# How to Choose Prompts to Track for AI Visibility

> Choose prompts to track for AI visibility with a scoring sheet for buyer intent, funnel stage and engine spread, plus how to prune the list after two weeks.

- Author: Simran Rawat, Growth Executive, Nurdd
- Published: 2026-11-08
- Publisher: Noma by Nurdd (https://nomahq.app)
- Original page: https://nomahq.app/blog/choose-prompts-to-track

To choose prompts to track, list the questions your buyers ask, score each one on buyer intent, funnel stage and engine spread, and keep the top 25 to 50. Run them daily across several AI engines, then prune after two weeks: drop prompts that add no signal and keep the ones that move.

## Key takeaways

- Score every candidate prompt on three things: buyer intent, funnel stage and engine spread. Keep the highest scorers, not the ones you find interesting.
- Start with 20 to 40 prompts. SE Ranking recommends this range and at least 30 days of tracking before drawing conclusions.
- Track unbranded category prompts separately from prompts that name your brand. They answer different questions.
- Phrase each prompt once and keep the wording fixed. Changing wording mid-test breaks your trend line.
- Review after two weeks. Cut duplicates, dead prompts and prompts the engines answer without naming any brand.

## How do I choose which prompts to track for AI visibility?

Treat it as a selection problem, not a brainstorm. You will generate 100 candidate questions in an hour. Most will tell you nothing you can act on. The job is to cut them to a set that is small enough to read, stable enough to compare week to week, and close enough to buying decisions that a change means something.

The reason to be strict is noise. SparkToro and Gumshoe.ai ran 12 brand-recommendation prompts, each 60 to 100 times per platform, and found nearly every response was unique in its brand list, its order and its length ([Search Engine Journal, January 2026](https://www.searchenginejournal.com/ai-recommendations-change-with-nearly-every-query-sparktoro/566242/)). A single run of a single prompt proves nothing. A fixed set, run repeatedly, can show a pattern.

The same study found something useful for selection. When 142 people wrote their own headphone prompts, the wording barely overlapped, yet Bose, Sony, Sennheiser and Apple appeared in 55 to 77% of the 994 responses. Intent matters more than exact phrasing. Choose prompts that represent an intent, then keep that wording fixed.

> "If you ask an AI tool for brand/product recommendations a hundred times nearly every response will be unique." Rand Fishkin, SparkToro, quoted by [Search Engine Journal, January 2026](https://www.searchenginejournal.com/ai-recommendations-change-with-nearly-every-query-sparktoro/566242/)

For the earlier overview of prompt types, see [which prompts to track](https://nomahq.app/blog/which-prompts-to-track). This guide adds a scoring method and a pruning rule. Chapter 3 of the [AEO guide on choosing prompts](https://nomahq.app/aeo-guide/choosing-prompts) covers the five prompt kinds in more depth.

## Step 1: Build a long list of candidate prompts

Collect 60 to 100 candidates before you score anything. Use four sources, in this order:

1. **Sales and support calls.** Copy the exact questions prospects ask, including the awkward ones about price and switching.
2. **Your search data.** Take Search Console queries that are questions or comparisons and rewrite them as conversational prompts.
3. **Competitor and category pages.** Headings on comparison and "best for" pages are often the question itself.
4. **A prompt research tool.** The [free AI Prompt Finder](https://nomahq.app/tools/ai-prompt-finder) lists questions buyers ask about your category, tagged by intent, which gives you a starting list in minutes.

AI engines publish no query volume, so do not wait for a volume number to decide. [AI prompt volume](https://nomahq.app/blog/ai-prompt-volume) explains why every figure is an estimate and how to size a category without one.

One structural point from Google: AI Mode uses what Google calls query fan-out, which in its words means ["breaking down your question into subtopics" (Google, May 2025)](https://blog.google/products-and-platforms/products/search/google-search-ai-mode-update/) and issuing many searches at once. A single prompt of yours can trigger many hidden sub-queries. Tracking the parent question is therefore enough; you do not need to track every sub-query.

## Step 2: Score each prompt on intent, funnel stage and engine spread

Give every candidate three scores, then add them with the weights below. The weights put buying intent first because a mention on a purchase question is worth more than one on a definition question.

| Criterion | What you score | Scale | Weight |
| --- | --- | --- | --- |
| Buyer intent | Would a person asking this be choosing a vendor or product? | 0 = pure information, 1 = problem research, 2 = category comparison, 3 = "best for", "alternatives" or "which should I buy" | x3 |
| Funnel stage value | How much does your business need to win this stage right now? | 0 = stage you do not sell at, 1 = awareness, 2 = consideration, 3 = decision stage where you already lose deals | x2 |
| Engine spread | Spot-check the prompt on four engines. On how many does the answer name brands at all? | 0 = zero or one engine, 1 = two, 2 = three, 3 = all four | x1 |

The maximum is 18 points (9 + 6 + 3). Keep prompts scoring 12 or more, review 9 to 11, and drop anything below 9.

### Why engine spread is in the sheet

A prompt that only one engine answers with brand names is a narrow test. A prompt where all four engines produce a brand list tells you whether your position is consistent or engine-specific. Engines also disagree: [AI engines disagree on the top brand](https://nomahq.app/blog/ai-engines-disagree-on-top-brand) for many buying topics, so spread shows you where a single-engine view would mislead you.

### Worked example: a project management tool for agencies

The prompts and scores below are an illustration of the method for a fictional vendor. They are not measured data.

| Candidate prompt | Intent (x3) | Funnel (x2) | Spread (x1) | Total | Decision |
| --- | --- | --- | --- | --- | --- |
| best project management software for marketing agencies | 3 = 9 | 3 = 6 | 3 | 18 | Keep |
| alternatives to [market leader] for small agencies | 3 = 9 | 3 = 6 | 2 | 17 | Keep |
| how do agencies track client work across projects | 1 = 3 | 2 = 4 | 2 | 9 | Review |
| what is project management | 0 = 0 | 1 = 2 | 0 | 2 | Drop |
| [our brand] pricing | 2 = 6 | 3 = 6 | 3 | 15 | Keep, in the branded group |

Notice that the generic definition prompt scores 2 even though it is the kind of question many teams track first. Engines answer it from general knowledge, rarely name a vendor, and a mention there would not change a purchase.

## Step 3: Balance the set by funnel stage and prompt type

Scoring picks individual prompts. Balance checks the whole list. SE Ranking suggests a split of 10 to 20 awareness prompts, 20 to 30 consideration prompts and 5 to 10 brand evaluation prompts tracked separately ([SE Ranking, as checked 7 November 2026](https://seranking.com/blog/how-to-choose-prompts-to-track/)). Use it as a sanity check, then shift weight to wherever your sheet says the business is losing.

| Group | What it measures | Typical share of a 40-prompt set | How to read it |
| --- | --- | --- | --- |
| Unbranded category | Whether you enter the shortlist unprompted | 20 to 25 prompts | Your headline mention rate and share of voice |
| Comparison and alternatives | Whether you are named against named rivals | 8 to 10 prompts | Where competitors take your place |
| Problem and use-case | Whether you appear before the buyer knows the category | 6 to 8 prompts | Early-funnel reach |
| Branded | What engines say about you when asked directly | 4 to 6 prompts | Accuracy and sentiment, not discovery |

The shares are a starting rule of thumb from the SE Ranking ranges, not a measured benchmark. Report branded and unbranded results apart. A branded prompt almost always names you, so mixing it in inflates your headline rate.

## How many prompts should I track for my brand in AI search?

For most brands, 25 to 50. SE Ranking says to start with 20 to 40 prompts, across two to three AI models, for at least 30 days. Fewer than about 20 and one noisy prompt moves your average. More than 50 and you spend your time reading rather than acting.

Use your scoring sheet to size it rather than a number alone:

- **One product, one market:** 25 to 40 prompts.
- **Several product lines or two to three markets:** 40 to 50 per line or market, tracked as separate projects so the results stay comparable.
- **Agency client:** 25 to 40 per client, with the same sheet reused for every account.

On Noma by Nurdd, every plan starts with a 3-day trial including 25 tracked prompts, and the Starter plan allows 50 tracked prompts and one project, per the [pricing page](https://nomahq.app/pricing). That fits the range above: use the 25 trial slots for your top scorers.

## Step 4: Write the prompts so they stay comparable

1. Write each prompt the way a buyer would type it, in plain language, with no brand name for unbranded prompts.
2. Include one qualifier that defines the buyer: size, use case or budget. "Best CRM for a 10-person agency" beats "best CRM".
3. Freeze the wording. If you change a prompt, retire the old one and start a new line, because the history no longer applies.
4. Tag every prompt with its funnel stage, group and market so you can slice results later.
5. Run each prompt more than once. The SparkToro data shows why: one run is an anecdote. See [how often AI answers change](https://nomahq.app/blog/ai-citation-stability) for the run counts that hold up.

## Step 5: Prune the list after two weeks

Two weeks of daily runs gives you around 14 answers per prompt and engine. That is enough to see which prompts behave and which are dead weight. It is not enough to claim a trend in your visibility; for that, follow SE Ranking's 30-day minimum. Pruning is about the quality of the prompt, not about your score.

| Pattern after 14 days | What it means | Action |
| --- | --- | --- |
| No brand named in any answer, yours or a rival's | Engines treat it as an information question | Drop it, or rewrite with a buyer qualifier |
| Same competitors, same sources, same order as another prompt | Duplicate intent | Keep the higher scorer, drop the other |
| You are never named but rivals are, consistently | A real gap | Keep, and move to the content backlog |
| You are named sometimes, and it swings between runs | A contested prompt where work can move the result | Keep, and watch first |
| You are named every run on every engine | Defended ground | Keep a few as a canary; retire the rest to free slots |
| Only one engine answers with brands | Narrow test | Keep only if that engine matters to your buyers |

Replace each dropped prompt with the next-highest scorer from your long list, not a fresh idea. That keeps the discipline of the sheet. Re-score the whole list once a quarter, or sooner after a product launch or a new model release.

Noma's own measurements suggest why you should judge sources, not just mentions: in Noma's count of 2,545 answers, [two in three cited URLs appeared in one run only](https://nomahq.app/blog/ai-citation-half-life). A prompt whose cited sources churn every day is a noisy place to measure change.

## Step 6: Turn the kept prompts into work

A prompt list is worth what you do with it. For each kept prompt in the gap or contested rows, note which pages the engines cite, then decide whether the fix is on your own pages or on other people's. Before you rewrite anything, run the target page through the [free AI SEO Checker](https://nomahq.app/tools/ai-seo-checker), which shows whether ChatGPT, Claude, Perplexity and Google AI Mode can read the page, with 30+ checks and the top fixes.

Then re-measure on the same prompts. The [AI Visibility Measurement Standard](https://nomahq.app/resources/ai-visibility-measurement) sets out how to state numerators and denominators and when a change exceeds the margin of error. For share of voice across the set, see [how to measure AI share of voice](https://nomahq.app/blog/measure-ai-share-of-voice).

## Common mistakes when choosing prompts

- **Tracking only your brand name.** You will be named every time and learn nothing about discovery.
- **Copying a vendor's template list.** Their prompts fit their sample category, not your buyers.
- **Editing prompts weekly.** Each edit resets the baseline.
- **Reporting an "AI ranking position" from one run.** Fishkin's published view in the SparkToro write-up is that a ranking position is not a reliable metric, while appearance rate across many runs is steadier.
- **Tracking one engine.** Engine spread is in the sheet because answers differ by engine.

## Where Noma fits

Noma by Nurdd runs your prompt list every day across 13 AI engines, including ChatGPT, Gemini, Perplexity and Claude, and reports visibility, position and sentiment per engine. It shows which competitors are named instead of you and which sources each answer cites, so you can apply the two-week pruning table to real results. [Prompt tracking](https://nomahq.app/capabilities/prompt-tracking) keeps the set fixed so a change in the answer is a change in the world.

See it on your own brand: [book a demo](https://nomahq.app/demo) or [start the 3-day free trial](https://platform.nomahq.app/signup).

## How we researched this

We reviewed vendor guidance on prompt selection (SE Ranking, Semrush), a published research study on answer variability reported by an industry publication (SparkToro via Search Engine Journal), and Google's own description of AI Mode. The scoring sheet, weights and pruning rules are our own method, not a published standard, and the worked example is illustrative. Facts, figures and prices were checked on 2026-11-07.

## What to do next

Build your long list this week, score it on the sheet, and load the top 25 to 50 into a tracker. Set a calendar reminder for day 14 to prune. If you want it running across 13 AI engines without manual checks, [book a demo](https://nomahq.app/demo) or [start the free trial](https://platform.nomahq.app/signup).