Build an AI Visibility Prompt Set

A Prompt Set Is A Research Instrument

A buyer-prompt set is a controlled sample of questions that potential customers may ask while defining a problem, choosing an approach, comparing options, and checking risk. Its purpose is to reveal patterns in generated answers. It is not a list of keywords with question marks added, and it cannot represent every prompt real users type.

Start by writing the decision the set should observe. 'Understand whether mid-market finance teams encounter our brand while evaluating invoice automation' is testable. 'Track AI visibility' is too vague. Define the audience, market, language, platforms, product category, competitors, and the business decisions the report will inform.

Keep the core set small enough to rerun consistently. Thirty to fifty well-chosen prompts, repeated under controlled conditions, usually teach more than hundreds of near-duplicates run once. Add an experimental set for new questions without rewriting the baseline whenever an interesting idea appears.

Research question: what decision or perception will this set observe?
Scope: audience, country, language, product category, platforms, and search setting.
Core set: stable prompts used for trend reporting.
Experimental set: emerging prompts tested without altering the baseline.

Collect Buyer Language Before Writing Prompts

Use evidence from sales calls, customer interviews, support tickets, onboarding notes, community discussions, product reviews, internal-site search, and paid-search query reports. Look for the situation around the question: company size, role, urgency, existing workflow, constraints, and fear of a bad choice. Those details make conversational prompts different from short search queries.

For each source question, record the underlying job and stage. 'Can this connect to NetSuite without engineering help?' is a compatibility and implementation-risk question. 'Best AP automation software' is a recommendation question. They may concern the same category, but they invite very different answers and should not be grouped as interchangeable variants.

Remove personal data, confidential customer details, and one-off scenarios that do not represent the intended segment. Do not paste private call transcripts into public AI products unless your organization's data policy explicitly permits it. The prompt set needs authentic language, not identifying information.

Capture the original wording, source type, buyer role, stage, and decision being made.
Keep constraints that materially change the answer, such as region, budget, company size, or required integration.
Merge true duplicates, but preserve questions that test distinct risks or selection criteria.
Mark assumptions supplied by the research team so they are not mistaken for customer evidence.

Map The Set Across Six Decision Clusters

A balanced set follows the buyer's decision rather than overloading the most commercially flattering prompt. Use six clusters: problem discovery, category and approach, selection criteria, recommendations, comparisons and alternatives, and implementation or risk. Add branded diagnostic prompts as a separate seventh cluster when you need to audit factual accuracy.

A practical 40-prompt baseline might contain six problem prompts, six category prompts, six criteria prompts, eight recommendation prompts, six comparison prompts, five implementation or risk prompts, and three branded diagnostic prompts. That distribution is a starting point, not a universal standard. Weight clusters using sales importance and observed buyer frequency before collecting results.

Keep branded prompts out of unbranded visibility rates. A model answering 'What does [Brand] do?' already received the entity name and therefore does not show discovery. Branded prompts are useful for checking positioning, current facts, perceived strengths, and recurring misconceptions.

Problem discovery: 'How can a five-person finance team reduce invoice approval delays?'
Category and approach: 'What types of tools automate invoice approvals for a mid-market company?'
Selection criteria: 'What should I check before choosing AP automation software for NetSuite?'
Recommendation: 'Which AP automation tools suit a 300-person UK manufacturer?'
Comparison and alternatives: 'Compare [Competitor A] and [Competitor B] for multi-entity approvals.'
Implementation or risk: 'What commonly goes wrong when introducing invoice automation?'
Branded diagnostic: 'Who is [Brand] designed for, and what are its main limitations?'

Write Neutral Prompts With Decision-Level Detail

Use a simple construction: buyer role or situation + task + material constraints + requested decision. For example, 'I lead finance at a 300-person UK manufacturer using NetSuite. Which invoice-automation tools should I shortlist if multi-entity approvals and a short implementation are priorities? Explain the trade-offs.' The context narrows the decision without telling the system which brand to choose.

Avoid leading phrases such as 'Why is [Brand] the best?' in discovery measurement. Avoid stuffing a prompt with your product's exact feature language, which creates an artificial fit. Do not instruct the answer to include a fixed number of brands unless list depth is the variable you intentionally want to control. Preserve natural phrasing even when it is less tidy than marketing copy.

Create variants only when they test a real hypothesis. A role variant can show whether the answer changes for a CFO versus an AP manager. A constraint variant can test the UK versus US market. Cosmetic rewrites that retain the same buyer, task, and constraints consume sample budget without adding much information.

Name the situation: role, organization type, or current workflow.
Name the task: learn, shortlist, compare, verify, plan, or troubleshoot.
Add only decision-changing constraints: region, budget, scale, integration, compliance, or time.
Ask for reasoning or trade-offs when you need to understand why brands appear.
Do not name your brand in prompts intended to measure unprompted discovery.

Define The Execution Protocol Before The First Run

The same prompt can produce different answers across platforms, product modes, locations, and runs. Decide whether web search is on or off, then keep that state fixed within a comparison. ChatGPT Search and Claude web search both surface web sources, but their interfaces and retrieval behavior are not interchangeable. Report each platform separately before showing a combined view.

For every observation, store prompt ID and version, exact prompt text, platform, displayed model or mode where available, search state, account state, country, language, date, time, run number, raw answer, and visible source URLs. Start a clean conversation for independent runs unless conversation context is deliberately part of the study.

Repeat each core prompt at least three times per platform for an exploratory baseline. More runs reduce the influence of a single unusual answer but do not turn the sample into platform-wide user data. Score all comparison brands in the same answers rather than issuing a more favorable set of prompts for your own brand.

Use a new conversation for each independent observation.
Keep locale, language, search state, and other known settings consistent.
Run prompt batches within a defined collection window and annotate platform changes.
Store invalid or refused runs and state how they affect the denominator.
Retain raw answers so mention, position, sentiment, accuracy, and citation labels can be audited.

Score, Review, And Version The Set

Write the scoring rubric before seeing the baseline. Define accepted brand aliases, what counts as a recommendation, how unordered prose is handled, which source types are owned or earned, and what makes a factual error material. Have a second person rescore a sample. Disagreement reveals ambiguous rules or prompts that need clearer classification.

Give every prompt a stable ID, owner, cluster, business weight, version, inclusion status, and rationale. Editing one word can change retrieval, so a materially changed prompt receives a new version. Preserve the old version and mark the date it left the core set. Do not splice results from two versions into one trend line.

Review the core set quarterly or when the market changes materially. Retire prompts when the buyer task disappears, not because the brand performs poorly. Promote experimental prompts only after confirming that they represent an important, repeatable decision. Publish additions and retirements in the methods note of the report.

Monthly: check failed runs, source parsing, aliases, and obvious factual changes.
Quarterly: compare the set with fresh calls, tickets, win-loss notes, and category language.
On change: increment the prompt version and annotate the reporting break.
Governance: one owner approves changes; commercial stakeholders can propose but not silently rewrite prompts.

Worked Example: From Buyer Question To Trackable Prompt

Suppose a sales call contains the question, 'Will this actually work with NetSuite across our three subsidiaries, or will implementation become another IT project?' The underlying job is not simply finding software. It combines compatibility, multi-entity workflow, implementation effort, and risk.

Split it into two prompts because each requests a distinct decision. Prompt AP-CRIT-04: 'What should a finance team verify when choosing invoice-automation software for NetSuite across three legal entities?' Prompt AP-REC-07: 'I run finance for a 300-person company with three entities on NetSuite and limited IT capacity. Which invoice-automation tools should I shortlist, and what implementation trade-offs should I expect?' The first measures criteria coverage; the second creates a qualifying recommendation opportunity.

Assign the prompts to their clusters, document the UK or US locale if it matters, set a business weight before testing, and run the same versions across the chosen platforms. If the brand is absent, inspect the stated criteria and cited sources before deciding on an action. The useful finding may be a missing integration page, weak third-party evidence, a true product gap, or simply unstable output. The prompt result diagnoses where to investigate; it does not prove the cause.

Good result: a prompt maps to one decision and has a clear scoring opportunity.
Weak result: a prompt asks for discovery, comparison, implementation, and pricing in one overloaded request.
Honest conclusion: repeated patterns guide research, but no controlled prompt set represents every real user or guarantees a future answer.

How to Build a Buyer-Prompt Set for AI Visibility Tracking

Key Takeaways