What does ChatGPT say about your business?
The only way to find out is to ask, several times, in fresh sessions, using the words a customer would use. There is no report to download and no single answer to find: the same question asked twice can produce two different replies, both plausible. What you can do is run a fixed set of prompts, record the answers word for word, note which sources were cited, and repeat it on a schedule. That turns a one-off anecdote into something you can compare against next month, which is the difference between checking and measuring.
How to check, in a clean session
Your own account is the worst place to run this test. It carries your chat history, any saved memory, your location, and a long record of you asking about your own company, all of which push the answer toward telling you what you already believe.
Set up a clean run instead. Log out, or open a temporary chat with memory and history turned off. Ask in the words a customer would use rather than your internal vocabulary. Do not correct the assistant mid conversation, because once you supply the right answer the rest of the session is contaminated. And start a new session for each prompt rather than stacking them.
If you want to know what the product does with your conversations and which controls exist, OpenAI's published privacy policy is the primary source, and it is worth reading once rather than relying on summaries.
Prompt testing: the set worth running
Five kinds of prompt cover most of what matters. Keep the wording fixed once you have written them, because changing the prompt changes the measurement.
| Prompt type | Example wording | What it reveals |
|---|---|---|
| Category | "Best [service] in [city]" | Whether you are named at all, and beside whom |
| Brand | "What is [business name]?" | The description a stranger receives |
| Comparison | "[Your business] vs [category alternative]" | How you are positioned against others |
| Trust | "Is [business name] legitimate?" | Which negative or cautionary material surfaces |
| Detail | "What does [business name] charge, and where are they?" | Whether operational facts are correct or invented |
Run each one at least three times in separate sessions. Where the assistant offers to search the web, run it both ways, because a retrieved answer and a generated one are different measurements of different things.
Answer variance, and why one reply is a sample
Variance has several ordinary sources: the generation step itself is probabilistic, the retrieval step reads whatever the live web returns at that moment, the model version can change without announcement, and personalisation shifts the result further.
The practical rule is to treat any single answer as one observation. Three consistent answers is a finding. One alarming answer is a prompt to run it again tomorrow. This is also why screenshots circulated as proof of what an assistant says about a company are weak evidence on their own: without the prompt, the date, the session state, and the number of runs, there is nothing to compare them against.
Source citations are the useful part
When an assistant searches the web, it usually attaches links. Those links are the most actionable output of the whole exercise, because they are the pages actually feeding the description of you.
Open every one. Note which are yours, which are directories or listings, which are review platforms, and which are third-party articles. Check whether the cited page actually supports the sentence it was attached to, because that connection is generated too and it is not always sound. OpenAI's documentation on web search in its models describes how those results reach the model, which is a useful frame for reading them: the assistant is summarising a small set of pages, not consulting a body of knowledge about you.
Where no citations appear at all, you are looking at a generated answer with no traceable source, which is a different problem and usually a harder one.
Wrong answers, sorted by cause
Errors are worth sorting before acting on them, because the cause decides the route and only some of them are fixable at all.
- Out of date. An old address, a closed location, a former service. Almost always a live page or listing that still says so.
- Confused with someone else. Similar names, thin or inconsistent identifying detail. Fixed by publishing clear, consistent facts everywhere.
- Sourced from something real but unflattering. A genuine bad review, a real news story. The route here is the platform's rules or the publisher, and where the material is accurate the honest answer is that it stays.
- Invented. No source exists for it. Vendor feedback, plus accurate published material for retrieval to find instead.
- Absent. You are not mentioned where you would expect to be. That is a visibility question, not an accuracy one.
Where the error rests on a page you want out of Search entirely, Google Search Central on removing information from Search sets out what the actual routes are and, just as usefully, what they do not cover.
Recording it so next month is comparable
- Date and time of the run.
- Which assistant, and the model version where the product shows it.
- Session state: logged out, memory off, web search on or off.
- The prompt, verbatim.
- The answer, verbatim. Paste the text; a screenshot is harder to search later.
- Every source cited, as a list of URLs.
- Errors flagged by cause, using the five categories above.
Keep it in one sheet, one row per run. After two or three months you can see whether anything moved, which is the only way to tell a real change from a good day. If you would rather have someone else run the first pass and sort the errors by what is realistically fixable, that is one of the things a reputation audit covers.
Questions about what does chatgpt say about your business?
What does ChatGPT say about my business?
The only way to know is to ask, several times, in fresh sessions, with the questions a customer would use. Answers vary between runs, so treat any single reply as a sample rather than a verdict.
How do I check what AI says about me?
Log out or use a temporary chat, run a fixed set of category, brand, comparison, trust, and detail prompts three times each, record the answers and their cited sources verbatim, and repeat monthly.
Why does ChatGPT give a different answer every time?
Generation is probabilistic, live retrieval returns whatever the web shows at that moment, and model versions change. Consistency across several runs is the signal worth acting on.
Should I correct the assistant when it gets something wrong?
Not during a test run, because the rest of that session will simply repeat what you told it. Record the error first, then use the feedback control afterwards.
Does asking about my own business change what it says?
Within your own account it can, because memory and chat history influence later answers. That is why a clean logged-out session is the more honest test.