AI Prompt Panel 2026:
How to Build, Tag and Run Prompts That Measure AI Visibility

Why does AI visibility tracking need a prompt panel?

Because AI engines publish no query data, you have to pick the questions yourself. A prompt panel is that pick. It is a fixed set of real buyer questions you run on a schedule, so the numbers mean the same thing each week.

In SEO you track a keyword and get a rank. AI search does not work that way. A buyer types a long question. The engine writes a fresh answer each time. SparkToro found there is less than a 1 in 100 chance that ChatGPT or Google AI gives the same brand list twice.

The same study asked people to write their own prompt for one need. Their prompts had a semantic similarity of just 0.081. So there is no single “right” prompt to track. You need a sample. A good sample covers the people, topics, markets and intents that make up your real demand. Understanding how Answer Engine Optimization (AEO) operates begins with this fundamental shift from keyword rank to probabilistic appearance, as documented in our comprehensive AEO guide.

Which brand signals should drive prompt research?

Twelve signals shape a good panel, starting with your brand kit, the topics you want to win, your personas, competitors, markets and buyer intent. Each signal adds a tag to every prompt. The tags are what let you slice results later.

Skip a signal and you get a blind spot. Leave out markets and you never learn you are invisible in Germany. Leave out personas and a CFO question and an office manager question get averaged into one number. The infographic below sorts all twelve into four groups: who you are, what they ask about, who is asking, and why and how they ask. Each signal has one short example.

How many prompts and runs does reliable LLM visibility need?

Size the panel as a product of six numbers: topics, times 5 to 10 prompts per topic, times personas, markets and languages, and buyer intents. Then multiply by the AI engines you track. A startup lands near 45 prompts. An enterprise can pass 10,000.

Bigger brands grow on the same few dials. They win more topics, write more variants per topic, and sell in more markets and languages, on more engines. The other signals stay about the same. Brand kit, competitors, pain points and prompt format shape the wording of each prompt and leave the count alone.

Every prompt and engine pair then needs repeat runs. University of St. Gallen researchers found one run carries a standard error of 0.370 on brand detection. That is close to a coin flip. Seven runs bring it under 0.10. At enterprise scale, many teams run the core set daily and rotate the long tail weekly.

Industry practitioners frame size in similar ways. Scrunch multiplies core topics by 5 to 8 questions each, and its worked example reaches 4,500 prompts. Peec AI suggests 30 to 50 prompts to start. Report share of voice as a range over two to four weeks, and ignore rank. As our empirical non-determinism research shows, the order of brands almost never repeats, which we also verified across our 90-day customer case study of 20 B2B brands. For an empirical evaluation of tracking cadences, panel sizing, and visibility scoring validity, see our fact-check on Prompt Volume, Frequency & AI Visibility Scoring Myths, and specifically our debunk of the Industry-Standard Minimum Prompts to Track.

How do you turn signals into prompts that test AI brand visibility?

Stack two to four signals into one sentence that sounds like a person typing. A persona, a topic, a constraint and an intent make one prompt. AI engines filter on each detail, and that filtering is exactly what you want to measure.

Write the way buyers talk. “Expense software” is a keyword. “I'm a controller at a 200 person SaaS on NetSuite. Which expense tool handles three entities?” is a prompt. Practitioner research from Radyant suggests 70% to 80% of a panel be long, context-rich prompts like that. Test each one two or three times by hand before it goes live, and drop any that return off-topic answers.

Prompt Research in ThriveStackcitedby builds this panel for you from your domain, keywords or topics and tags every prompt by topic, persona, intent and market.

How should search intent shape the prompt mix?

Weight the panel toward the intent closest to revenue. Our default is 20% informational, 40% commercial and 40% transactional, whereas tools like Peec AI default to 25%, 50% and 25%.

• Informational prompts ask why, how and what. They test whether AI treats you as the expert in Retrieval-Augmented Generation (RAG) pipelines.
• Commercial prompts ask for best-of lists, comparisons and alternatives. They test the shortlist.
• Transactional prompts carry pricing, trials, seats or dates. They sit one step from a sale.

A software brand that sells on demos should lean transactional. A publisher that earns from traffic should lean informational. Whatever you pick, keep the mix fixed for at least a quarter. Change it mid-stream and your trend line breaks. For empirical proof on why prompt panels must capture all 5 intent types (including brand and instructional queries) rather than branded keywords alone, review our fact-check on Do You Need Five Prompt Types for Full AI Visibility?.

How should branded vs non branded keywords split in a prompt panel?

Keep 80% to 90% of prompts unbranded, and track branded prompts in their own group. Unbranded prompts measure discovery: does AI bring you up when nobody asked for you? That is the number that grows pipeline.

Branded prompts do a different job. One branded control prompt, such as “Northwind vs Ramp”, checks that the engine knows you at all. A small brand safety set checks what AI says about your pricing, security and reviews. Peec AI advises tracking brand prompts separately so they do not inflate your visibility score, while Radyant research puts brand prompts at 10% to 20% of the panel. For broader structural trust across LLMs, see our study on AI trust signals engines reward.

How do you rank prompts when prompt volume is hidden?

Borrow demand from the signals you already have: Search Console, sales calls, support tickets and panel-based volume tools. OpenAI, Google and Perplexity do not publish what people type, so every prompt volume number is an estimate.

Two kinds of estimate exist in the market: consumer panel extrapolations and estimated intent modeling (e.g., from Otterly.AI or Gumshoe). Both are directional. Use them to rank topics, and let your own first-party data pick the exact wording. For ranking the broader software ecosystem, see our benchmark of the best AI visibility tools ranked for 2026.

How does each AI visibility tool build a prompt panel?

Every major AI visibility tool now helps you build the panel, but each starts from a different signal. Otterly.AI starts from your keywords. Peec AI starts from your website and markets. Gumshoe starts from real prompt data and synthetic user clusters. Scrunch starts from personas and topics. ThriveStackcitedby is the only provider that allows all such combinations on a single Prompt Panel builder (explore Prompt Research →).

They agree on more than they differ. All platforms tag prompts by topic and country. All suggest starting wide, then pruning. Scrunch, for example, advises running prompts for two weeks before cutting any, maintaining a 95% coverage floor while dropping queries that surface redundant domain citations.

Which one is best depends on what you need to prove. Pick a tool that builds prompts from your own brand signals, tags every prompt and runs each one many times, because answers change from run to run. For a few hundred prompts, most tools cost about $300 to $500 a month, and Profound sells by quote. ThriveStackcitedby prices by answers: 325 prompts run weekly on 3 engines cost $363 a month, with a 7-day free trial (prices checked September 2026, see the pricing page).

Which metrics show AI search visibility once the panel runs?

Track mention rate, share of voice, citation rate and sentiment, each cut by the tags you set up. The tags are the payoff. They turn one score into answers like “we win CFO prompts in the US but vanish in Germany.” For marketing leadership reporting frameworks, see our executive guide to AI visibility metrics for CMOs.

Cross-engine gaps are normal. In our research analyzing AI citation sources, we found only 11% source overlap between ChatGPT and Perplexity. Brand mention rates range from about 18% to 50% by country. So always report by engine and by market, never as one blended number.

What does a prompt tracking setup look like step by step?

Five steps: add your brand kit, pick your signals, generate prompts, test and trim, then run daily. Most teams can get a first panel live in an afternoon. The hard part is choosing signals, and that is where a tool saves the most time.

ThriveStackcitedby Prompt Research takes your brand kit, topics, personas, competitors, markets, languages and intent mix in one pass. It returns a tagged panel you can edit before it starts running. You can also analyze crawler pickup in our server logs study or connect full pipeline telemetry on the ThriveStack Platform.

Agencies running panels for many clients can follow the weekly, monthly and quarterly rhythm in our AEO agency workflow guide.