Building a Prompt Set: How Many Prompts Do You Need?

A prompt set is the fixed list of questions you run in AI assistants to measure visibility, and its size should follow from the decisions you need to make: enough prompts per topic to read a trend, split by intent, and frozen long enough to compare periods. Most teams start with 15 to 50 prompts and grow from there.

On this page

• What belongs in a prompt set?
• Worked example: Ledgerline's first prompt set
• Where do good prompts come from?
• How do you turn search queries into prompts?
• What intent mix should you aim for?
• How many prompts do you need?
• How do you write prompts that resemble real use?
• How do you check a prompt before adding it?
• How do you keep the set stable?
• How do you tag prompts for analysis?
• Worked example: a 120-prompt starter set
• How do you document the set?
• How do you review the set each quarter?
• From MentionX data: how intent and repeats change results
• How does MentionX approach prompt sets?
• Common questions
• Common mistakes
• How do you know the set is good enough?
• Checklist
• Where to go next
• Sources

Statistics checked against the NIST/SEMATECH handbook on 30 October 2026. "Our view" marks MentionX opinion. Ledgerline is fictional.

What belongs in a prompt set?

Report each group separately. Blending branded prompts with discovery prompts inflates visibility, because branded prompts almost always mention you.

Worked example: Ledgerline's first prompt set

Ledgerline is a fictional accounts-payable tool. Figures are illustrative. The marketing lead collects 140 candidate questions from sales call notes, support tickets, Search Console and two Reddit threads. She cuts them to 36:

Most cuts were near-duplicates or questions nobody on the team could act on yet. She keeps the problem-led group even though she expects low numbers, because that's where buyers first describe their pain. The first month bears this out: Ledgerline appears in most comparison answers and almost none of the problem-led ones, which becomes the content plan for the next quarter.

Where do good prompts come from?

• Sales and support conversations: the exact questions prospects ask.
• Site search and on-site chat logs: real wording from visitors.
• Search Console queries: longer question-style queries often map to AI prompts.
• Community threads: Reddit, forums and review sites show how people describe problems.
• Your team's own AI usage: ask colleagues what they type when researching tools.

See how to do prompt research for a full process.

How do you turn search queries into prompts?

Search Console is one of the quickest sources, because it shows real wording (Google). A simple routine:

1. In the Performance report, filter queries containing words such as "best", "vs", "alternative", "how to" and "for", across the last three months.
2. Export the list and group near-duplicates.
3. Rewrite each group as the full question a buyer would ask an assistant. "ap automation mid size" becomes "What's the best accounts payable automation software for a mid-sized company?"
4. Add context that buyers give an assistant but rarely type into search: team size, country, current tools, budget.
5. Tag the prompt with the original query, so you can later compare AI visibility against search performance for the same topic.

Assistants get longer, more specific questions than search boxes. Keep some short prompts too, because not everyone writes long ones.

What intent mix should you aim for?

Our view, for a B2B brand's core set:

Adjust to your market. A brand with lots of brand-name demand may need more accuracy prompts. A new brand may want more problem-led ones, because that's where it can stand out.

How many prompts do you need?

Precision depends on observations per reported group. Using the worst-case margin from statistical confidence for AI visibility tracking:

Observations = prompts × runs × engines (reported separately per engine). Illustrative: 4 topics, 25 prompts each, 2 runs each = 50 observations per topic per engine. That's enough to tell "rarely mentioned" from "usually mentioned", not to detect a 5-point change.

Two cautions:

1. Distinct prompts beat repeats. Runs of the same prompt are correlated. Prefer 50 prompts × 1–2 runs over 10 prompts × 10 runs.
2. Size the groups you report. If leadership wants a number per product line, each product line needs its own adequate sample.

How do you write prompts that resemble real use?

• Use natural language and full questions, as buyers type them.
• Vary specificity: some broad, some with constraints (budget, team size, industry).
• Avoid leading wording such as "Why is [Brand] the best…".
• Include context where buyers would give it: "for a 10-person agency in Germany".
• Keep a record of exact wording; tiny changes alter results.

How do you check a prompt before adding it?

Run each candidate through five quick tests:

• Would a buyer ask this? If you can't point to a source, such as a call note, ticket or query, it's a guess.
• Is it neutral? No brand names in discovery prompts, and no wording that leads to an answer.
• Is it distinct? If it's a near-copy of an existing prompt, it adds little.
• Can you act on the result? There should be a page, a source or a message you could change.
• Does it need a location or language? If the answer depends on the market, set it explicitly.

Then run it once in each engine and read the answer. If the answer is off-topic, for example the engine reads "ledger" as an accounting textbook, reword it before it enters the set.

How do you keep the set stable?

How do you tag prompts for analysis?

Add columns for topic, funnel stage, prompt type, persona and priority. Tags let you answer questions like "Are we visible for late-funnel comparison prompts in our core product line?" without rebuilding the set.

Illustrative structure:

Worked example: a 120-prompt starter set

Illustrative. A B2B HR software company sells to companies with 50–500 employees and has three product lines: payroll, time-off and onboarding. Leadership wants a monthly number per product line in ChatGPT and Perplexity.

At 60 observations, each product-line estimate has a margin of roughly ±13 points, enough to show whether the company is broadly absent or present. The team agrees to report quarterly trends rather than monthly ones, pooling three months to reach 180 observations per line. Branded prompts are used for accuracy checks, not as a visibility rate.

How do you document the set?

Keep the prompt set in a shared sheet or repo, not only inside a tool. For each prompt, record the exact wording, tags, source, date added, version and the action you'd take on a bad result. Add a short method note: engines, modes, location, language, runs per prompt and schedule. When someone asks in six months why a number moved, this note answers half the question before anyone opens a dashboard. Give the sheet a named owner, and keep a changelog row for every edit to the core set, with the date, the change and the reason. It also makes it possible to switch tools without losing comparable history.

How do you review the set each quarter?

Once a quarter, spend an hour on maintenance:

• Relevance: Do prompts still reflect what buyers ask? Check new sales and support questions.
• Coverage: Are new products, segments or competitors missing? Add them to the explore group first.
• Balance: Is any group too small to report on? Add prompts or pool periods.
• Drift: Has an engine changed behaviour, for example giving shorter answers? Note it beside the trend line.
• Promotion: Move proven explore prompts into the next core version, and restate history only when you have to.

From MentionX data: how intent and repeats change results

Method. Read-only aggregate queries over 54 MentionX scans from 22 August to 29 October 2026 (UTC): 2,154 completed runs across 18 anonymised projects and six engines.

Set size. Scans ran a median of 14 prompts, ranging from 9 to 50.

Intent drives the headline rate.

A set with 30% branded and comparison prompts will show a much higher overall mention rate than one made only of discovery prompts, with no difference in actual visibility.

Repeats aren't independent. We repeated 28 prompt-engine pairs 643 times between 4 September and 19 October. 11 pairs were sometimes mentioned and sometimes not; the rest were stable. Repeats tell you how stable a prompt is, but they add less information than new prompts.

Limits. A convenience sample, mostly B2B software. Intent labels were set by each project and may not be applied the same way everywhere.

Our view: report unbranded discovery, recommendation and problem prompts as the headline. Show comparison and branded prompts beside it, never blended in.

How does MentionX approach prompt sets?

We start from real buyer language, tag every prompt by intent and topic, and keep a versioned core set with a separate explore group. Unbranded prompts drive the headline metrics. Comparison and branded prompts are reported separately as positioning and accuracy checks.

Common questions

Should prompts include the user's context, such as company size or country? Some should, because buyers include it. Keep a mix: broad prompts show category visibility, specific ones show fit for your segment.

Can I use AI to generate prompts? As a starting point for wording variety, yes. Then check each against real buyer language. Generated prompts tend to be tidier and more keyword-like than what people type.

How often should I add prompts? Add new ones to the explore group whenever sales or support hear a new question. Promote them to the core set at the quarterly review, as a new version.

What if my category has few searches? Lean on sales calls, support tickets and community threads. Low search volume doesn't mean buyers aren't asking assistants; they may describe the problem rather than the category.

Should competitors' names appear in prompts? Only in the comparison group. Elsewhere, a named competitor changes the question and skews the result.

How many prompts is too many? When nobody reads the results for some groups. A set that's twice as big but half ignored is worse than a smaller one the team acts on.

How do I handle multiple languages? Write native prompts for each market rather than translating. Report each language as its own group.

Common mistakes

• Choosing prompts the brand already wins, which turns tracking into reassurance.
• Mixing branded and unbranded results.
• Changing wording every month.
• Reporting tiny subgroups as if they were precise.
• Copying keyword lists without rewriting them as questions.

How do you know the set is good enough?

Show it to someone in sales or support. If they recognise most prompts as questions buyers really ask, and can't name an obvious missing one, the set is ready to use.

Checklist

• Prompts sourced from real customer language
• Types separated: discovery, problem, comparison, branded, feature
• Sample sized per reported group
• Core set versioned and frozen
• Tags for topic, stage, type and persona
• Run metadata logged

Where to go next

• Statistical confidence for AI visibility tracking
• Recommendation rate: the metric that matters
• How to do prompt research

Sources

• NIST/SEMATECH, Confidence intervals for proportions
• Google Search Console Help, Performance report
• OpenAI, Prompt engineering guide