How we measure which businesses AI mentions

Each measurement gathers questions about one trade in one city and records which businesses appear in the answers. This page explains the engines we query, the configuration, the repetitions and how we count mentions.

Results describe the sample collected on the dates shown. Queries go through the providers’ APIs and can differ from the apps. How often a business is mentioned says nothing about the quality of its service.

Agents Explore method, version 1, in force since October 27, 2026, the date of the first published measurement. Every research page states which version it used and when the answers were collected. The full text of version 1 is published in Portuguese at /br/metodo/v1/.

1. What we measure, and what we do not

We measure presence. For each trade in each city, we ask ChatGPT and Gemini the questions people ask when they look for a company, and we count in how many answers each business appeared. The result is the list on each research page: business name, number of answers it appeared in, and by which engine.

We do not measure:

• quality of service, price, lead time or customer care;
• customer reviews;
• whether a business is licensed, registered or in good standing;
• what each person sees in the ChatGPT or Gemini app, which changes with the account, the history and the place the question comes from.

Appearing on the list means the AI mentioned the business in the answers we collected. It does not mean the business is better than another, nor that Agents Explore endorses it. AI assistants build answers from what they find on the web and what the model learned, and that comes with errors: wrong names, companies from another city, old phone numbers, businesses that have closed. We publish what the AI answered, with a date, not what the business is.

2. Engines and models

A measurement uses one model per engine. The model name and version are recorded with the measurement and shown on each page. Gemini with search (grounding) is allowed by its terms only to be shown to the person who asked, not stored or published; that is why it appears only in live checks, never on public pages. Other AI assistants that read the web are not measured in this version.

3. Configuration

• Every question is a new call: no previous conversation, no personal account, no memory, no history.
• Language of the published measurements: Brazilian Portuguese, written as a person would type it, with no system instruction and nothing that asks for, avoids or favors any business.
• ChatGPT: search available, approximate city location, provider defaults for generation. Gemini: no search and no location.
• Every answer is stored in full, with date, time, requested model, responding model and the configuration sent.

4. The 10 questions

Each trade and city combination receives the same 10 questions, in the same template, changing only the vocabulary of the trade for what people use when they search. The intent of each position is the same across trades. The template, with the variables in braces (the published measurements use the Portuguese wording):

1. Who does {service} in {city}?
2. Name {service} companies in {city} with good reviews.
3. Which is the best {service} company in {city}?
4. I need {service} in {city}. Give me 3 options with a phone number.
5. {service} company in {city} for {specialty A}.
6. {service} company in {city} for {specialty B}.
7. How much does {service} cost in {city}, and who does it?
8. {service} near me.
9. I am going to ask for {service} quotes in {city}. Which companies should I ask?
10. Which {service} companies in {city} have a website and answer on WhatsApp?

Specialties A and B are the two most searched specialties of the trade, measured by monthly search volume. Question 8 does not name the city; the city goes only in the location passed to the API, so it runs only on ChatGPT. Questions that name a specific business never enter research pages; they exist only in live checks, labeled as “recognition”, and the result is shown only to the person who asked.

5. Repetitions and calendar

Each question runs more than once, on different days, because the same question can bring different names from one day to the next. Repeating shows what holds. The rule is 3 times, on 3 different days (hypothesis). A measurement is a complete collection cycle; each page shows the month, the start and end dates of the collection and the number of valid answers. Measurements older than 60 days leave the search index until the next collection (hypothesis).

6. How we count

• Valid answer: the API returned a complete answer. Network errors, timeouts and provider limits are retried up to 5 times; an answer that never arrives does not count.
• Mention: a business name in the answer text, counted once per answer, as the AI wrote it.
• Name resolution: variants become one name only with evidence that it is the same business (same website, phone or address); a person reviews every merge. Same names of different businesses stay separate.
• Type: local, chain or franchise, and portals (quote sites, aggregators and apps), which are listed apart because they are not businesses of the trade.
• Order: by the number of answers a business appeared in; ties in alphabetical order. Payment, profile or plan never enter the count.
• Reading the number: “appeared in 12 of 57 answers” (example). Each engine has its own column, and the columns are not added up.

7. Corrections

Any business mentioned can ask for a correction or removal of its data, with a protocol and a deadline. Requests go to contato@agentsexplore.com. Corrections are recorded on the page with the date.