The basic test fits in one sentence: ask ChatGPT the questions a customer would ask, several times, and note who shows up. For instance: "which company would you recommend for [your trade] in [your city]?". If your name comes back regularly, you are in the game; if it never appears, you know where you stand. The trap is concluding from a single answer: answers vary from one run to the next, and an isolated appearance (or absence) proves nothing.
The manual test, in fifteen minutes
- Write five to ten questions exactly as a customer would ask them, without naming your company: direct recommendation ("who would you recommend for..."), comparison ("what are the best..."), concrete need ("I am looking for a... in...").
- Open a session without history (private window, or a fresh account): your past conversations personalise the answers and would skew the test. What you need is the answer a stranger would get.
- Ask each question two or three times, in separate conversations, and note the names cited each time, along with the displayed sources when there are any.
- Repeat the exercise on another engine (Claude, Gemini, Perplexity): answers differ from engine to engine, and your customers are not all on the same one.
At the end, you have a first picture: the names that keep coming back, the one that may be missing (yours), and the pages engines draw their answers from.
Why a single answer proves nothing
Generative engines do not return the same answer twice: the same question, asked twice to the same engine, can cite different names. That is a structural property, not a passing defect. In our weekly measurements (four engines, repetitions per question, several markets, since June 2026), a company can appear in one answer out of two, or one out of ten: those two situations look identical if you test only once, and are worlds apart in reality.
That is why any serious measurement is expressed as a citation rate (the share of answers where your name appears), computed over repetitions, never as a yes or no. That is the principle of our measurement protocol: each question is asked several times per engine, every week, and every full answer is archived so the figure stays verifiable.
Two different questions: what the AI says about you, and whether it recommends you
Do not conflate two measurements:
- Reputation: what the AI answers when asked about your name ("what do you think of [your company]?"). That is the ground of AI reputation monitoring: tone, restated facts, possible errors.
- Recommendation: your presence when a customer asks for a recommendation without naming you. That is conquest, and it is an entirely different ground.
Our measurements show that one does not predict the other: a company can be treated very well on the direct question and remain invisible in spontaneous recommendations. So test both, separately.
Track over time, do not photograph once
AI answers move: consulted sources change, engine behaviours evolve, your competitors publish. A one-off measurement ages within weeks. What really informs you is the series: your citation rate, engine by engine, week after week, before and after your actions.
The manual test above is an excellent starting point, and it is free: do it. Its limit is time: redoing it properly every week, on several engines, with repetitions and archived answers, is exactly the repetitive work an automated tracker does for you, promising nothing beyond the measurement itself.
Read next
If the test reveals that your competitors appear and you do not, "Why ChatGPT recommends your competitors" explains the mechanisms, and "How to appear in AI answers" walks through the levers that can be worked on.