How to test, in 20 minutes, whether AI assistants cite your company
You can find out today, with no paid tool, whether AI assistants cite your company. This article hands over the protocol ReBo uses in the first round of a diagnostic: four questions, five engines, a private session and a seven-column spreadsheet. That is twenty runs of about a minute each.
How do I test whether ChatGPT cites my company?
Open ChatGPT in a private window, logged out, and ask the question your customer would ask, without your company's name. Note whether the brand appears, in what position and which sources the answer shows. Then repeat the same question in the other engines.
The whole test fits in four steps. The first is writing four questions in the customer's vocabulary rather than the company's. Each one tests something different, and the table below shows the four types.
| Question type | Template | What it measures |
|---|---|---|
| Problem | "How do I [solve the problem your company solves]?" | whether your educational content gets retrieved |
| Choice | "Who provides [your service] in Brazil?" | whether the AI lists the company among the options |
| Selection with criteria | "Build a list of three [companies in your sector] with [your customer's criterion]." | whether the AI picks the company when it acts as an agent |
| Brand | "What is [company name] and who are the partners?" | whether the basic facts about the company are right |
Source: ReBo AI Visibility baseline protocol, September 2026.
The first three questions leave out the company's name, because a new customer does not know it yet. The fourth includes it, to check whether what the AI states about the company matches reality.
The second step is running the four questions in five engines: ChatGPT, Gemini, Claude, Perplexity and Copilot. When an engine offers web search, leave it on. The search layer picks the cited pages, and that is what the test wants to observe.
The third step is recording each answer with the same rules. The fourth is reading the result. The two sections below cover each one.
Why does the test have to run in a private session?
Because the assistant personalizes answers with the account's history and memory. Someone who searches for their own company often teaches the engine to show it. A private, logged-out window brings the result closer to what a new customer would see.
Personalization happens in two ways. The first is memory: several assistants keep preferences and topics from earlier conversations. The second is search history, which weighs on what gets shown to someone who has already shown interest in a name.
When an engine requires an account, as Claude does, use an account with no history and turn memory off in the settings. Always run from the same country, too. Answers change by region, and comparing rounds run in different places mixes two variables.
One method caveat applies to the whole test. AI answers are not deterministic: the same question, in the same engine, can bring different sources within minutes. A round is an indicative snapshot, and confidence comes from monthly repetition with the same questions.
What should I record in each round so the result works as a baseline?
Date, engine, exact question, whether the brand was cited, in what position, which competitors appeared and which sources the answer displayed. With those seven fields, next month's round has something to compare against, and the test becomes a time series.
| Field | Example entry | Why it matters |
|---|---|---|
| Date and engine | 09/24/2026 · Perplexity | answers change over time and across engines |
| Exact question | copied, unedited | one different word changes the answer |
| Brand cited | yes or no | the basis of the citation rate |
| Position | 2nd of 4 brands | appearing first weighs more than appearing last |
| Competitors cited | every name, separated by semicolons | without them there is no share of voice |
| Sources shown | the URLs linked in the answer | they show which pages the AI read |
| Evidence | share link or screenshot | lets anyone check the entry later |
Source: ReBo AI Visibility baseline protocol, September 2026.
The definition of "cited" has to be strict. A citation is the company's name written in the answer or a link to its domain. A generic mention of "companies in the sector" counts as no. Loose rules inflate the first round and erase the progress you want to show later.
A filled-in spreadsheet yields two metrics. The citation rate is the number of citations divided by the number of runs. AI share of voice is the company's slice of all brand citations, including those of the competitors you recorded.
My competitor shows up and I don't. What does that mean?
It means the AI found content from the competitor that it could read and use, and did not find yours. The cause is usually technical: a format machines read poorly, a blocked AI crawler, a page with no direct answer or few external mentions.
The sources column shows where the gap is. When the links point to the competitor's site, its edge is first-party content a machine can read. When they point to press, rankings and directories, the edge is external signal: third parties writing about it on readable pages.
There is also the case where no company in the sector appears. In the test ReBo published on September 15, 2026, the same question about Brazilian private credit, asked of five AIs in a private session, returned the same two names in all five answers: XP and Fitch, according to the sources the models declared. A few repeated sources suggest few readable pages on the topic.
The four answers also place the company in the 4-level AI Visibility framework. No citation in the unnamed questions, plus wrong facts in the brand question, is the picture of Level 1, Invisible to AI. The mechanism that decides who makes it into the answer is covered in how ChatGPT chooses the sources it cites.
Keep reading
To understand each metric in the log, read the entries on citation rate and AI share of voice. To see which level the company is at and what changes from one level to the next, see the ReBo framework. If the test came back with zero citations and you want the exact cause, ask for an AI Visibility diagnostic.
Frequently asked questions
Do I need a paid tool to run this test?
Not for the first round. The entry-level versions of the five engines and a spreadsheet are enough for twenty runs. A monitoring tool starts to pay off when the set grows to dozens of questions repeated every month across several engines.
How often should I repeat the test?
Once a month, with the same questions, in the same engines and with the same recording rules. AI answers vary between runs, so a single round is indicative. The monthly series shows whether the company's presence is changing.
Why does the AI say wrong things about my company?
Because it builds the answer from what it found published, and the easiest data to read is not always the most recent. An old figure on a forgotten page, a third-party profile nobody reviews and a fact with no stated date are possible origins. The sources column shows which one was used.
Does the test explain why my company doesn't appear?
It shows where to look. The sources column reveals whom the AI read instead of your company. The exact cause calls for a technical audit of the site, from AI crawler access to content format, which is the first stage of a ReBo diagnostic.