Which Prompts to Track for AI Search Visibility

Which prompts to track for AI visibility comes down to one rule: track a representative sample, not everything a buyer could type. Cover category prompts that never name your brand, brand prompts that do, and competitor prompts, across every buyer stage. Start with 25 to 50, pick them from real sales evidence, and review them monthly.

That sounds simple, and the hard part is the picking. AI engines have no keyword volume, no fixed rankings and no two identical answers, so the usual SEO habit of tracking whatever has the biggest number does not work. This guide gives you a method to choose prompts, a scorecard to rank them, a worked example, a sizing and cadence table, and a checklist you can run this week. If you are new to the discipline, start with what generative engine optimization is.

Key takeaways

• Split your prompt set into three groups: category (unbranded), brand, and competitor. Report them separately, because mixing them distorts every metric.
• Map each prompt to a buyer stage: awareness, consideration, evaluation or decision. Put most of your effort where buyers are still deciding and your visibility is uncertain.
• Begin with 25 to 50 prompts, then expand only when a month of data shows which topics need depth.
• Source prompts from closed-won and closed-lost deals, sales calls, support tickets and FAQs, not from brainstorming alone.
• Collect answers daily, judge trends monthly, and change the prompt list quarterly so you are not reacting to noise.
• Score every candidate prompt on revenue proximity, uncertainty, stability and coverage before you commit to it.

Why is choosing prompts harder than choosing keywords?

Keyword tracking has a stable unit: a phrase, a volume and a rank. Prompt tracking has none of those. SE Ranking's guide to choosing prompts notes that prompt tracking has no volume data, no ranking positions and no static results, which is why it argues that choosing the right prompts matters more than tracking more of them (SE Ranking, 2026).

The second problem is that answers move. In a study where 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times, SparkToro reported a less than 1 in 100 chance that ChatGPT or Google's AI would give the same list of brands in any two responses, and closer to 1 in 1,000 for the same list in the same order (SparkToro, 2026). A single run of a single prompt tells you almost nothing.

SparkToro's finding is not a reason to stop tracking. It is a reason to track many runs of a deliberately chosen set, and to read the share of answers that include you, not whether you were first in one answer.

The same study found that visibility percentages were still meaningful. For example, when Google's AI was asked to recommend digital marketing consultants with e-commerce expertise, one agency appeared in 85 of 95 responses (SparkToro, 2026). Repeated sampling of a well-chosen prompt gives a usable signal, even when ordering is random.

What are the three kinds of prompts you should track?

Every useful prompt set is built from three groups. Semrush frames them as branded, category and competitor prompts (Semrush, 2026), and HubSpot uses branded, unbranded and competitor (HubSpot, October 2026). The labels differ, but the logic is the same and it is worth being strict about it.

The shares in the last column are our own suggested starting split, not a published benchmark. They reflect one idea: HubSpot argues that branded prompts should show the brand anyway, so most effort belongs where visibility is uncertain and buyers are still deciding (HubSpot, October 2026).

Why category prompts carry the most weight

Category prompts are the only ones where a buyer has not already chosen you. HubSpot describes unbranded prompts as the clearest measure of whether answer engine work is reaching new buyers (HubSpot, October 2026). If your brand is missing from them, the brand prompts will look healthy while the pipeline of new buyers stays thin.

Why you must report the groups separately

SE Ranking advises tracking brand comparison prompts separately because they skew category metrics if mixed in (SE Ranking, 2026). A competitor prompt that names you will almost always mention you, which inflates an overall visibility figure. Keep three dashboards, or three filtered views, and never average them.

How do buyer stages change which prompts matter?

Prompts at different stages answer different business questions. Semrush recommends tracking both awareness-stage and decision-stage prompts because each tells you something different about your position in the customer journey (Semrush, 2026). HubSpot notes that decision prompts are narrow and packed with constraints such as budget, integrations, team size or contract terms, and that their answers usually name a few vendors, which shapes the shortlist (HubSpot, October 2026).

SE Ranking lists five prompt types: informational, comparative, instructional, brand-specific and transactional, and says most brands over-index on comparative prompts and ignore the rest (SE Ranking, 2026). As a rough check of our own, if more than half your candidate list starts with a comparison, you probably have a hole in awareness and instruction.

Buyers also rarely stop at one question. HubSpot recommends representing follow-up intent, for example narrowing with questions about which option is cheapest (HubSpot, October 2026). Add a handful of these so your set reflects a conversation, not only a first question.

How should you word a prompt?

Write each prompt the way a buyer would type it to an assistant: a full sentence, with the constraint that matters. A buyer does not type a bare keyword; they ask for something for a person, a budget or a situation. Prompts built on real phrases from sales calls and support tickets will match how your market actually speaks.

• Include one constraint. Skin type, team size, budget, country or integration turns a generic prompt into a decision prompt.
• Keep one intent per prompt. A prompt that asks for price, quality and availability at once is hard to score and hard to fix.
• Avoid your own jargon. If customers never use a term in tickets or reviews, do not use it in the prompt.
• Name the market. Add a country variant where products, prices or regulations differ.
• Freeze the wording. Once a prompt is in the set, do not edit it, or your trend line will break. Add a new prompt instead.

How do you choose prompts to track, step by step?

This is the method we recommend. Each step says what to do, what you get, and why it works. It draws on the published guidance above, and the four-factor scorecard in step 5 is Nurdd's own framework.

Step 1: Mine revenue evidence for raw prompts

Action: Collect 100 to 150 candidate questions from closed-won deals, closed-lost reasons, sales call notes, support tickets, on-site search, reviews and your FAQ. Result: a long list written in buyer language, not marketing language. Why: HubSpot recommends starting with closed-won deals, then mining closed-lost reasons and competitors, and weighting toward high-value segments (HubSpot, October 2026). Prompts that come from deals connect visibility to revenue.

Step 2: Add research-tool prompts you would not think of

Action: Check a prompt research source for topics your list misses. Semrush's Prompt Research Report is described as keyword research for the AI era, helping you discover and prioritise the prompts people use in AI search (Semrush, 2026). Ahrefs' Brand Radar says it already tracks brand mentions across more than 400 million search-backed prompts, which maps the breadth of your AI funnel (Ahrefs Academy, 2026). Result: new angles for your long list. Why: your own team's vocabulary is the biggest blind spot.

Step 3: Tag every candidate by group, stage and persona

Action: Add three columns: group (category, brand, competitor, follow-up), stage (awareness to decision) and persona or market. Result: a table you can filter. Why: seoClarity describes segmenting prompts by topic, persona, buyer-journey stage and brand versus non-brand intent so you can answer questions such as how you perform for one audience or stage (seoClarity, 2026).

Step 4: Cluster variations into one representative prompt

Action: Group near-duplicates, such as five ways of asking for the best option for sensitive skin, and keep the clearest wording as the tracked prompt. Result: a shorter list in which each prompt stands for an intent. Why: Semrush notes that clustering simplifies reporting, so you report on 12 intent categories instead of 80 individual prompts (Semrush, 2026).

Step 5: Score each cluster with the four-factor scorecard

Action: Score each cluster from 1 to 3 on the four factors in the table below, then add them up. Result: a ranked list with a clear cut-off. Why: it forces a choice between a prompt that is merely interesting and one that can move revenue.

Keep clusters scoring 9 or more, and hold 6 to 8 on a reserve list. These cut-offs are our own. The uncertainty factor is borrowed from HubSpot's point that effort belongs where visibility is uncertain (HubSpot, October 2026). Prompts where you always win are monitored, not worked on.

Step 6: Size the set to your question

Action: Pick a set size using the next section. Result: a final list you can afford to review. Why: a list nobody reads is worse than a small one.

Step 7: Set engines, markets and cadence

Action: Choose which AI engines your buyers use and which countries matter, then decide how often to collect and review. Result: a tracking plan. Why: HubSpot advises choosing engines by goals and focusing on priority markets (HubSpot, October 2026). Our guide to Perplexity vs ChatGPT vs Gemini explains how engines differ for a brand.

Step 8: Review quarterly, and prune

Action: Every quarter, add prompts for launches, new competitors and new customer language, and remove dead ones. Result: a set that stays aligned with the business. Why: Semrush suggests a quarterly review that adds prompts for a launch or new category and retires prompts that generate no signal (Semrush, 2026).

How many prompts should a brand track in ChatGPT and other engines?

The honest answer is that the number depends on the question you are asking. One published guide separates three jobs: 5 to 10 prompts checked by hand for a diagnosis, 10 fixed prompts run nightly on every AI service for a trend, and the rest monthly for coverage (Be Discoverable, 2026). Semrush suggests around 25 to start, enough to cover core branded, category and competitor prompts (Semrush, 2026), and HubSpot recommends 25 to 50 (HubSpot, October 2026).

Those numbers agree more than they appear to. Hand checks diagnose, a core set of 25 to 50 trends, and expansion follows evidence. The table turns that into a decision.

The per-engine question matters too. The count applies to the prompt list, not to each engine separately: a tool that runs one list across several engines multiplies the answers you collect without multiplying the list you maintain. Noma by Nurdd tracks 13 AI engines, including ChatGPT, Gemini, Perplexity and Claude, polled daily, so a 40-prompt list yields a large sample of answers for you to read at the stage and group level.

Why not track 1,000 prompts?

You can, but the reporting gets hard to act on. HubSpot warns against the trap of tracking everything and says a small starting set spots patterns without creating noise (HubSpot, October 2026). Breadth is useful for discovery and for finding gaps, and depth is useful for decisions. Track a small set for decisions and use a research source for breadth.

What does a worked example look like?

Here is an illustrative example, with a fictional brand and no real performance numbers. Imagine a direct-to-consumer mineral sunscreen brand for sensitive skin, sold in the US and the UK, with two main rivals. The team has 118 raw questions from support tickets, reviews and sales notes, and wants a 40-prompt starting set.

1. Cluster. The 118 questions collapse into 22 intent clusters, such as reef-safe options, sunscreen under makeup, and white cast on darker skin.
2. Score. The team scores each cluster on the four factors. Reef-safe and white cast score 11 and 12, because they appear in lost deals and reviews. A cluster about the history of SPF scores 5 and is dropped.
3. Allocate. 40 slots: 16 category, 10 competitor, 6 brand, 4 follow-up and 4 reserve slots for the second month.
4. Split by market. Six of the category prompts get a UK variant, since wording and available products differ.
5. Baseline. The team runs the set for a month, then reads results by group and stage, not as one number.

The allocation looks like this.

Our own write-up of asking ChatGPT, Perplexity and Gemini to recommend a sunscreen shows what category prompts look like in practice for a single product category. The month-one read-out is simple. If the brand appears in many brand prompts but few evaluation prompts, the problem is discoverability, and the fix is content and sources, not messaging.

How often should you track and review prompts?

Collection should be frequent and decisions should be slow. HubSpot recommends daily collection because answers vary even on the same day for identical prompts, monthly reporting to smooth that variation, and holding bigger decisions such as rewriting pages or replacing prompts for monthly or quarterly reviews (HubSpot, October 2026). It also notes that a decline is sometimes industry-wide, in which case waiting beats acting.

If you want a week-by-week routine, our guide to tracking your AI search visibility score week over week covers the reporting side, and the weekly ChatGPT tracking experiment shows what movement looks like in practice.

What should you read in the results?

Read answers at four levels, in this order. First, mention rate: the share of answers for a prompt group in which your brand appears. Second, citation: whether your pages are linked as a source. Third, description: whether the engine states facts about you correctly. Fourth, competitor presence: who appears when you do not.

Semrush suggests validating a prompt set with brand mention and citation frequency, and prioritising prompts where competitors appear but you do not (Semrush, 2026). Those are the gaps with the clearest fix. For ideas on the fix itself, see why your competitor is cited and you are not and the sentence structure that gets quoted.

A decision rule for each prompt after a month

• You appear often, cited: keep it as a monitor, review quarterly.
• You appear rarely, competitors often: this is a priority gap. Compare the sources the engines cite, then improve or create the page that answers the prompt.
• Nobody relevant appears: an open space. Publish a clear, quotable answer first.
• You appear but described wrongly: fix the source pages and your brand description. See our guide on writing a brand description AI engines can use.

What are the most common mistakes when choosing prompts?

• Tracking only branded prompts. They flatter you and say nothing about new buyers.
• Writing prompts like keywords. Buyers type full questions with constraints. Write them that way, or use real phrases from sales calls.
• Judging one run. SparkToro's data shows individual lists vary heavily, so read rates over many runs (SparkToro, 2026).
• Mixing groups into one score. Competitor prompts that name you inflate the average.
• Reacting to every dip. Volatility can be industry-wide, as HubSpot notes (HubSpot, October 2026).
• Never pruning. A stale set tracks last year's market.
• Ignoring the page behind the prompt. A prompt is only winnable if a page exists that answers it. Our guide to checking your site for AI answers helps here.

Checklist: build your first prompt set in one afternoon

1. List your three or four most valuable products, segments or markets.
2. Pull 100 or more raw questions from closed-won and closed-lost deals, support tickets, reviews and your FAQ.
3. Add prompts from a research source such as a prompt research report or a competitor's FAQ.
4. Tag each by group, stage and persona or market.
5. Cluster near-duplicates into one representative prompt.
6. Score each cluster on revenue proximity, uncertainty, stability and coverage.
7. Keep the top 25 to 50 and hold the rest on a reserve list.
8. Check the mix: roughly half category, a quarter competitor, a small share brand and follow-up.
9. Choose engines and markets, and confirm answers are collected daily.
10. Set a monthly review and a quarterly prune, and write both in the calendar.

Before you start, the GEO readiness scorecard gives a 12-factor check of whether your pages are ready to be quoted, so that the prompts you choose are winnable. The AEO audit checklist for D2C brands goes deeper on page-level fixes.

How does Noma help you track the right prompts?

Noma by Nurdd is an AI search visibility dashboard at nomahq.app. It tracks whether a brand appears in answers from 13 AI engines, including ChatGPT, Gemini, Perplexity and Claude, polled daily. It shows citation frequency by category, competitor comparison and recommendations for improving AI visibility, so a prompt set built with this method becomes a trend line you can act on.

Book a demo at nurdd.club/demo. Bring your draft prompt list and you will see how your own brand and your rivals show up across the engines, where the category gaps are, and which recommendations Noma gives for them.

Noma has a 3-day free trial. A card is taken at signup and billing starts on day 3, so you are not charged if you cancel before then. If you are still comparing tools, see the comparison of AI search visibility tools, or the head-to-head pages for Noma vs Profound and Noma vs Semrush.

Which prompts suit different business types?

The method is the same, but the mix shifts by industry. A software company will weight evaluation and decision prompts about integrations and pricing, as on our B2B SaaS industry page. A skincare brand will weight ingredient, concern and routine prompts, as on the beauty industry page. Local services care most about location prompts, and marketplaces care about comparison prompts.

For published category ranges, our AI search visibility benchmarks article collects what research shows. Use it to sanity-check expectations, not as a target.

How we researched this

We searched for the primary keyword and read the top guides on prompt selection, including HubSpot's, Semrush's and SE Ranking's, and compared their sections, sizing advice and cadence advice. We also reviewed a published SparkToro study on answer consistency, help pages from Semrush and Ahrefs, and practitioner guides from seoClarity and Be Discoverable. All sources were checked on 29 October 2026.

We did not run hands-on tests, interview customers or use private data. The worked example is a fictional illustration with no performance figures. Every third-party statistic comes from a linked source; the scorecard thresholds, candidate counts and group shares are our own framework.

Frequently asked questions

Which prompts should I track for AI visibility?

Track a representative set across three groups: category prompts that do not name you, competitor prompts, and brand prompts. Spread them over buyer stages from awareness to decision, and add a few follow-up questions. Choose them from closed-won and closed-lost deals and support questions, not guesses.

How many prompts should a brand track in ChatGPT?

Most guides we reviewed start at 25 to 50 prompts per product line or market. A smaller set of 5 to 10 checked by hand works for a first diagnosis. The count applies to your list, and a tool that polls many engines will multiply the answers collected from it.

How do I choose prompts for AI search tracking?

Collect raw questions from sales and support, tag them by group and stage, cluster duplicates, and score each cluster on revenue proximity, uncertainty, stability and coverage. Keep the highest scorers. Review the set every quarter.

Should I track branded prompts at all?

Yes, but as a smaller share. Brand prompts show how engines describe you and whether the facts are right. They rarely show growth, because the buyer already knows you.

How often should I check my prompts?

Collect answers daily with a tool, look for anomalies weekly, and judge trends monthly. Change the prompt list quarterly. Answers vary between runs, so acting on a single day's result is a common error.

Can I use keyword research to pick prompts?

It is a good source of topics, but not a direct translation. Keyword volume does not exist for prompts, and buyers write longer questions with constraints. Turn keywords and FAQs into full questions, then score them like any other candidate.

Do I need different prompts for different AI engines?

Usually no. Keep one buyer-phrased list and run it across engines, then compare the engines in your results. Add engine-specific prompts only if your buyers use an engine in a distinctive way.

When should I remove a prompt?

Remove it at your quarterly review if it has produced no mention or citation signal, if the product it refers to is gone, or if another prompt covers the same intent better. Semrush likewise suggests retiring prompts that generate no signal.

Ready to track the prompts that matter?

You now have a method: mine revenue evidence, tag, cluster, score, size, schedule and prune. The next step is seeing your own brand against it. Book a demo of Noma and we will show you where your brand appears across 13 AI engines, where competitors appear instead, and what to fix first. You can also browse the free Nurdd tools or the blog for more guides.