Brand Monitoring in ChatGPT: How to Actually Check

The first thing almost everyone does, when they finally get curious about how AI talks about their company, is open ChatGPT and type: what do you know about [brand]?

The answer comes back fluent, plausible, mostly correct. Maybe it describes a product you sunset years ago, and you make a mental note to fix that. Then you close the tab, mildly reassured, and go back to work.

That test is worthless. Not weak, not incomplete. Worthless. And it is worth understanding why, because the reason points at the test you should be running instead.

Why the vanity prompt always lies

When you name your brand in the prompt, you have already done the model's hardest job for it. You handed it the entity. All it has to do now is say something about a string it has been given, and a language model will essentially always find something to say. The answer is reassuring by construction.

It tells you the model can produce text about you when text about you is what you asked for. It tells you nothing about whether the model would ever bring you up on its own, which is the only thing that matters. Nobody who does not already know you exists is typing your name into ChatGPT. Your name in the prompt is the one condition that never holds in the situation you care about.

The prompts that measure something

Real brand monitoring in ChatGPT starts from the questions a buyer asks before they have heard of you. They fall into three rough shapes, and you want all three.

Categorical

"What is the best project management tool for a mid-sized manufacturing company." "Which suppliers should I look at for industrial coatings in Europe." This is someone building a shortlist from nothing. The model answers with a handful of names. Either you are one of them or you are not, and there is no partial credit.

Comparative

"Compare the main options for warehouse automation software." "What are the alternatives to [the market leader in your category]." These force the model to produce an explicit set, which makes them brutal and useful. They also reveal what the categorical prompts do not: the company the model treats as the default in your space, the one everyone else gets compared against.

Problem-shaped

"Our returns process is slow and customers keep complaining, how do we fix it." Here the buyer does not name a category at all. They describe a pain. The model picks the category, then picks the vendors. This is the earliest moment in the funnel, and no keyword tool ever showed it to you.

Write these questions in the words your buyers use, not the words your marketing team uses. Then ask them cold, in a fresh chat with no history. What comes back is your real position.

What ChatGPT is actually doing when it answers

Two mechanisms sit behind that reply, and confusing them is how people end up chasing the wrong fix.

From training. Part of the answer comes from what the model absorbed while it was trained: a compressed, statistical impression of an enormous slice of the web, frozen at a knowledge cutoff. No document is being read, so there is no source to cite. If your brand lives in this layer, it is because the web talked about you enough, and consistently enough, before that cutoff. A page you published last week is invisible to it.

From browsing. When ChatGPT searches the web mid-answer, it fetches pages in the moment and answers from them, usually with links. This layer moves fast, and a new page can start appearing in it within days. It is also where citations come from, which is the only route by which any of this becomes a click.

The two modes give genuinely different answers to the same question. A brand can be recommended confidently from training and never linked, or cited while browsing and absent the moment browsing is off. Test both: run your prompt set once in a plain conversation and once with search explicitly invoked, and record the results separately. Merging them hides the mechanism you would need to act on.

GPTBot, and the rule you set and forgot

The browsing half of that picture depends on OpenAI being allowed to fetch your pages. That is governed by a crawler called GPTBot, and you control it exactly where you would expect. Open yourdomain.com/robots.txt right now and look for this:

User-agent: GPTBot
Disallow: /

A lot of sites have that rule. It was added during the scraping panic, often by someone who has since left, and never revisited. It is a legitimate choice: you may not want your content used as training material, and the crawler directive is one of the few levers you have. But it has a cost, and it should be made on purpose rather than inherited.

Blocking GPTBot means OpenAI's systems do not fetch your pages. Your competitors' pages, their review sites, their comparison articles, the forum threads about them: all still readable. Yours are not. You have not made yourself invisible to ChatGPT, because other people's writing about you remains fair game. You have made yourself unable to speak for yourself.

One answer is an anecdote

Here is where most brand monitoring quietly fails. Someone runs the good prompts, gets a bad result, panics. Or gets a good result and relaxes. Both reactions are unjustified, because language models are probabilistic: the same prompt, asked twice, can produce two different shortlists. A single answer is a sample of one, drawn from a distribution you have not measured.

So repeat. The same prompt set, several times in a sitting, and again over the following weeks. You are not looking for a verdict but for a frequency: how often you show up, not whether you showed up once. For every run, record four things:

• Named? Whether your brand appeared in the answer at all.
• Linked? Whether there was an actual citation to your domain, or only a mention. Different outcomes, different causes.
• Who instead? The names that appeared where yours did not. Usually the most uncomfortable and most useful artifact of the exercise, because it is rarely the competitive set you think you are in.
• Described how? The exact framing. Being called "a budget option for small teams" is a visibility win and a positioning problem at once.

A spreadsheet handles this fine for one language and a handful of prompts. It stops working the moment the same set has to run across several languages and markets on a schedule, which is where PSentry exists: it repeats a prompt set across ChatGPT and the other major assistants, in each language you sell in, and logs those same columns so that a comparison over time is actually a comparison.

What you cannot do

Be clear about the ceiling, because the market around this topic is not. There is no submission form. You cannot register your brand with ChatGPT. There is no index to be added to, no ranking factor to adjust, no tag that makes a model like you.

What you can influence is the only input the system has: what exists on the open web about you, how clearly it says what you do and who you are for, and how consistently it says the same thing across the pages, forums, reviews and articles a model can reach. A slow lever, and an indirect one. It is also the real one, and anyone offering you a faster one is describing a product that does not work.

Monitoring will not fix your visibility. It will tell you the truth about it, in the only terms that matter: not what ChatGPT says about you when asked, but whether it says anything about you when it is not.

Frequently Asked Questions

Does checking from my own account bias the result?

It can. Memory lets the assistant carry context between conversations, so if you have discussed your company before, you are no longer running a clean test. Use a temporary or logged-out chat, and treat results from your everyday account as contaminated.

Does blocking GPTBot remove my brand from ChatGPT?

No. It stops OpenAI's crawler from fetching your pages. It does not remove what the model already absorbed in training, and it does nothing about the rest of the web writing about you. Blocking controls your own content, not your reputation.

How often should I run this?

Models do not change their mind about a brand overnight, so daily checking is theatre. A pass every few weeks is enough to see a trend, provided each pass is many prompts run several times, not one prompt run once.

ChatGPT describes my product incorrectly. Can I report it?

There is no correction channel that works the way you want. The practical route is to make the correct information easy to find and hard to misread on pages a crawler can reach, and to accept that it propagates slowly.

Should I check the other assistants, or is ChatGPT enough?

ChatGPT is the largest and a reasonable place to start. It is not representative. Systems that retrieve and cite live sources behave differently from ones answering mostly from training, and a brand can be strong in one and missing from another. If you check only one, know you are looking at a slice.

Does any of this show up in my analytics?

Only the citations do, and only when someone clicks. Mentions without links leave no trace, which means a traffic tool systematically under-reports your presence in AI answers. The measurement has to happen on the answer side, not the visit side.