How to Build a Prompt Panel That Actually Tracks Your Category

Most AI visibility dashboards come with a list of prompts, and most of those lists fail in the same way: they are built around the brand instead of the category. If a query already contains your company name, the only thing it can prove is that an engine has heard of you. It cannot tell you whether the engine would ever recommend you to a buyer who has never seen your name. A prompt panel closes that gap, and it takes an afternoon to build one.

Why the usual prompt list is a vanity list

The typical starting point is a handful of searches the team already runs by hand: “Brand X reviews”, “Brand X pricing”, “Brand X vs Competitor Y”. Every result looks reassuring, because every query is already about you. The problem is structural, not cosmetic:

• A brand query cannot reveal whether you are absent from category-level answers, which is where most buyers start.
• Ad-hoc queries drift week to week, so nothing is comparable over time.
• Lists built from whatever the team thought of that morning are biased toward the vocabulary of insiders, not customers.
• A single-engine, single-query snapshot conflates one model’s answer with “the market” and invites overreaction.

None of this means brand queries are useless. They belong in a panel, at the end, as a verification layer. They just cannot be the whole panel.

A panel is a fixed, versioned sample

Treat the panel the way a research team treats a survey instrument. It is a fixed set of prompts, stratified by how buyers actually search, sampled on a schedule, and versioned so that every later comparison means the same thing. The goal is not to collect the most queries. It is to collect the same queries often enough that a change is a signal rather than a coincidence.

A panel is not “every query we can think of”. It is 40 to 80 prompts that never change silently, run across every engine you care about, in the languages your buyers use.

Five layers every panel needs

Stratify the panel by intent rather than by keyword volume. Five layers cover almost everything a B2B or SaaS buyer does before a purchase decision:

Layer	What it captures	Prompt patterns
1. Category	Whether engines can describe the category at all, and who they name as its players	“what is [category]”, “how do you [category job]”, “what does [category] mean”
2. Problem	Whether engines reach for you when someone describes a symptom, not a solution	“why is my brand missing from AI answers”, “track mentions in ChatGPT answers”
3. Comparison	Head-to-head positioning against the competitors buyers actually weigh	“[brand] vs [competitor]”, “difference between [brand] and [competitor]”
4. Alternatives	Whether you appear in recommendation lists where no brand is named	“best [category] tools for [segment]”, “alternatives to [competitor]”
5. Verification	Whether the story engines tell about you is accurate and current	“is [brand] legit”, “[brand] pricing”, “[brand] reviews”

Keep at least 70 percent of prompts brand-free. Layers 1, 2 and 4 are where most of the commercial truth lives, and they are the layers a brand-only list never measures. Layers 3 and 5 are deliberate brand entries; keep them, but do not let them dominate the denominator of share of voice.

Four rules that keep the panel honest

• Size it before you grow it. Eight to ten prompts per layer is enough to see direction. A panel of 300 prompts does not buy precision, it buys maintenance burden and a slower weekly read.
• Freeze the text, version the panel. Change a prompt’s wording and you have changed the instrument. Keep v1 intact, add new prompts as v2, and log every edit with a date and a reason.
• Refresh quarterly, not weekly. Replace roughly 20 percent each quarter to keep up with how buyers phrase questions. Weekly churn makes every chart unreadable.
• Match locale and persona. Run each language on its own panel, in the phrasing native buyers use. B2B procurement questions and consumer questions are different instruments, even inside one category.

Read movement, not absolute numbers

A single week’s snapshot is a mood; three weeks is a trend. Watch three columns together, because each one fails in a different way:

• Mention rate: how often the brand appears in answers to a layer of prompts. This is the headline number, and the easiest to inflate with brand queries.
• Rank: where the brand sits when several are named. A drop in rank with stable mentions means competitors are being recommended ahead of you.
• Cited sources: which domains engines rely on for these prompts. This is the diagnostic column; it explains why the other two moved.

Looks like	Usually means	What to do
One week dip on one engine	Model refresh or sampling noise	Note it, wait for the next run
Sustained drop across two or more engines	Content or source gap the engines can see	Check cited sources for the affected layer
Mentions stable, rank slipping	Competitors are being named first	Compare their cited pages against yours
Brand layer fine, category layers empty	Brand queries masking a weak category presence	Rebuild the panel around layers 1, 2 and 4

Segment by engine before you draw conclusions. Engines retrieve and cite differently, and a panel run across several of them will regularly show one engine leading and another lagging for the same prompt. That is information, not contradiction.

Five mistakes that quietly break a panel

• Only brand prompts. It produces a beautiful report about a question nobody asks.
• Hundreds of prompts. Maintenance cost rises faster than insight; the weekly review stops happening.
• Editing prompts mid-quarter. Every comparison after the edit is between two different instruments.
• Ignoring locale. An English panel says nothing about how a Spanish or Japanese buyer is answered.
• Treating one engine as the market. It turns a distribution into a single point estimate and invites panic.

A starter panel you can copy

Layer 1, category: what is [category]; how do I [category job]; is [category] worth it; [category] explained; who provides [category]; [category] trends 2026.

Layer 2, problem: my brand is not mentioned by AI assistants; how to track AI answers about my product; why do AI engines ignore small brands; how to appear in ChatGPT recommendations; best way to monitor AI search visibility; AI visibility for [segment].

Layer 3, comparison: [brand] vs [competitor]; [brand] or [competitor]; is [brand] better than [competitor]; [brand] compared to [competitor] pricing.

Layer 4, alternatives: best [category] tools for [segment]; alternatives to [competitor]; [category] tools under [budget]; top [category] platforms 2026.

Layer 5, verification: is [brand] legit; [brand] pricing; [brand] reviews; [brand] support quality; [brand] case studies.

Fill in the placeholders once, per locale, and freeze the file. That is your v1.

From the weekly read to an actual action

The panel earns its keep only if each movement maps to one of three fixes:

• Content gap: the prompt intent has no page of yours that answers it clearly. Fix the coverage first; structure matters as much as substance, as covered in our guide to quotable content.
• Source gap: engines cite third-party pages for that layer, and you are absent from them. Earn presence on those sources instead of rewriting your homepage again.
• Entity gap: the facts engines repeat about you are inconsistent or stale. Align profiles, pricing pages and directory listings before publishing more content.

Where the panel fits in measurement

The panel is the input layer. Metrics such as mention rate, rank and sentiment are the output layer, and they are only comparable when the input is stable. If you are still deciding what to measure, start with the metrics that actually matter, then come back and fix the prompt set they run on.

Daily monitoring across multiple engines only pays off when the prompts underneath it are stable and representative. A panel built this way turns a dashboard from a scoreboard into a diagnosis tool, which is the point of tracking in the first place. If you want to see how a fixed panel is scanned across engines, compare the monitoring plans and start with the panel above.