peham

CEO notes · Prompt research11 min read

AI prompt research.

A good AI prompt set is small enough to review every month, tagged by intent and topic, and weighted toward questions that can change a buying decision. It is not a dump of every keyword variant. I keep a prompt only if it passes four filters: buying or reputational intent, an answer that names vendors or methods, a source we can create or fix, and a team that will look at it again.

A 69-second walkthrough of the method in this article. Captions are burned in; a text transcript is below.
Transcript

A good AI prompt set is small enough to review every month, and weighted toward questions that can change a buying decision. Here is how I decide what is worth tracking, and what I throw out.

Take two example prompts. A research question: What is an AI hallucination? And a buying question: Which tool provides real prompt volume data rather than estimates? Only the buying prompt goes on my list.

I run every candidate through four filters. Does it have buying or reputational intent? Would the answer name a vendor or a method? Can we actually create or fix a source for it? And will we look at it again on a cadence?

Here is a simple example table. Two of these can change a content decision this month, so I track them. The other two I throw out: one is a research question, and one is too broad.

If a prompt cannot change a content decision or a product-page edit this month, I do not track it. The full playbook is in the article on my site. I'm Thomas Peham, CEO and co-founder of OtterlyAI.

Key takeaways

  • Prompts are not keywords. Buyers brief AI assistants in full sentences, and the engine turns one prompt into several searches.
  • Four filters decide what goes on the list: intent, whether the answer names a vendor or method, whether you can influence it, and whether you'll review it.
  • Prompt volume is always an estimate. AI platforms don't publish counts, so use volume as a direction, not a verdict.
  • Tag every prompt by funnel stage and topic, and keep brand-name prompts separate.
  • A prompt earns its place when a gap on it can turn into an article, a FAQ or a product-page edit this month.

The short answer

AI prompt research is finding the questions your buyers ask AI assistants about your category, then choosing which of them to track over time. It's the prequel to measurement: the prompt list is the denominator of every AI visibility number you'll ever report.

A note on the term, because Google is confused about it too. Search "prompt research" and most results are libraries of prompts for writing research papers with ChatGPT. Here it means the AI-search version: the questions people put to ChatGPT, Perplexity, Gemini, Claude, Copilot and Google's AI Overviews and AI Mode.

My process in one paragraph: collect candidate prompts from sales calls, support tickets, Search Console queries and our own tests in the engines; run every candidate through four filters; tag the survivors by funnel stage and topic; freeze the list as a version; and review it monthly against what we actually shipped.

At OtterlyAI, the result is a fixed set of roughly 500 prompts we track in the US for our own category, the same set behind the worked example in my monthly AI search report.

Prompts are not keywords

A keyword list under-covers AI search because buyers brief assistants in full, specific sentences, and the engine then breaks each prompt into several searches of its own. You need both views, but the prompt is the unit the buyer actually types.

Three differences matter in practice:

Keyword data is still useful as raw material. Your Search Console queries show the language buyers already use; question-shaped and long queries are often the closest thing to a prompt you have.

The four filters I use to decide what to track

A candidate prompt goes on my list only if it passes all four filters. Most candidates fail the first or the third.

1. Does it have commercial or reputational intent?

The prompt must be close to a decision, or about us directly. "What is an AI hallucination?" is a research question: useful to answer, not worth tracking for us. "Which tool provides real prompt volume data rather than estimates?" is a buying question, because the answer is a shortlist.

Reputational prompts qualify too: what we cost, who we're for, whether we're a fit for a use case. Those are where wrong facts hurt most, which I cover in how to stop AI search from getting your company wrong.

2. Would the answer name a vendor or a method?

If a good answer names products, brands or a specific method, our presence in it can be measured and can matter. If a good answer is a definition with no names in it, there is nothing to win except a citation, and I track citations separately.

3. Can we create or fix a source for it?

This is the filter people skip. If no page we own, and no third-party source we could realistically earn, could change the answer, tracking the prompt only produces a number we can't move. I keep prompts where we can write the article, improve the feature page, or earn the review or mention the engines cite.

4. Will we look at it again on a cadence?

A prompt nobody reviews is a cost, not an insight. If a prompt can't change a content decision or a product-page edit this month, I don't track it.

Candidate promptIntentNames vendors or methods?Can we influence it?Track?
"Which tool provides real prompt volume data rather than estimates?"BuyingYesYes (this article, our feature pages)Yes
"Can I organise tracked AI prompts into topics and funnel stages in any platform?"BuyingYesYesYes
"What is an AI hallucination?"ResearchNoHardlyNo
"Best AI tools"Too broadYes, but not our categoryNoNo

"Volume" only helps if you know what was counted

Every prompt volume number is an estimate, because the AI platforms don't publish how often each prompt is asked. OtterlyAI's own Prompt Research page says it plainly: OpenAI doesn't disclose prompt volume data directly, which is why we show an Estimated Intent Score as a direction rather than a count.

That doesn't make volume useless. It means you have to ask what was counted before you trust it:

  • Where does the number come from? A model, a panel of users, search data, or a mix? Each has blind spots.
  • Which engines and which country? A figure built on one assistant in one market says little about another.
  • Is it per prompt or per topic? Phrasing varies so much that per-topic demand is usually more meaningful than per-sentence demand.
  • Does it agree with your other signals? Sales calls, support tickets and Search Console are first-party evidence. If a "high-volume" prompt never comes up there, I'm sceptical.

Google demand is a useful sanity check, and a humbling one. In the US, "ai prompt research" averages 40 searches a month and "prompt volume" 40, according to Google Ads search volume data I pulled on 16 November 2026. Small numbers. I track these questions anyway, because the buyers who ask them are exactly the people who choose a tool, and low Google volume tells you little about how often the same question goes to an assistant.

My rule: volume breaks ties; intent and influence make the list.

How I organise the prompt list

I give every prompt one funnel stage, one topic, and where useful a persona and a source tag, and I keep brand-name prompts in their own group. Tags turn a flat list into a report you can read by topic and stage.

This is the structure we use in our own OtterlyAI workspace:

Tag groupOur tagsWhy it exists
Funnel stageToFu, MoFu, BoFuShows which part of the buying journey we're losing
PersonaAgencies, Marketing TeamsBuyers in each group phrase the same need differently
TopicMonitoring, Analytics, Features, Competitors, AI search experimentsLets us report coverage per theme and hand gaps to the right owner
SourceGSC, Survey, YouTubeRecords where the prompt came from, so we can judge its quality later

Four habits keep the list honest:

  • One funnel stage per prompt. If you can't decide, the prompt is probably two prompts.
  • Separate brand prompts. Prompts that contain your name inflate coverage. I explain why in my monthly AI search report.
  • Version the set. Add new prompts in a "new" group and merge them at a version change, or month-over-month comparisons break.
  • Size it per topic. Single AI answers are noisy. OtterlyAI's guidance is 50 to 100 prompts per topic or market: in our stability study, tracking 100 prompts daily for 30 days cut the variation in brand coverage by about 73% compared with tracking 10.

From prompt list to content plan

A tracked prompt pays off when a gap on it turns into a specific piece of work: a new article, a FAQ, or an edit to an existing page. The list is a backlog, not a dashboard.

I sort every gap into one of three moves:

What the gap looks likeMoveExample
Nobody has a good, specific answer to the questionNew articleBoth of my articles published this week started as prompt gaps: one on fixing wrong AI answers, and this one
A page of ours ranks in Google but isn't cited in AI answersImprove the existing page: answer first, add the missing specifics, update factsA feature or pricing page that already ranks, rewritten so the first paragraph answers the prompt
The answer is right but buried in a long pageAdd a FAQ or a clearly headed section that answers the prompt in its first sentenceThe FAQ blocks at the end of my articles

One warning from Google that applies to every GEO team: don't create a separate page for every prompt variation or fan-out query. Google's AI optimization guide says doing that primarily to manipulate rankings or AI responses violates its scaled content abuse spam policy. One strong page per topic beats twenty thin ones.

After publishing, re-measure the same prompt set on the same engines and country about four weeks later, comparing windows of the same length. For where this fits in the bigger strategy, see my GEO guide and AI search optimization playbook.

Disclosure, and how I do prompt research in OtterlyAI

I'm the CEO and co-founder of OtterlyAI, so weigh this accordingly. OtterlyAI Prompt Research generates prompt suggestions from a seed topic or URL, grouped by intent, funnel stage and topic, includes a query fan-out view, and can turn Google Search Console queries into prompts. Any prompt can be promoted into a brand report and tracked with the AI Prompt Tracker across seven AI engines, daily or weekly. Plans start at $29/month.

The short version: collect widely, filter hard, tag everything, freeze the set, and let every gap become a piece of work. If a prompt can't change what you ship this month, leave it off the list.

FAQs

How do I do prompt research for AI search?

Collect the questions buyers ask about your category from sales calls, support tickets, Search Console queries and your own tests in ChatGPT, Perplexity and Google. Keep only prompts with commercial or reputational intent, where the answer names vendors or methods, where you can create or fix a source, and that you will review on a cadence. Tag them by topic and funnel stage, then track the same set across engines.

How is AI prompt research different from keyword research?

Keyword research measures short search terms typed into Google. AI prompt research maps the longer, conversational questions people ask AI assistants, including follow-ups. Engines then turn one prompt into several searches, so a keyword list alone under-covers what buyers actually ask.

How do I tell which AI prompts have commercial intent?

Ask whether the answer would name a vendor, a product or a shortlist, and whether a buyer could act on it this quarter. Prompts that compare options, ask for recommendations, or ask about cost, fit or setup are commercial. Definitions and general how-it-works questions are research intent.

Do I need real prompt volume data, or are estimates enough?

Estimates are enough for prioritising, as long as you know how they were made. AI platforms do not publish prompt counts; OtterlyAI notes that OpenAI doesn't disclose prompt volume directly. Treat any volume figure as a direction, and weight intent and your ability to influence the answer more heavily.

How should I organise AI prompts into topics and funnel stages?

Give every prompt one funnel stage (top, middle or bottom of funnel), one topic, and optionally a persona and a source tag. Keep brand-name prompts in a separate group. That lets you report visibility per topic and stage and see which part of the buying journey you are losing.

What is a sensible number of prompts for a B2B SaaS team to track?

Enough per topic that single-answer noise averages out, and few enough that someone reviews the list every month. OtterlyAI's guidance is 50 to 100 prompts per topic or market, based on a study where tracking 100 prompts daily for 30 days cut variation in brand coverage by about 73% compared with tracking 10.

Keep
reading.

All writing →
OtterlyAI

See how your brand shows up in AI search.

Track mentions, citations and share of voice across ChatGPT, Perplexity, Gemini, Google AI Overviews, AI Mode and Copilot.

Try OtterlyAI ↗

Talks · advisory · press

Want Thomas on your stage or podcast?

Get in touch →