Updated on 19 October 2026, by João Leite, Founder of murmur.

Build the question list (prompts) for measuring brand presence in ChatGPT from the doubts the buyer has before buying. A prompt is the question or request a person writes to the AI. Presence is how many answers the brand appears in, divided by the total number of answers. Each question is written the way the buyer writes it and annotated with its type, the purchase stage and the reason it is on the list. Questions with your brand name form a separate list, and the market list only changes by addition, with a date.

  • The questions come from the doubts the buyer has already left in plain sight: with the sales team, in customer service, in Google search and in forums.
  • Each question goes in whole: people do not reduce a request to the AI to 2 to 5 keywords, writes Rand Fishkin, of SparkToro.
  • The list covers the purchase stages, with problem, category, comparison, alternative and price questions.
  • The question with the brand name stays out of the presence calculation: it shows what the AI knows about the company, in a list of its own.
  • The list grows by addition: a new question enters with a date, and a question that leaves gets an exit date and stays on file.

What the question list needs in order to measure brand presence

A list that measures brand presence gathers buyer questions, written the way the buyer writes them and without your brand name. The questions do not change from one round to the next, so that the difference between two measurements comes only from the answers. A round is each time the whole list is run again, and within it each question is asked more than once, on different days. The list measures what AI models respond to those questions, not how many people ask each one.

murmur is a GEO agency in Brazil that places brands in the answers of ChatGPT, Claude, Google AIO and other AI agents. On Google, the AI answers in the AI Overview, the summary that opens many searches, and Google AIO is its short name. On 2 September 2026, murmur took 228 questions from the Fly Vet market to ChatGPT, a veterinary marketing agency and a client of the firm. None of them cites Fly Vet by name. In a list like that, every answer that brings up Fly Vet shows that ChatGPT remembered it on its own.

How to find the questions your customer asks ChatGPT

Find the questions your customer asks ChatGPT through the doubts they already bring to the sales team and to customer service. The most valuable source is the buyer’s own sentence, kept in emails, WhatsApp conversations and meeting notes. Copy each doubt the way it came, without correcting the grammar or swapping their word for your team’s word. If the company started measuring because of a screenshot of an AI answer, the question from that screenshot goes on the list word for word. The path from the screenshot to the list is in the guide on why ChatGPT does not recommend your company.

Google autocomplete, the suggestions that appear while a person types, shows what people search about a subject. According to Google Search help, these suggestions reflect real searches made on Google. The suggestions change with the language, the place, rising interest in a subject and the previous searches of whoever is typing. So run the query without logging in, in an incognito window. Use the suggestions to find subjects, not to copy the question. A Google search tends to be short, and the request to the AI comes in whole.

Forums and communities show the doubt before it reaches your team: Reddit threads, comments on LinkedIn and YouTube, industry groups. ChatGPT itself also gives clues, because it tends to offer a next step after answering. These offers show questions it links to the subject, and each one becomes a candidate for the list when it matches a buyer doubt. None of these sources shows the exact text each person types into ChatGPT. So the list covers the doubts, written as close to the buyer as possible.

How to choose and write each question on the list

Choose one question per buyer need and write each one as a full sentence, with the context the buyer puts in. Rand Fishkin, of SparkToro, an American audience research firm, writes that people do not reduce a request to the AI to 2 to 5 keywords. For him, asking the AI is not like searching on Google: people write in a creative, strange and very specific way. “Sistema restaurante” (“Restaurant system”), in the buyer’s mouth, becomes “qual sistema eu uso num restaurante pequeno que também faz delivery” (“what system do I use in a small restaurant that also does delivery”).

Each question on the list corresponds to a buyer need, not to one more way of asking the same thing. In a SparkToro test with headphones, requests written in very different ways for the same need brought back a similar group of brands. When Fishkin asked for headphones for gaming, for podcasts and with noise cancellation, the answers were very different.

Choose the questions before the first measurement, based on what the buyer asks, not on the answers in which your brand already appears. A list built after seeing the answers shows the brand strong where it is already strong and hides the questions where it is missing. The comparison with competitors also comes out skewed, because the list starts to favor one side.

The question types, from the problem to choosing the supplier

The list covers the purchase stages with five question types: problem, category, comparison, alternative and price. Each type catches the buyer at a different moment, from the annoyance they cannot yet name to switching suppliers. In a company that sells a system for restaurants, the five look like this:

  • Problem: the question from someone who is not yet looking for a supplier, such as “como controlar o estoque do restaurante sem planilha” (“how to control restaurant inventory without a spreadsheet”).
  • Category: the search for the type of solution, such as “melhor sistema para restaurante com delivery” (“best system for a restaurant with delivery”).
  • Comparison: two suppliers side by side, such as “sistema X ou sistema Y para restaurante, qual é melhor” (“system X or system Y for a restaurant, which is better”).
  • Alternative: switching suppliers, such as “alternativa ao sistema X para restaurante pequeno” (“alternative to system X for a small restaurant”).
  • Price: how much it costs and how it is charged, such as “sistema para restaurante cobra mensalidade ou taxa por pedido” (“does a restaurant system charge a monthly fee or a fee per order”).

The question type says what the buyer wants to know, and the purchase stage says where they are. The stages run from discovering the problem to getting to know the solutions and, finally, choosing the supplier. A list of choice questions only measures the brand in front of those already deciding and ignores those still looking for the type of solution. Peec AI, a monitoring tool that sells this feature, classifies each question as informational, commercial or transactional. These are the names SEO gives to search intent.

murmur saw two clients’ brands start to appear in questions from people who did not yet know them. At Mindo, a motion design studio in São Paulo, ChatGPT started recommending Mindo also in general questions, such as best animation studios. The buyer asks these questions before having any name in mind. At Visio, retail software, the questions that unlocked things were searches for alternatives to and competitors of other systems, and general questions about retail management. Without questions of these types, the two clients’ lists would not have shown these gains.

What to note for each question: the list spreadsheet

Each question takes one row of the spreadsheet, with number, exact text, type, purchase stage, business line, list, reason and dates. An ordinary spreadsheet is enough, with one tab for the list and another for the answers. In the table, the example row is from a company that sells a system for restaurants.

columnwhat to note downexample
numbera number that never changes and is never used again14
questionthe exact text, as the buyer writes it“qual sistema eu uso num restaurante pequeno que também faz delivery” (“what system do I use in a small restaurant that also does delivery”)
typeproblem, category, comparison, alternative or pricecategory
purchase stagediscovering the problem, getting to know the solutions or choosing the suppliergetting to know the solutions
business linethe product line or the audience of the questionrestaurants with delivery
listmarket, without the brand name, or with the brand namemarket
reasonwhy the question was added, written before the first measurementa question that delivery clients bring to the sales team
entry and exitthe date the question was added and, if it left, the exit dateadded on 2 October 2026

The number of each question links its answers across every round, and it is never used again, not even when the question leaves. The reason, written before the first measurement, makes it possible to check later whether the list came from the buyer’s questions or from favorable results. The type, purchase stage and business line columns show what the list already covers and what is still missing.

In the answers tab, each answer from each AI model becomes a row, with the date, the brand, the competitors and the sources cited. The brand’s presence, the main KPI, comes from this tab. A KPI is the number a company chooses to track to know whether a piece of work is delivering results. The six columns of this tab and the other KPIs are in the step-by-step guide to measuring brand presence in ChatGPT.

Questions with the brand name: a separate list, to see whether the AI understands the company

Questions with your brand name go on a separate list, which shows whether the AI understands the company. This list answers what the AI knows about the company: what it does, who it sells to, how much it charges and whether the information is correct. Brand presence in the market is counted only with the questions without the name. The answer to a question that already mentions the company almost always talks about it. Adding the two lists together inflates brand presence, and it is one of the points to check in any presence number.

The list with the brand name takes every question that carries the company name, alone or alongside a competitor. It includes questions such as “o que a empresa X faz” (“what does company X do”), “a empresa X é confiável” (“is company X trustworthy”) and “empresa X ou empresa Y” (“company X or company Y”). A question that mentions only the competitor stays on the market list, as a comparison or an alternative. The same list shows whether the AI confuses your company with another one of the same name.

Peec AI, a monitoring tool, also classifies each question as branded or unbranded. When the list with the brand name finds wrong information, the fix starts with the sources the AI read. Every step of that correction is in the guide to wrong information in ChatGPT.

How to change the list and keep measuring the evolution of brand presence

To measure the evolution of presence, change the list only by addition: a new question enters with a date, and the one that leaves gets an exit date. The question that leaves stays in the spreadsheet, with the answers it had, because the numbers of the old rounds were calculated with it. The comparison between two rounds uses only the questions present in both, and a new question counts from its second round onward. After any change, an extra round right away becomes the new starting point for comparisons.

The list changes when the business changes: a new product, a competitor that appeared, a new question from the sales team. The question about a product the company no longer sells leaves, with the date. Replacing the whole list at once erases the comparison, because the next measurement no longer has any question in common with the previous one.

At murmur, the question list guides the content and stays open to the client

At murmur, each piece of content comes from a gap measured in the question list, not from a publishing calendar. A gap is a question on the list where the AI answers without bringing in the client’s brand. The questions come from the sector study that murmur carries out as part of the work, and the same set goes back to the AI models week after week. The questions are sent to the AI models by a murmur server in Brazil dedicated only to this task, which never changes its internet address and uses a Portuguese-language browser, not a team member’s computer. For that reason, no answer comes out marked by the city of whoever is measuring, and rounds from different weeks can be compared with each other.

The murmur client follows the whole list in the dashboard, which comes with the service. Next to each question appear the purchase stage and the number of competitors competing for the answer. With the list in view, the client checks whether the questions are the ones their buyer would ask and whether they cover every purchase stage. The following rounds show, question by question, whether the brand started appearing where it was missing.

Frequently asked questions

How many questions does the list need to have?

From dozens to hundreds, and the number for your market comes from the business lines and the purchase stages the list needs to cover. A bigger list also makes the number steadier: ten times more questions, 100 instead of 10, brought the variation in measured presence down by close to 73%, in a September 2026 study in which Otterly.AI, whose business is selling AI monitoring, tracked questions daily across seven AI models. Repeating the same question on different days does not replace a new question on the list. The guide to list size shows why ten questions are not enough and how much less presence fluctuates on a bigger list.

Can I ask an AI to write the list?

It depends on the review: the AI serves as a first draft, and the final list is only valid after you read it, question by question. Rand Fishkin, of SparkToro, considers that visibility measured across many questions should be a reasonable approximation of how brands appear in real answers. For him, this holds both for questions written by people and for those generated by AI. Monitoring tools help too. Peec AI suggests questions from the brand’s context, and Otterly.AI has a prompt research tool. In the review, each question has to match a doubt the buyer actually has and earn its reason for being on the list.

Does a question that mentions a competitor go on the list?

It does, on the market list, as a comparison or alternative question, because it does not carry your brand name. It is in these questions that you see who the AI recommends when the buyer already has another supplier in mind. What to do with these answers, starting from the sources the AI cited, is in the guide on the AI that cites the competitor.

Can I change the wording of a question after the first round?

Not in place: the corrected version enters as a new question, with another number and date, and the old one is removed with an exit date. A single changed word can already change the AI’s answer, and that is why the answers to the new version are not compared with those of the old one. The round in which the new version enters is its starting point.

Start with the questions your sales team already hears

Start with the questions clients asked the sales team last month, copied verbatim, with a number on each one. While this list grows, the free diagnostic from murmur gives you a starting point, with no sign-up and in less than two minutes. With your website address, it builds questions for your market and brings what the AI models answered to each one. The sector studies from murmur show which brands the AI models recommended in various markets, each study with the questions from a single round.

Read also

References