To check what ChatGPT, Gemini or Perplexity says about your company, ask each assistant the questions a sceptical buyer would ask, not just your name; run them in every mode the assistant offers; record the full answer and every citation; and repeat the run on a schedule. One question in one assistant on one day proves almost nothing, because answers vary between runs, between browsing and memory modes, and between surfaces. The procedure below takes about an hour the first time and is the same one we run at the start of an AI reputation audit; the audit adds coverage, repetition over days and the source tracing that turns a bad answer into a repair plan.
Why asking once is not a check
An AI answer about your company is assembled on the spot from two layers: a frozen snapshot of the public web learned during training, and whatever the assistant retrieves live when it searches. Switch browsing on and the answer can change completely. Ask the same question an hour later and the wording shifts again. Ask “is it legit” instead of “reviews” and a different set of pages gets pulled in. A single screenshot, good or bad, is a coin flip. What you want is the pattern across prompts, modes and days, because the pattern is what your buyers, investors and recruiters statistically meet.
How to check what AI says about your company, step by step
Step 1: Build the prompt set
Write the questions people actually ask about you, in the words they use. The same set we use in ChatGPT reputation management work:
- your company name alone, and your name with your city or industry;
- “is it legit”, “is it safe”, “is it a scam”;
- “reviews”, “complaints”, “lawsuit”, “controversy”;
- comparisons with your two or three closest competitors;
- the awkward questions your sales and hiring teams already hear.
Keep the set to ten or fifteen prompts and use exactly the same wording every time you run it, so later checks compare like with like.
Step 2: Run the set in ChatGPT, with and without browsing
Run every prompt twice: once in the plain chat mode, which draws on model memory, and once with search or browsing switched on, which pulls in live pages. Record both answers in full. In browsing mode, open and note every citation ChatGPT shows, because those links are the raw material for any fix. Where the two modes disagree, the memory answer tells you what the model believes and the browsing answer tells you what the web is currently saying.
Step 3: Run it in Google AI Overviews and AI Mode
Type each prompt into Google search in a private window with your market’s location set. For many brand queries an AI Overview appears above the results; the AI Mode tab gives a longer answer to the same question. Record the summary text and every linked source. AI Overviews are built from what already ranks for the query, so they double as a map of your page-one results; our post on why AI Overviews show negative reviews explains the mechanics.
Step 4: Run it in Gemini
Ask each prompt in Gemini and then ask it which sources it used, or use its response-check feature where available, so you capture both the answer and what grounded it. Gemini leans on Google Search for current facts and on model memory for the rest, which is why its answer can differ from the AI Overview for the same query. Our Gemini reputation management page covers what drives its answers.
Step 5: Run it in Perplexity
Perplexity footnotes every page it reads, which makes it the fastest way to see which URLs are feeding the negative sentences across all the assistants. Run the set, record the answer, and copy out every numbered citation with the sentence it supports. This citation trail is how we build the source map in an audit; our Perplexity reputation management page explains how it is used.
Step 6: Run it in Microsoft Copilot
Copilot retrieves from Bing’s index, and in our experience ChatGPT’s browsing leans on Bing too, so a page that ranks on Bing but not on Google can still drive two assistants. Run the set, record the answer and the links it shows.
Step 7: Record everything in one sheet
Keep one row per prompt per assistant per mode: the date, the verdict (clean, drifting or damaging), the damaging sentence copied verbatim, and the cited URLs. Take a dated screenshot of each answer. This sheet is your baseline. Without it you cannot tell next month whether anything changed.
Step 8: Repeat monthly, with the same prompts and modes
Re-run the whole set once a month, or sooner after any press, dispute or review spike, and compare each cell against the baseline. A new sentence or a new citation is your early warning that a fresh source has appeared, and it is far cheaper to deal with a source before the assistants have settled on it. This is what AI brand monitoring does as a standing service, on a schedule and across many more prompt variants than most teams run by hand.