Prompts to Test Your Brand's Visibility in AI Search

A good prompt set for testing your brand in AI search has 20 to 30 questions that do not mention the brand name, split by decision stage (discovery, comparison, decision), by persona and by city, plus a few control questions. Each prompt should run several times, in new conversations, with everything logged. The math on how many answers are enough comes further down, along with ready-made templates.

Quick summary

• Test questions a customer would ask without knowing you. Prompts that include the brand name measure knowledge, not recommendation.
• Cover three stages (discovery, comparison, decision), at least two personas and the cities where you sell.
• Include controls: one question about your own brand, one about a competitor and one that checks for description errors.
• Run each prompt 3 to 5 times in a new conversation, with the same configuration, and log the date, tool, search mode and the raw answer.
• With 150 answers and an observed rate of 20%, the plausible range is still about 14% to 27%. Few answers cannot support fine-grained conclusions.

If you have never run a visibility test, start with the short walkthrough in how to find out if AI recommends your brand. This post goes deeper on the questions and the protocol, within the landscape described in SEO, AEO and GEO.

How to read this post: if you are not technical, copy the templates from the prompt bank section and use the run protocol. If you are a professional, focus on the controls, the minimum-sample table and the log format.

How do people really ask AI?

In a longer and more contextual way than in traditional search. Ahrefs reports, in its evidence review on LLM search (September 2025), that ChatGPT prompts average 42 words and that about 75% are commands, not questions. That changes the test: "best accounting" is a weak prompt; "I need to start an LLC for my freelance design work in Austin and want an affordable online accountant, what options exist?" looks like what someone actually types.

Beyond length, there is variety. In SparkToro's study on answer consistency, 142 people each wrote, in their own way, a prompt asking for headphone recommendations, and the average semantic similarity between the texts was low (0.081), according to Search Engine Journal's coverage. Even so, some brands appeared in 55% to 77% of answers. In short: different people ask in different ways, and strong brands show up anyway.

In plain English: there is no single "the question" the customer asks. There are dozens of ways to ask the same thing. That is why the test uses several versions, not just one.

To understand how these questions behave, see how people search in AI and query fan-out, the mechanism by which one question becomes several searches behind the answer.

How do I build the prompt bank?

Build it in layers and combine them. Each one answers a different question.

Layer	What it measures	Prompt template
Discovery	Whether the brand enters when the customer is still exploring the problem	What are the options for [category] for [persona] who [need]?
Comparison	Whether the brand appears next to competitors	Compare the best [category] options for [persona] considering [criterion 1] and [criterion 2].
Decision	Whether the brand is recommended when the choice is close	Which [category] do you recommend in [city] for [specific situation]? Explain why.
Local	Whether the brand appears in the cities where it operates	Where can I find [service] in [neighborhood/city] with [attribute]?
Persona	Whether the recommendation changes depending on who asks	The same decision prompt, swapping [persona] (beginner, small business, manager)
Brand control	What AI knows and whether it is correct	What do you know about [your brand]? Does it serve [type of customer]?
Competitor control	The comparison yardstick	What do you know about [competitor]?

The first four rows form the core, the persona row is a variation applicable to any of them, and the last two are controls that do not enter the recommendation rate. Keep them separate.

For SEO professionals: include variations of format (short question, long command, prompt with a budget constraint), of intent (informational, commercial, transactional) and of language/region when relevant. To decide the intent cut, draw on what you already have in your keyword research and Search Console queries. Prompts derived from real queries are more defensible than invented ones.

How many prompts of each type?

An initial split that works for 24 prompts: 8 discovery, 6 comparison, 6 decision and 4 local, plus 3 controls. Adjust it to your business: a national e-commerce store can skip the local ones, and a neighborhood clinic can have half the list be local prompts.

Hypothetical example: the bank for an online accounting firm

Hypothetical example: "Faro Accounting," a fictional online accounting firm for freelancers and small service businesses, serves the whole United States and has an office in Austin. The bank below shows eight of the 24 prompts.

Layer	Prompt
Discovery	What are the options for a freelance designer to set up an LLC and keep the books in order?
Discovery	I need an accountant for a service business with 2 partners and up to $10,000 a month in revenue. What kinds of services exist?
Comparison	Compare online accounting firms and traditional offices for a small service business, considering price and support.
Comparison	Which online accounting firms are best for someone with little time and no experience with taxes?
Decision	Which online accounting firm do you recommend for a sole proprietor who is about to form an LLC? Explain why.
Local	Accounting for small businesses in Austin with support by text message.
Persona	I own a small med spa and want to switch accountants. Where do I start and which options should I consider?
Control	What do you know about Faro Accounting? Does it serve med spas?

Note three things: no recommendation prompt mentions the brand; the personas appear inside the text, the way a real person would write; and there is a control that checks a specific claim ("does it serve med spas?"), which is where errors tend to show up.

How do I run the prompts without contaminating the result?

Always follow the same protocol, so the results are comparable. The steps:

1. Open a new conversation for each run. Context from earlier conversations influences the answer.
2. Fix the configuration: the same tool, the same selected model, the same web search state and the same memory/custom instructions state. Note each one.
3. Paste the prompt exactly as it is in the bank. Changing one word creates a different prompt.
4. Run each prompt 3 to 5 times. Spread the runs across different times or days instead of repeating them back to back.
5. Keep the raw answer (text and links), in addition to your notes. You will want to reread it when you change the classification rule.
6. Classify each answer with fixed rules: mentioned (yes/no), with a link to your domain (yes/no), described correctly (yes/no/partial), and which competitors appeared.

A simple log format, in CSV:

date,tool,model,web_search,layer,prompt_id,run,brand_mentioned,domain_link,description_correct,competitors,raw_answer
2026-10-04,chatgpt,default,yes,decision,D3,2,0,0,,"Competitor A|Competitor B",<pasted text>

For SEO professionals: keep the prompt_id stable and versioned. If you change a prompt's text, create a new ID. Otherwise the historical series mixes different questions and the trend stops meaning anything.

How many answers do I need to draw conclusions?

More than it seems. SparkToro suggests that visibility percentage across dozens to hundreds of prompts run multiple times is a reasonable metric, according to the original study. The table shows how uncertainty shrinks with the number of answers, assuming an observed mention rate of 20% and the Wilson interval method, at 95% confidence:

Answers	Mentions	Plausible range of the true rate
30	6	9.5% to 37.3%
60	12	11.8% to 31.8%
150	30	14.4% to 27.1%
300	60	15.9% to 24.9%

Two important caveats. First, the numbers assume each answer is independent; repeats of the same prompt are not independent of each other, so the real uncertainty tends to be larger. Second, the range reflects sampling only: it does not include the effect of having picked the wrong prompts. Details and formulas are in how to measure AI visibility.

In plain English: if you observed 6 mentions in 30 answers, the truth could be anywhere between "almost 1 in 10" and "more than 1 in 3." Only celebrate or worry about changes bigger than that range.

Which mistakes ruin the test?

The five most common:

• Naming the brand in the recommendation prompt. AI answers about the brand because you mentioned it.
• Prompts that already contain the answer, like "why is [category X] the best?". They lead to the conclusion.
• Only one prompt per intent. A single prompt is one person's opinion, not a pattern.
• Mixing configurations (with and without search, logged in and logged out) without noting it.
• Concluding from list position. SparkToro itself considers "ranking position in AI" an invalid metric; count presence, not order.

To choose where to automate all this, compare the options in tools to monitor AI citations. Contentor's AI Visibility runs, for the questions you define, the check of whether ChatGPT cites your brand. It covers ChatGPT only.

Frequently asked questions

Can I ask ChatGPT to generate the test prompts for me?

You can, as a starting point, but review them. Models tend to generate questions that are too "clean," unlike the ones customers type. Better sources: Search Console queries, questions asked to support, complaint wording in reviews and sales conversations. Use AI to vary the wording, not to decide what matters.

Do I need to test in more than one language?

Test in the language and region of your customers. If you serve only the United States, English is enough. If you sell abroad, create a bank per language, because brands and sources change between them. Do not mix results from different languages in the same rate.

How many times should I repeat each prompt?

3 to 5 times is a practical minimum. More important than repeating one prompt a lot is having many different prompts. Twenty questions run 5 times give more information than 5 questions run 20 times, because they cover more ways of asking.

Do the control prompts count toward the rate?

No. Questions with the brand name or a competitor's name measure knowledge and description errors. If they enter the calculation, they inflate the rate. Keep them in a separate column or tab and track the change independently.

How often should I redo the bank?

Review the bank every quarter, but keep the same prompts between measurements so the series stays comparable. If you swap prompts, preserve a fixed core and add the new ones with new IDs. Always record the date and the configuration.

Next step

With the bank ready and the protocol defined, the next step is turning the logs into indicators. In how to measure AI visibility: metrics, share of voice and attribution, you learn to calculate mention rate, share of voice and margins of error, and to tie it all to traffic.

Read also:

• How to find out if AI recommends your brand
• Tools to monitor AI citations
• Query fan-out: why one question becomes several searches