AI Visibility Prompts: Which Questions to Track in ChatGPT?

What is a prompt panel, and why does everything depend on it?

A panel is the fixed list of questions you ask AI assistants on a regular schedule to find out whether your brand gets cited. It is your measuring instrument: if the questions are poorly chosen, every number it produces is wrong, no matter how precise it looks.

The scenario is common. The managing partner of an accounting firm types “what does ChatGPT think of [her firm]?” and reads a flattering paragraph. She concludes the AI knows her. The next day, a prospect asks “which accountant for a small LLC in Austin?”: three firms come up, not hers. She measured the wrong thing.

A prospect writes a full sentence to an assistant, not a string of keywords. And a single question can trigger several searches behind the scenes: Google says AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources to build a response (the “query fan-out” technique, described in AI features and your website). So your panel should look like the full sentences your customers write, not like a list of SEO keywords.

Why should you never put your brand in the question?

Because a question that names your brand forces the assistant to talk about it: the mention is guaranteed, so the measurement is worthless. The panel must measure whether your name comes up on its own when a prospect looks for a solution without knowing you.

The same rule applies to your competitors. “What’s an alternative to [competitor]?” automatically inflates that competitor’s visibility on every run and skews the exact comparison you want to make. Read those answers out of curiosity if you like, but they have no place in the panel.

Branded questions do have a use: checking how the assistant describes you (pricing, target customer, strengths) and spotting false information. Track them separately, in a document, and never mix them into your visibility numbers. If you find your competitor shows up everywhere and you don’t, the guide why ChatGPT recommends your competitor helps you find the cause.

What types of questions belong in the panel?

A good panel mixes several angles: the general category, use cases, customer profile, selection criteria, problems to solve and, if you sell locally, the geographic area. Each angle surfaces different brands.

Tonaily’s AI Visibility Index follows this logic: its industry rankings rely on 30 neutral buying questions, split into six families (discovery, use cases, company profile, pain points, features, decision criteria). The 30 questions for each industry are published in full (in English) on the AI Visibility Index, for example for CRM software: a good template to adapt.

One condition for every question: it must be able to get brands cited. “How does payroll work?” calls for an explanation, not a recommendation. “Which tool should a small business use to run payroll?” calls for names.

How should you spread questions across the buying journey?

Spread them across three moments: the prospect discovering the category, the one comparing options, and the one torn between two or three names on a specific criterion. The brands cited change from one moment to the next, and the middle is often where the difference is made.

For a 15-question panel, a split that holds up:

1. Discovery (4 questions): “what are the best…”, “which tool for…”.
2. Consideration (6 questions): use cases and customer profiles, where your offer is strongest.
3. Decision (5 questions): selection criteria and problems, “the cheapest”, “the simplest”, “no commitment”.

To find the wording, don’t start from your internal jargon. Reread quote requests, questions asked in sales meetings, support tickets and customer reviews. The terms your customers use are the ones they will type into ChatGPT. The guide GEO: how to get cited by ChatGPT, Gemini and Grok then explains what to publish to answer those questions.

flowchart TD
  A[Reread quotes and tickets] --> B[Write with no brand at all]
  B --> C[Cover the 3 stages]
  C --> D[Test each question]
  D --> E{"Any brands cited?"}
  E -- No --> B
  E -- Yes --> F[Freeze the panel]
  F --> G[Measure on a fixed schedule]

Should you specify the city, country or language?

Yes, as soon as your customer base is local or national: the question should be asked in your customers’ language and, for a local service, mention the city or region. An answer about “the best plumber” with no location measures nothing for a plumber in Denver.

Location also matters on the tooling side. OpenAI’s web search API lets you specify an approximate user location (country, city, region, timezone) to refine results based on geography. Writing the city into the question makes your measurement less dependent on that setting.

Language changes the brands cited, too. The AI Visibility Index says so itself: its current editions are measured in English, on a global market, and a French edition will rely on questions written directly in French. If you sell in two countries, or in two languages, build two panels rather than one mixed panel nobody can make sense of.

How many prompts should you track?

Between 10 and 20 well-chosen questions are enough for a business with one main offer and one market. Beyond that, the panel gets hard to review, and a question nobody reviews is useless.

The right number depends on your scope, not on a standard. Roughly:

• 10 questions: one offer, one market, one language.
• 15 questions: one main offer and two or three strong use cases.
• 20 questions: several offers or several cities.

For several countries, several languages or a dozen product lines, you need one panel per market: the total quickly climbs to several hundred questions. At that scale, platforms built for large-scale AI tracking are a better fit than a 20-question panel: the comparison of AI visibility tracking tools shows when each one has the edge.

Why does the same question give different answers?

Because text generation is not deterministic: asking the same model the same question twice can produce two different lists of brands. A single answer, positive or negative, proves nothing.

In other words, OpenAI acknowledges that the content a model generates varies from one call to the next. Even when you set the seed parameter, its API reference states that “Determinism is not guaranteed” (Create chat completion). Research backs this up: a study published on arXiv tested five models configured to be deterministic, with 10 runs per task, and found accuracy variations of up to 15% from one run to the next (Atil et al., 2024).

On top of this variability come documented differences between products: Google writes that AI Mode and AI Overviews may use different models and techniques, so the responses and links they show will vary (source). In practice, consumer assistants also factor in the conversation context: when you test by hand, use a session with no history.

Three rules follow:

1. Run it again. A question is measured over several runs, not just one.
2. Read frequencies. “Cited in 6 answers out of 10” tells you something; “cited this morning” tells you nothing. The guide on AI share of voice walks through the calculation.
3. Look at the trend over several weeks, across the whole panel, rather than one question’s score on a given day.

That is also why a panel beats a single question: the research paper that introduced the term GEO evaluates its methods on a benchmark of diverse queries spanning several domains, not on a single query (Aggarwal et al., KDD 2024).

Can you change the questions along the way?

Yes, but as little as possible: a stable panel is what makes measurements comparable over time. Every question you add, remove or rewrite changes what you measure, and a rise can come from the panel rather than from your visibility.

What justifies a change:

• a typo or awkward wording (fix it right away);
• a question that never gets any brand cited, on any engine, after several runs;
• a new offer or a new market.

What doesn’t: a question where you are never cited. That is precisely the information you are looking for. Swapping it for a more flattering question makes the curve look better without changing reality. Note the date of every change so you can read the curve correctly later.

Which questions should you track in your industry?

Start from the templates below, replace the brackets with your category, your target customer and your city, then delete anything that doesn’t match what you sell. Each template covers discovery, use cases and decision.

B2B SaaS (example: project management tool)

• What are the best [category] tools for a small business?
• Which [category] tool for a remote team of [size] people?
• Which [category] tool integrates well with [tool your customers commonly use]?
• Which [category] software is easiest to roll out without an IT department?
• My [category] tool is too expensive, what are the alternatives for a small team?
• Which [category] tool hosts data in Europe?

E-commerce (example: natural cosmetics store)

• Where can I buy [product] [criterion: organic, locally made] online?
• Which [product] brands do you recommend for [need: sensitive skin]?
• What’s the best [product] under [price]?
• Which online stores deliver [product] within 24 hours in [country]?
• Which [product] brand should I choose as a gift?

Agency or local service (example: web agency, plumber, architect)

• Which [service] agency do you recommend in [city]?
• Who should I contact for [specific need] in [city] or [county]?
• Which [service] provider in [city] for a small business?
• How do I choose a [trade] in [city], and which ones are best rated?
• Which [trade] in [city] handles emergencies on weekends?

Training (example: digital marketing training provider)

• What are the best [topic] courses for a beginner?
• Which [topic] courses are eligible for [funding: a government training program, an employer training budget]?
• Which online [topic] course for an employee who works full time?
• Which [topic] training provider in [city]?
• Which [topic] course for a career change?

Consulting firm (example: HR transformation consulting)

• Which [field] consulting firms for a mid-sized company in [country]?
• Which firm supports [project: a transformation plan, a compliance program] in the [industry] sector?
• Which [field] consultant for a small business with [size] employees?
• How do I choose a [field] consulting firm, and which ones are well regarded?
• Which [field] consulting firm in [city]?

Before adding a question to the panel, ask it once to several assistants: if none of them cites a brand, rephrase it. Tonaily’s free AI visibility test puts one question to ChatGPT, Claude, Gemini and Perplexity, no account needed, three times a day: enough to sort through your candidate questions.

How do you validate your panel before you start measuring?

Review each question against the list below. A question that fails even one point gets rephrased or removed before the first run, not after three months of measurements.