How to Check What ChatGPT Says About Your Brand

Most brand audits of ChatGPT are worthless, and they fail in the same three places every time. Somebody types the company name into their own logged-in account, gets a pleasant paragraph back, screenshots it, and drops it into a deck captioned “what ChatGPT says about our brand”.

Every part of that is wrong. The account is personalised to them. The prompt is not one a buyer would use. And one run of a non-deterministic system is not a finding.

If you want a real answer to what does ChatGPT say about my brand, here is the version that actually tells you something. It takes about an hour the first time, and it will not flatter you.

First, stop testing on your own account

ChatGPT personalises. Memory retains things across conversations, custom instructions steer tone and framing, and the model has your entire chat history with you as context for who you are and what you want to hear.

If you have ever discussed your company in that account, and you have, the model knows whose side it is on.

The most flattering review your brand will ever get from ChatGPT is the one you run on your own logged-in account. It is also the least useful.

So before anything else: use a temporary chat, the mode that is excluded from memory and history, or sign out entirely and use the logged-out product. Both give you something much closer to what a stranger sees.

Two more things worth setting up before you start:

Check which model you are on. Different models inside ChatGPT answer differently, and a reasoning model will often produce a more careful, better-sourced reply than a fast one. Record which you used. Comparing this month’s reading on one model against last month’s on another tells you nothing about your reputation.

Notice whether it browsed. Some answers are written from training memory and some are assembled from live pages it just read. When it browses, you get citations, and those citations are the most actionable thing in this whole exercise. When it does not, you are looking at a slower-moving picture that new coverage will not shift for a long time.

Ask what a buyer would ask

Your brand name on its own is the least informative prompt available. Nobody researching a purchase types a single company name and stops.

Use these five, replacing the bracketed parts. They are ordered deliberately, from the question where you are most likely to be missing entirely to the one where you are most exposed.

1. Category discovery. “I am a [role] at a [size and type of company] in [market]. Who are the best providers of [category] for us?”

This is the one that matters most, and the one brands skip because it does not mention them. If you are not in the answer, nothing else on this list is your biggest problem.

2. Comparison. “How does [your brand] compare with [named competitor] for [specific use case]?”

Comparison prompts are where inherited framing shows up. Watch which of you gets the qualifier and which gets the plain statement.

3. Direct brand. “What is [your brand], who owns it, and what do they sell?”

The dull one, and the one that catches stale facts: old leadership, a retired product, an ownership structure that changed two years ago.

4. Objection. “Is [your brand] reliable? Have they had any controversies or complaints?”

Ask it. Your buyer will. Better you read the answer first.

5. Procurement. “What should I ask [your brand] before I sign a contract with them?”

This is a free read on how the market frames your weak points, and it is frequently the most useful paragraph of the hour.

Adapt the wording to your category, but keep the shape. The point is that four of the five never assume you are the answer.

Run every prompt more than once

This is the step that separates a measurement from a screenshot, and almost nobody does it.

ChatGPT is not deterministic. Ask the identical question three times in three fresh temporary chats and you will get three different replies. Sometimes the difference is only phrasing. Sometimes you are recommended in one and absent in the next.

Run each prompt at least three times, five if you intend to report it. Then record the spread, not just the best run.

The spread is genuinely informative:

• Named in five out of five is a strong position.
• Named in two out of five is a coin toss, and it will read very differently to two different buyers on the same afternoon.
• A claim that appears in every run is load-bearing, probably sitting in the model’s memory, and slow to shift.
• A claim that appears in one run out of five is probably retrieved from a single page, and that page is findable.

That last distinction is worth the extra twenty minutes on its own, because it tells you whether you are facing a source problem you can fix this quarter or a memory problem measured in years.

Write down four things, every time

Not the screenshot. Four fields, in a sheet:

1. The prompt, word for word, so next month is genuinely comparable.
2. The engine and model, plus the date.
3. The answer in full. Paste the text, not an image, so you can search it later.
4. Every source it named. This is the column that pays for the whole exercise.

The source list is your brief. When an answer cites a review site, a two-year-old trade article and a forum thread, those three pages are currently doing your brand communications, and you now know that. Read them. That list will change what you work on more than the sentiment of the answer ever will.

Add a fifth column if you are being rigorous: claims that are wrong, itemised. Not “the tone was negative” but “says our head of engineering is X, who left in 2024”. Specific, checkable errors are the ones you can actually do something about.

Reading what comes back

Four questions, in this order.

Were you there at all? Presence in the category prompt is the headline number. It is also the one most monitoring tools cannot see, because a tool built to find mentions has nothing to report when you are not mentioned.

Is it accurate? Go through claim by claim. Accuracy problems are the most fixable of everything on this page and the most reliably ignored, because they are boring.

How are you characterised? Look at the adjectives and the qualifiers. “A solid option for smaller teams” and “an established choice for enterprises” are both positive and they are not the same answer.

Who wins when it has to choose? If the reply names a single recommendation, whose name is it? Track that across runs. It is the closest thing to a scoreboard this medium has.

When the answer is wrong

You cannot edit ChatGPT. Nobody can, and any vendor who says otherwise is selling something that does not exist.

What works is source work, in this order:

Ask the chat where the claim came from. Follow up with “which sources support that?” If it browsed, you get the pages. If it did not, search the claim in the model’s own phrasing and the origin usually surfaces quickly, because the wording tends to echo whatever it read.

Publish the correct version somewhere models retrieve from, starting with your own site. One question per page, the answer in the opening lines, the claim stated plainly and dated.

Get it corroborated off your own domain. A correction that exists only on your site is a claim. The same fact in trade press or a reference source is a fact, and it is weighted like one. This is ordinary media relations pointed at a new audience.

Re-test on a schedule and write down the result. Retrieval-based answers can move within weeks. Answers coming from training memory will not move for far longer, and knowing which one you are dealing with, from the run-to-run spread above, tells you which timeline to promise.

The full version of that sequence, including what to do about being absent rather than misdescribed, is in the pillar on AI reputation management.

Turning it into something repeatable

Doing this by hand once is genuinely valuable, because you will read the answers properly rather than skim a dashboard, and you will come away with a much sharper sense of how your category is being described.

Doing it by hand every month is where it dies.

Count the work: five prompts, five runs each, is twenty-five readings for ChatGPT alone. A serious prompt set is fifteen or twenty prompts, not five. Then repeat the whole thing for Gemini, Claude, Perplexity and Google’s AI Overviews, because your buyers are not all using the same assistant. That is several hundred readings a month, each one needing the answer stored, the sources extracted and the claims checked against what is true today.

That arithmetic is the entire reason AI reputation tracking is built as a daily automated check with the prompt set held constant and the source chain captured every time. Not because the manual version does not work. Because it works exactly once, and then quietly stops happening in month three.

Either way, run it by hand first. An hour spent finding out what ChatGPT says about your brand to somebody who has never heard of you is rarely an hour wasted, and it tends to end the internal argument about whether any of this matters.

Common questions

How do I find out what ChatGPT says about my company?

Start a temporary chat, or sign out, so your account history and memory cannot shape the reply. Ask the questions a buyer would ask rather than your brand name on its own: who leads the category, how you compare with a named rival, and whether you are reliable. Run each question at least three times, because the answers vary between runs, and write down every source it names.

Why does ChatGPT give me a different answer than my colleague gets?

Three reasons, and all of them are ordinary. The model is not deterministic, so repeated runs of one prompt differ. Memory and custom instructions personalise replies to each account. And ChatGPT browses for some questions and not others, so one of you may be reading live pages while the other reads training memory.

Should I test ChatGPT logged in or logged out?

Logged out, or in a temporary chat, for any reading you intend to record. A logged-in session carries your memory, your custom instructions and your past conversations, all of which push the answer towards what you already believe. That is the single most common way a brand audit flatters itself.

How many times should I run the same prompt?

At least three, and five is better for anything you plan to report. A single run is an anecdote. Running the same prompt several times tells you whether a good answer is reliable or a lucky draw, and the spread between runs is usually more informative than any one reply.

ChatGPT is saying something false about my company. What now?

You cannot edit the output, so work on the sources. Ask the chat to list where the claim comes from, read those pages, then publish a clear dated correction on your own site and get it corroborated somewhere independent. Re-test the same prompt weekly afterwards, because changes surface over weeks rather than days.