Blog
What ChatGPT says about your company: how to check in ten minutes
A practical method to find out what ChatGPT, Gemini and Perplexity say about your company, how to read the answers, what to note down and when the signal deserves a real analysis.
The question "what does ChatGPT say about my company" has an answer you can get in ten minutes. It is not an analysis, but it is enough to understand whether you have a problem of absence, old facts or tone. This article explains how to do it well, what to note down and how to tell a signal from noise.
Before you start: turn on web search
Generative engines answer in two ways. Without browsing they use training memory, frozen at a date and built on what was repeated across many sources. With browsing they search the web in real time and cite sources. For a useful test you need both modes: the first tells you what the model "knows" about you, the second what it finds today. On ChatGPT turn on web search; Perplexity always browses; Gemini shows AI Overviews with sources.
Ask each question in a new chat, with no previous context: the model must not know who you are or what you expect.
The right questions
Questions should be divided into four types, because they measure different things.
Branded: they contain your name. "What do you know about [Company]?", "What is it like to work at [Company]?", "Is [Company] reliable?". They measure accuracy and tone.
Category: they do not contain your name, but should make you appear. "Which are the best [industry] companies in [country]?", "Which [service] provider would you recommend to a [size] company?". They measure presence where shortlists are decided.
Comparison: "[Company] or [Competitor]: which one should I choose?". They measure the arguments the model uses for and against.
Risk: "Are there controversies linked to [Company]?", "What are the weaknesses of [Company]?". They bring out dated narratives.
The full list, with ten ready-made questions, is on the DIY test page.
What to note down
For every answer note four things in a table:
| What | Why |
|---|---|
| Are you present? In what position? | In category questions there are three or four slots |
| Which facts are cited, and from which year? | The median year of the facts is your freshness indicator |
| The tone: positive, neutral, mixed, negative | It determines the decision of whoever reads |
| Which sources are cited, and do you look after them? | Sources are where you act |
If the engine does not show sources, ask for them: "Which sources did you use to answer?". The answer is not always complete, but it often reveals the Wikipedia page, the article or the review that weighs most.
Repeat, and change engine
Generative answers vary: the same question asked three times produces three different texts. What comes back in all three is the consolidated perception; what appears only once is noise. Repeat each question at least twice.
Then change engine. ChatGPT and Copilot read the Bing index, Gemini the Google one, Perplexity its own, Claude relies more on training memory. If one engine knows you and another does not, the problem is not your website: it is the coverage of the sources that engine uses.
How to read the result
Three outcomes are common.
Absence from category questions. The most frequent and the most underestimated. Branded answers are correct, but when someone asks "which supplier" or "where to apply" your name does not appear. The cause is almost always absence from the third-party sources models use for rankings: industry directories, communities, comparison articles.
Old facts. The answer cites a restructuring, a lawsuit, a product from years ago, and mentions nothing recent. The cause is an authoritative source frozen in time (often Wikipedia or a news article) that nothing more recent has contradicted.
Mixed tone for no obvious reason. Phrases such as "some reviews mention" appear. The cause is dated reviews left unanswered, which the model summarises as the current state.
When a real analysis is needed
The ten-minute test shows one answer per question, at one moment, on one engine at a time. A GEO Assessment repeats dozens of questions several times on five engines, traces every source with URL and weight, compares with competitors and measures again after 90 days. If the test showed you one of the three outcomes above, the next step is the free audit: in 48 hours we present the full picture.