithmetic is simple. Take 40 prompts: 10 branded, 30 unbranded. The brand appears in 100% of branded answers and 20% of unbranded ones. The blended mention rate is (10 x 100% + 30 x 20%) / 40 = 40%. The number that reflects how often AI recommends you to people who have not heard of you is 20%. Branded prompts doubled it.

Our own data shows why the branded flag has to live on each prompt. Across 968 successful runs of 42 prompt-engine pairs tracked in AI Rank Lab between February 13 and October 14, 2026 (only pairs with 5 or more runs), Gemini mentioned the tracked brand in every single run for 10 of its 13 pairs (274 runs). ChatGPT did so for 1 of 11 pairs, and never mentioned the brand in 4 of them; Claude did so for 6 of 11. Our aggregate does not record whether each of those prompts named the brand, so we cannot tell how much of Gemini's always-on rate is genuine visibility and how much is prompt wording. That is the point: without a branded tag on every prompt, nobody can. The full method is in our AI answer stability study.

Practical rules:

• Headline metric = unbranded prompts only. Report the branded group separately for accuracy, sentiment and which sources engines rely on to describe you.
• Use Google's definition of branded. The Search Console filter counts misspellings and unique product names as branded. Apply the same rule to prompts, including category terms only you use.
• Keep competitor-named prompts in their own cluster. "Alternatives to [competitor]" is unbranded for you, but it behaves differently from a generic category prompt. SE Ranking and Peec both recommend tracking comparison prompts separately so they do not skew category metrics.
• Write the split down and keep it fixed. Nightwatch suggests roughly 30% branded and 70% unbranded; Profound recommends keeping branded prompts a minority. Whatever you choose, changing the ratio changes the blended number.

Step 6: How many prompts should you track, and how many runs?

Plan on at least 50 unbranded prompts per engine for a brand-level number, 100 or more if you want to compare segments or engines, two to three runs per prompt per engine in each reporting window, and windows of two to four weeks. More prompts beat more runs of the same prompt.

The reason is noise. In the same AI Rank Lab dataset, 18 of 42 prompt-engine pairs (42.9%) flipped between mentioning and not mentioning the brand at least once, and two consecutive runs of the same prompt shared a median of only 41% of their cited domains. A mention rate is a proportion, so its uncertainty follows the standard interval in the NIST Engineering Statistics Handbook. Using distinct prompt-engine pairs as the sample size, a 50% mention rate carries a 95% margin of about +/- 13.9 points at 50 pairs, +/- 9.8 at 100, +/- 6.9 at 200 and +/- 4.9 at 400. A month-on-month move smaller than the margin is not a finding.

Published guidance on set size varies by a factor of ten, and almost none of it comes with reasoning:

The small numbers are fine for exploration: 25 prompts will tell you which competitors keep appearing and which sources engines lean on. They are not enough to report a percentage and claim it moved. The reconciliation with SparkToro's advice to ask "at least 60-100X" is that their figure is about knowing one prompt's full recommendation set; for a portfolio number, spreading runs across many prompts is more efficient. The stability study has the full runs-versus-margin tables.

Step 7: How often should you update your prompt set?

Freeze the tracked set for each reporting period, run discovery monthly, and change the tracked set in one versioned batch per quarter, starting a new baseline each time. Editing prompts mid-period is the most common way teams fake a trend without meaning to.

1. Monthly discovery. Re-run the evidence ladder: new sales-call themes, new Search Console queries, new grounding queries, and any brand the engines named that is not on your competitor list. Add candidates to a backlog, not to the live set.
2. Quarterly change window. Promote backlog prompts, retire weak ones, and record the change in a log with the date, prompt, action and reason.
3. Add, never edit. A reworded prompt is a new prompt. Retire the old one and keep its history, otherwise its trend line silently changes meaning.
4. Re-baseline. Report the first period after a change as a new baseline, or report the old and new sets side by side for one period.
5. Break the cycle only for real events: a product launch, entering a new market, a rebrand, or a competitor that suddenly appears in a large share of answers.

At each quarterly review, a prompt keeps its place only if it passes these checks:

• It traces to evidence (ladder rows 1-8) or has proven itself since it entered as a hypothesis.
• It maps to a product you sell and a buyer you want.
• Someone owns an action for it. Similarweb suggests rebuilding the set if more than half the prompts have no specific action attached.
• It is not a near-duplicate of another prompt in the same cluster.
• It still produces a useful signal. A prompt where you have been at 0% for a quarter is a content opportunity, not a candidate for deletion; a prompt where every brand is always present may be safe to retire.

Once the set is frozen, the runs themselves can be repeated in one tool: AI Rank Lab's AI visibility tracking re-runs the same prompts on demand across ChatGPT, Claude, Gemini and Perplexity, with separate trackers for Google AI Overviews and AI Mode, and records mentions, position, sentiment, competitors and cited URLs for each run.

What is prompt volume, and how accurate is it?

Prompt volume is an estimate of how often people ask AI assistants a given prompt or topic. It is always modeled. As Ahrefs notes, apart from Microsoft Copilot no major AI platform shares query data, so vendors build estimates from consumer panels, clickstream data or Google search volumes. Treat the number as a relative signal inside your own prompt set.

For transparency: AI Rank Lab's AI Prompt Research also shows an estimated monthly intent volume next to suggested prompts. It is a rough figure produced by the AI model during generation, not a panel or clickstream measurement, so the same caution applies to ours as to everyone else's.

How to use prompt volume without being misled:

• Rank, do not total. Use volume to order prompts within one cluster or compare topics. Ahrefs warns that summing topic volumes within a prompt can produce overinflated results.
• Evidence beats volume. If a question shows up in sales calls every week and a tool says "low volume", keep the prompt. Panels under-sample niche B2B buyers.
• Compare tools only on trends. Different methods give the same prompt very different numbers. Watch direction within one tool, not levels across tools.
• Remember most AI use is not search. The NBER working paper How People Use ChatGPT (September 2025) found practical guidance, seeking information and writing make up nearly 80% of conversations, and Semrush measured web search switched on for 34.5% of ChatGPT queries in February 2026. Only part of any volume figure is the kind of question where brands get recommended.

Can you get real prompt data from Google, Bing or ChatGPT?

Not the prompts themselves. Each platform gives you a partial view, and together they are still the best evidence you have.

• Google Search Console: the Generative AI performance report shows impressions in AI Overviews and AI Mode by page, country, device and date, but no queries. Our Search Console generative AI report guide walks through it. Use it to see which pages already earn AI impressions, then use the regular Performance report with the non-branded filter to find the queries those pages rank for.
• Bing Webmaster Tools: grounding queries are the closest first-party view of AI query language. Bing describes them as a sample, and they reflect how the AI retrieved your page, not the user's words.
• ChatGPT, Claude, Gemini, Perplexity: none provides site owners with query logs. What you can see is referral traffic: pages that already receive AI visits tell you which topics work, and Ahrefs suggests building prompts backwards from those landing pages.
• Your own customers: the "what did you ask the AI?" field on a lead form is the only source of real AI prompts most companies will ever get. Add it.

Worked example: a 65-prompt set for a payroll software company

Here is how the method plays out for a hypothetical mid-market payroll software company selling to US businesses with 50 to 500 employees. The set has 50 unbranded prompts (14 problem, 20 category, 10 comparison and alternatives, 6 purchase and objection) that feed the headline metric, and 15 branded prompts (10 brand evaluation, 5 post-purchase) reported separately. Run on four engines, the unbranded set gives 200 prompt-engine pairs, which puts a 50% mention rate at about +/- 6.9 points. Twelve representative rows:

Notice what the set does not contain: no prompt uses the company's own category name, no prompt is phrased to flatter the product, every prompt names its evidence, and the two competitor-vs-competitor prompts sit in their own cluster so they cannot inflate the category number.

Common prompt research mistakes

• Pasting the keyword list in. Three-word keywords miss the context that changes AI answers. Convert them.
• Letting an LLM write the whole set. Generated prompts skew short and commercial. Use them for coverage, then check them against real evidence.
• Blending branded prompts into the headline. It can double your apparent visibility.
• Reading single runs. With 42.9% of pairs flipping in our data, one run per prompt is an anecdote.
• Editing prompts in place. It breaks the trend line without anyone noticing.
• Deleting prompts where you score 0%. Those are often the most valuable prompts in the set: they show where competitors win.
• Treating prompt volume as a market size. Even the vendors who sell it say not to.

Tools that help with prompt research

Disclosure: AI Rank Lab is our product. We list it first because this is our site, and we've kept the same criteria for every tool below. We compared what each tool contributes to discovery and selection, not overall platform quality, and did not compare prices here.

Limitations of this method

• A prompt set is a sample of your design, not of real conversations. No tool sees what real users typed or were shown, and memory, location and conversation history change answers.
• The evidence ladder is a judgment. We rank sales-call questions above search data because they are verbatim buyer language, but they come from a small, skewed group of people who already found you.
• Funnel shares and the 50-prompt floor are recommendations. They are informed by the studies above and the margin math, not proven by our 42-pair dataset on its own.
• Vendor studies have conflicts of interest. Otterly, Peec, Profound, SE Ranking, Similarweb and Nightwatch all sell tracking, and so do we. SparkToro's study was run with Gumshoe.ai, which also sells AI tracking, and was not peer reviewed.
• Features change fast. Search Console, Bing Webmaster Tools and vendor details were checked on October 18, 2026.

The bottom line

Prompt research decides what your AI visibility numbers mean. Start from what buyers actually asked, convert it into the longer, more personal language people use with AI, tag every prompt by stage, persona and branded status, keep branded prompts out of the headline metric, size the set so changes beat the margin of error, and change it only in versioned batches. Treat prompt volume as a tiebreaker, not a foundation. If you want the discovery and the repeated runs handled in one place, AI Rank Lab's brand visibility tracking drafts prompts from your site and keywords and runs them across six AI surfaces, so your time goes on checking the evidence and acting on the gaps.