Noma by Nurdd (NomaHQ)Noma
Prompts

Prompt Research for AI Search: Group by Intent and Stage

Prompt research for AI search, done by hand: find the questions buyers ask AI, group them by intent and funnel stage, and pick which to track.

Simran Rawat8 min read
Geometric cones in shades of purple and white arranged on speckled stone steps against a purple background.

Prompt research for AI search is the work of finding the questions your buyers actually ask AI assistants, sorting them by intent and funnel stage, and choosing a fixed set to track and write for. You can do it without a volume database: collect, tag, score, then test each prompt on the engines.

Key takeaways

  • Tag every candidate prompt on four fields: funnel stage, intent type, branded or unbranded, and context modifiers such as budget or location.
  • Real prompts are short. In one survey, two-thirds of respondents wrote 15 words or fewer, so write your set the way buyers type, not as long templates.
  • Weight your tracked set towards unbranded Compare and Decide prompts, where a brand can still be added to the shortlist.
  • Score each prompt on commercial value, evidence of demand, current gap and winnability, then track the top scorers and write pages for the gaps.
  • Keep the tracked set fixed for at least a month. Changing prompts mid-test makes any movement impossible to read.

Why did Profound's Prompt Analysis launch matter to buyers?

On 12 November 2026, Profound published its Prompt Analysis launch. Per Profound, November 2026, the feature splits a keyword's conversations across Discovery, Consideration, Purchase and Support, lets you filter to branded or unbranded prompts, and gives example prompts to track (as checked on 2026-11-14). The method in this guide needs no volume database.

The question behind the launch is still the one every team has: which prompts deserve a place in a tracking set that costs money to run? That decision does not depend on one vendor's dataset. This guide gives you a method you can run in a spreadsheet, plus a way to check it against live answers. For background on the terms, see what AEO is and how to choose prompts to track.

Other published takes exist. SISTRIX describes clustering over 62 million real user questions into topics instead of keywords (SISTRIX, August 2026). That explains why prompts are worth grouping. It does not give you a scoring rule for deciding which prompts earn a slot, nor a way to check whether the prompts you picked can be won. The steps below do.

How do I find which prompts people ask AI about my product category?

Start with sources you already own, then widen. Build a raw list of 80 to 150 candidates before you cut anything.

  1. Sales and support language. Pull the exact phrases from call notes, demo requests and support tickets. Buyers repeat the same problems in the same words.
  2. Search Console queries. Export queries that are already questions, or that contain "best", "vs", "price" or "for". These are the closest thing you hold to proven demand.
  3. People Also Ask and autocomplete. Type your category and seed terms and note the question variants.
  4. Community threads. Reddit, forums and review sites show the unpolished version of the question, including the constraints people add.
  5. A prompt discovery tool. The free AI Prompt Finder lists questions buyers ask about a category, tagged by intent, which gives you a starting list to edit.
  6. Fan-out sub-questions. AI engines split one question into several searches. Google's documentation says AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches (Google Search Central, December 2025). List the sub-questions behind each prompt; this query fan-out worksheet shows how.

Write the prompt the way a buyer types it

Marketers tend to write long, tidy prompts. Buyers do not. A Stella Rising survey of 524 active LLM users in January 2026, reported by Search Engine Land, found two-thirds wrote prompts of 15 words or fewer, and about 60% phrased them as questions (Search Engine Land, 2026). The same article reports a median of 8 words when respondents wrote the prompt they would send to buy shoes, and attributes to Otterly.AI a median of 12 words for real prompts. The author discloses being VP of SEO at Stella Rising, and the survey's margin of error is about 4.3%.

So write each core prompt in two forms: a plain version of 6 to 12 words, and one context variant with a budget, location, role or constraint added. The same survey reports that 28% of prompts mention price or budget, 16% are location-based and 32% include personal attributes. Those figures tell you which modifiers to test first.

How do I group AI prompts by intent and funnel stage?

Tag every candidate on four fields. Keep the vocabulary small so two people tag the same prompt the same way.

FieldValuesRule for tagging
Funnel stageLearn, Compare, Decide, UseLearn: the buyer is defining the problem. Compare: weighing options. Decide: price, fit or vendor choice. Use: already a customer, needs help.
Intent typeDefinition, Recommendation, Comparison, Price, How-to, TroubleshootingTag by what the buyer wants the AI to do, not by the first word.
Brand stateUnbranded, Competitor-branded, Own-brandUnbranded means no brand name appears in the prompt.
ModifierNone, Budget, Location, Role or size, ConstraintOne value per prompt. Pick the strongest.

This is deliberately close to the stage logic Profound describes, so your results compare with theirs if you later look at their data. The difference is that you own the tags and the definitions, and you can add a Use stage for post-sale prompts that your support team cares about.

Worked examples

Prompt (illustrative)StageIntentBrand stateModifier
what is a headless CMSLearnDefinitionUnbrandedNone
best headless CMS for a small teamCompareRecommendationUnbrandedRole or size
Contentful vs SanityCompareComparisonCompetitor-brandedNone
headless CMS pricing under $500 a monthDecidePriceUnbrandedBudget
how do I migrate from WordPress to a headless CMSUseHow-toUnbrandedConstraint

The brands in these rows are examples of the tagging method only; they say nothing about how any engine ranks them.

How do I decide which prompts to track and which to write for?

Tracking and writing are different decisions. A tracked prompt costs a slot every day. A page costs hours. Score each tagged prompt from 1 to 3 on four criteria.

Criterion1 point3 pointsWhere the evidence comes from
Commercial valueLearn-stage, no path to a productCompare or Decide, matches a product you sellSales team judgement
Demand evidenceYou invented itAppears in Search Console, sales calls and a community threadYour own data, counted by source
Current gapYou are named in most answersCompetitors are named and you are notA first run on the engines
WinnabilityAnswers cite only analysts or publishers you cannot influenceAnswers cite pages you could publish or earnThe cited sources in each answer

Total of 10 to 12: track and write. Seven to nine: track only, and watch. Six or below: drop it. Demand evidence is the weak spot of any manual method, because you have no volume figure. Count independent sources instead: a prompt that shows up in sales calls, Search Console and a forum is better evidenced than one from a single brainstorm.

Allocate your tracked set by stage

Weight the set towards unbranded Compare and Decide prompts, because that is where a shortlist is still being built. Here is an illustrative split for a 50-prompt set. It is a planning example, not a benchmark.

BucketPromptsPurpose
Unbranded Compare16Are we on the shortlist?
Unbranded Decide (price, fit)10Do engines state our price and fit correctly?
Competitor-branded Compare8How do we appear in head-to-head answers?
Learn8Are we present early, when people define the problem?
Own-brand5Is the description accurate?
Use3Do help answers send people to our docs?

Noma by Nurdd's Starter plan includes 50 tracked prompts, and the trial includes 25, so a split like this fits a first project. Check current limits on the pricing page.

How do I test a prompt before I commit a slot?

Run each shortlisted prompt on ChatGPT, Gemini, Perplexity and Claude, three times each, in fresh sessions. Record three things in your sheet: whether you are named, which brands are named, and which URLs are cited. Engines vary between runs, so a single run proves nothing; how often AI answers change explains the spread.

The cited URLs matter most for the "write for" decision. If the answers keep citing a competitor's comparison page, a review site and a forum thread, your options are a better page of your own, a listing on that review site, or a presence in that thread. If they cite pages you could never influence, lower the winnability score.

Then check whether your candidate page can be read at all. The free AI SEO Checker shows whether ChatGPT, Claude, Perplexity and Google AI Mode can read a page, with its top fixes. Run it on the page you plan to rank for each high-scoring prompt before you rewrite anything; a blocked or script-only page will not respond to better copy. For a deeper audit of where answers come from, see where ChatGPT gets its information.

Where does Noma fit in prompt research?

Noma by Nurdd is the tracking layer after the research. It puts your chosen prompts to 13 AI engines, including ChatGPT, Gemini, Perplexity and Claude, every day, and reports whether your brand is named, how high, and with what sentiment, plus the sources cited and the topics where competitors appear and you do not.

See it on your own brand: book a demo or start the 3-day free trial. A card is taken at signup so billing can start on day 3; cancel before then and nothing is charged.

What mistakes make prompt research misleading?

  • Tracking only branded prompts. You will see your own name and learn nothing about the shortlist.
  • Inventing long, perfect prompts. The survey data above points to short, keyword-like prompts as common. Test both forms.
  • Treating one run as the answer. Run each prompt several times before scoring the gap.
  • Editing the set weekly. Change at most 10% of prompts per month, and log each change with a date.
  • Skipping the Use stage. Post-sale questions shape retention and are easy to ignore.
  • Reading a rank into the result. An AI answer is not a ranked list; judge presence, position and the sources cited, as in the share of voice method.

What should I do this week?

  1. Build a raw list of 80 candidates from sales calls, Search Console and community threads.
  2. Tag all four fields and score each prompt out of 12.
  3. Run the top 30 on four engines, three times each, and log names and citations.
  4. Pick your tracked set using the stage split above, and set it up in Noma.
  5. For each prompt with a gap and a winnable source, run the page through the AI SEO Checker, then write or fix it.
  6. Re-run after four weeks and compare presence by stage.

How we researched this

We reviewed Profound's Prompt Analysis announcement, SISTRIX's prompt research article, Search Engine Land's coverage of the Stella Rising survey, and Google Search Central's AI features documentation. Facts about those pages and all figures were checked on 2026-11-14. Examples labelled illustrative are ours and are not data. We did not run experiments for this post.

Next step

Run the manual method on one product line first. When you want the chosen prompts tracked daily across 13 AI engines, book a demo or start the free trial. For the wider programme, read the AEO guide chapter on choosing prompts.

See what AI engines say about your brand

Free · no account · or start a 3-day trial

Questions people ask

How do I find which prompts people ask AI about my product category?

Collect candidate questions from sales calls, support tickets, Search Console queries, People Also Ask and community threads. Add a free prompt discovery tool such as the Noma AI Prompt Finder for a starting list. Then run each candidate on the engines to confirm it returns a relevant answer.

How do I group AI prompts by intent and funnel stage?

Tag each prompt on funnel stage (Learn, Compare, Decide, Use), intent type (definition, recommendation, comparison, price, how-to, troubleshooting), brand state and one modifier. Keep definitions short so two people tag alike. Profound's Prompt Analysis uses a similar stage split of Discovery, Consideration, Purchase and Support.

How long should the prompts I track be?

Short. A Stella Rising survey of 524 LLM users, reported by Search Engine Land, found two-thirds wrote 15 words or fewer. Track a plain 6 to 12 word version and one version with a budget, location or role added.

Which prompts should I track first?

Start with unbranded Compare and Decide prompts, because buyers are still choosing brands there. Then add competitor-branded comparisons, a few Learn prompts, and a handful of own-brand prompts to check accuracy.

Do I need prompt volume data to do prompt research?

No. Without volume, count independent evidence of demand: a prompt that appears in sales calls, Search Console and a community thread is better supported than one idea. Score it alongside commercial value, current gap and winnability.

How often should I change my tracked prompts?

Rarely. Keep the set fixed for at least a month and change no more than about 10% at a time, logging each change with its date. Frequent edits make it impossible to tell whether the answers moved or the prompts did.

← Back to the blog