Building a Representative Prompt Panel: The Complete Method

In brief: A representative prompt panel is built in four stages — broad collection from diverse sources (customer interviews, support, forums, AI auto-suggestions), qualification by volume and business potential, stratification by buying stage and persona, final sizing between 50 and 300 prompts. Quarterly reviews prevent obsolescence. A poorly constructed panel permanently skews decisions — which underscores the importance of methodology. Best practices recommend balanced coverage: 30% TOFU, 40% MOFU, 30% BOFU, with at least three distinct personas and diverse question formats (questions, comparisons, recommendation requests).

The prompt panel is the most valuable asset of a GEO monitoring setup, and probably the most neglected. Many teams start with a list of 30 prompts built in an hour, use it for six months, then are surprised that metrics don't reflect business reality. The cause is almost always the same: an improvised panel measures improvised things.

Building a serious panel takes two to three weeks upfront, then a few days per quarter for review. This initial investment determines the quality of all subsequent measurement. Here's how to approach it methodically.

Stage 1 — Broad Collection

The goal of this first stage is to gather 300 to 500 candidate prompts, without filters. You're looking for volume and diversity; sorting comes later.

Sources are varied. Customer interviews provide natural phrasing — how did your customers describe their problem before seeking a solution? What questions did they ask before buying? Customer support gives access to in-use queries — what language recurs in tickets, chats, emails? Industry forums (Reddit, professional communities, LinkedIn groups) are filled with naturally phrased questions.

Auto-suggestions from the LLMs themselves are valuable. When you type a keyword into Perplexity or ChatGPT, associated questions surface — this is the conversational equivalent of "people also ask." These suggestions reflect questions actually being asked by users.

Finally, input from sales and pre-sales teams rounds out the picture. Questions asked in client calls, demos, and webinars provide formulations often more precise than public sources.

Stage 2 — Qualification

At this point, you have 300 to 500 candidates. Qualification narrows this to 150-300 retained prompts, based on two criteria.

Estimated usage volume. Not all prompts are equal. A question asked by 1,000 buyers per month deserves more weight than one asked by five. Prompt analysis tools (Profound, Otterly, AthenaHQ and others) allow you to estimate these volumes with increasing accuracy. Absent that, the expert judgment of an experienced sales professional provides a usable approximation.

Business potential. A high-volume prompt with no commercial intent ("what is a CRM in IT?") deserves less weight than a lower-volume prompt with strong purchase intent ("which CRM should I choose for a 50-person construction SMB?"). Qualification distinguishes between these two profiles.

At the end of qualification, you have a weighted list. The highest-scoring prompts (volume × business potential) enter the main panel, others go into a secondary panel or passive monitoring.

To build a solid GEO measurement infrastructure, this qualification stage separates amateur programs from professional ones. Rigor here determines the relevance of all subsequent measurement.

---

AI visibility score: test your site Discover if your brand appears in responses from ChatGPT, Claude, and Gemini. Free audit in 2 minutes. Automated paid actions. Launch my free audit

Stage 3 — Stratification

Once the list is qualified, you stratify it along three axes.

Buying stage. TOFU (top of funnel — problem discovery), MOFU (middle of funnel — solution evaluation), BOFU (bottom of funnel — final decision). The ideal split hovers around 30% TOFU, 40% MOFU, 30% BOFU. A panel too heavy on TOFU measures awareness but not conversion; too heavy on BOFU ignores upstream phases where the brand must be present.

Persona. End buyer, technical influencer, user, economic decision-maker. At least three distinct personas, more if the buying journey is complex.

Prompt type. Informational question ("what is…"), recommendation request ("what's the best…"), comparison ("X vs Y"), review search ("what reviews exist for…"), procedural ("how do I choose…"). Healthy diversity in types avoids blind spots.

Stage 4 — Final Sizing

How many prompts should you include in the end? Practical rule: between 50 and 300, depending on company size and market complexity.

For a micro or small business single-market: 50 to 100 prompts suffice. For a small to mid-sized business with multiple segments: 100 to 200. For a multi-brand or multi-geography group: 200 to 500, or more in segmented sub-panels.

Below 30 prompts, statistical variation makes metrics too volatile to be actionable. Above 500, simulation costs explode without proportional marginal gains.

How to Maintain the Panel Over Time?

Quarterly review prevents obsolescence. Three actions at each review:

Remove obsolete prompts — those generating no relevant responses or whose usage has dropped sharply. Add emerging prompts — new formulations spotted in support, forums, AI auto-suggestions. Adjust stratification — if your brand has shifted business focus, TOFU/MOFU/BOFU weighting must follow.

A well-executed quarterly review typically modifies 10 to 20% of the panel. That's healthy. A panel that never changes slowly drifts from reality.

Two Real-World Sector Examples

A project management SaaS editor started with an 80-prompt panel built internally in two weeks. Initial distribution was unbalanced: 60% TOFU, 30% MOFU, 10% BOFU. After three months, the team noticed metrics improving upstream but not on conversion. Panel redesign to 130 prompts with 30/40/30 distribution: three months later, BOFU citation share became measurable and actionable.

An outdoor furniture brand built a panel solely from Google keywords converted to questions. After two months, analysis revealed their customers asked ChatGPT very different questions than they Googled — longer, more contextualized, sometimes centered on specific constraints (pets, climate, space). Panel rebuilt via customer interviews and Reddit listening: the new panel reflected real usage and measurement became actionable.

In summary: a representative prompt panel is built in four stages — broad collection, qualification, stratification, sizing. Sources combine customer interviews, support, forums, AI auto-suggestions, sales teams. Target size ranges from 50 to 300 prompts depending on company scale. Ideal distribution by buying stage hovers around 30/40/30. Quarterly review prevents obsolescence. A poorly constructed panel permanently skews decisions — initial investment conditions the entire quality of subsequent measurement.

In Brief

• Four stages: collection, qualification, stratification, sizing.
• Multiple sources: customers, support, forums, AI, sales teams.
• Target size: 50 to 300 prompts depending on context.
• Ideal split: 30% TOFU, 40% MOFU, 30% BOFU.
• Mandatory quarterly review to prevent obsolescence.

Conclusion

Time spent building the panel is rarely wasted. One day invested here saves weeks later, by avoiding fuzzy metrics, misguided decisions, and indefensible reports. The best GEO teams review their panel like they review a budget — seriously, regularly, with multiple perspectives. This discipline separates programs that last from those that collapse.

---

Are you visible on ChatGPT? Check now Discover if your brand appears in responses from ChatGPT, Claude, and Gemini. Free audit in 2 minutes. Automated paid actions. Launch my free audit