How to See If ChatGPT Mentions Your Brand

A client asks the obvious question: does ChatGPT mention us? Nobody on the team actually knows. Someone tried it once on their phone, got a nice answer with the brand name in it, took a screenshot, and that became the official position for the next six months. Then a prospect mentions that ChatGPT recommended a competitor instead, and the screenshot stops meaning anything.

This is a measurement problem, not a visibility problem. Getting mentioned in AI search and getting cited by ChatGPT and Google AI Overviews cover the tactics. This post is about something narrower and, honestly, more neglected: how you actually check, on a schedule, with a method that produces the same kind of answer twice.

Why one lucky screenshot doesn't tell you anything

ChatGPT answers are not deterministic. Ask the same question twice, even back to back in a fresh chat, and you can get two different sets of recommended brands, two different phrasings, sometimes a mention in one run and a complete omission in the next. That's not a bug you're going to fix by asking more carefully. It's how a model that's sampling from a probability distribution works, and it means a single test run is closer to a coin flip than a verdict.

Your account state changes the answer too. Logged in versus logged out matters, because a logged-in session can pull from saved memory and past conversations. Which model is answering matters (free-tier users and paid users may not be on the same default model). Whether web search fired for that specific prompt matters a lot, because a model answering from training data alone can genuinely not know your product exists, while the same prompt with live search turned on might pull a current page that names you directly.

None of that means testing is pointless. It means you need a method that controls for the variables you can control, and that treats one run as a data point, not an answer.

Set up a clean testing environment first

Before you touch the prompt list, lock down the account state so your results are comparable from one week to the next.

Log out, or at minimum use a temporary chat, which OpenAI documents as a session that doesn't use or create memories and doesn't appear in your history. A temporary chat is the closest thing to a neutral, first-impression test: no saved facts about you, no prior conversation nudging the model toward or away from a brand it's seen you mention before.

Turn memory off for your baseline runs. If ChatGPT remembers that you work at a company, or that you've asked about a specific competitor before, that context leaks into later answers in ways that make your test non-representative of what a stranger would see. OpenAI's memory settings let you view and clear saved memories, or turn the feature off entirely under Settings, Personalization; do that before a monitoring session, then decide separately whether you also want to test the personalized experience, which is a different and valid thing to track, just not the same thing as a cold, generic check.

Test both a logged-out state and a logged-in-with-memory-off state if you can. Logged out shows you what anyone encountering ChatGPT without an account sees. Logged in but memory-cleared shows you the experience most of your actual prospects are having, since most people researching a purchase are logged into some account.

Decide what you're doing about web search deliberately, not by accident. By default, ChatGPT decides on its own whether to run a live search for a given prompt, based on whether it judges the question needs current information, and OpenAI's own documentation on web search describes this as an automatic behavior, not something you have to enable per message. You can override it by clicking the search control in the composer, or by phrasing the prompt in a way that makes a live lookup obviously necessary ("what are the current options for..." tends to trigger it more reliably than a vaguer phrasing). Run each prompt once letting ChatGPT decide and once forcing search on, and note which happened. A model answering purely from training data is really answering "what did ChatGPT learn about this during training," which is a different and often stale question from "what does the live web say about this right now."

Build a prompt set that matches how people actually ask

A single prompt ("tell me about [brand]") tells you almost nothing useful, because nobody buying anything phrases their research that narrowly. Build a list that covers the actual shapes of question your buyers use, then run the whole list on a schedule rather than cherry-picking one favorite prompt forever.

Cover at least these shapes: a direct brand question ("what is [brand] and what does it do"), a category recommendation with no brand named ("best [category] for [specific use case]"), a comparison against a specific competitor ("[brand] vs [competitor]"), an alternative-seeking prompt ("alternatives to [competitor]"), and a problem-first prompt that never mentions a product category at all ("how do I solve [the problem your product solves]"). That last one is the hardest and most honest test, because it's the one closest to how someone searches before they know what kind of tool exists to fix their problem.

Vary the specificity, too. "Best CRM" and "best CRM for a 12-person real estate brokerage that needs text message follow-up" will often surface different brands entirely, because the second one gives the model something concrete to match against rather than a generic popularity contest.

Run every prompt more than once per session if you want any sense of consistency. Three runs of the same prompt, noting whether the mention shows up zero, one, two, or three times, tells you far more than one lucky hit.

A sheet you can actually keep up

Logging results in a spreadsheet is the unglamorous part that makes the whole exercise worth anything. Without it you're back to vibes and screenshots. Here's a structure that holds up over months:

Track your competitors in the same sheet, using the same prompts. A flat line for your own mention rate reads very differently depending on whether competitors are climbing, flat, or also dropping out. If three competitors are losing mentions in the same window you are, that's a model update or a shift in what sources it's drawing from, not something specific to your content. If you're the only one losing ground, that's worth investigating on your end.

Watch for near-misses that aren't really mentions. ChatGPT sometimes names a parent company instead of the product brand, or gets the name slightly wrong (an old product name you've since rebranded away from, a common misspelling), or confuses you with a similarly-named company in an unrelated category. Decide ahead of time whether a near-miss counts in your tracker as a partial mention worth a note, or as a non-mention, and apply that rule consistently so your trend line means the same thing from week to week.

Run this weekly or every two weeks, not daily. Day-to-day noise from model sampling will make daily data look volatile for no real reason; a two-week cadence shows you an actual trend instead of noise. Keep model version as a column if you can tell which one answered, since OpenAI updates default models periodically and a mention pattern that holds for months can shift after a model update for reasons that have nothing to do with anything you changed.

Referral traffic from chatgpt.com in GA4

Everything above is manual and front-end. There's also a back-end signal worth checking: whether people who saw your brand in ChatGPT are actually clicking through to your site.

ChatGPT has been appending utm_source=chatgpt.com to outbound citation links for a while now, and since May 2026 Google Analytics made this easier to isolate on its own: GA4 added a native "AI Assistant" channel to the Default Channel Group, which automatically reclassifies sessions from recognized AI referrers, including ChatGPT, out of generic Referral traffic and into their own bucket, with no configuration required on your end (Google Analytics Help). Before that change, you had to manually dig through Acquisition, Traffic acquisition, and filter the Session source dimension for chatgpt.com to find this traffic at all, and that manual path still works if you want to see the raw referrer rather than the rolled-up channel.

Verify this is actually capturing what you think it's capturing before you trust it. Check Reports, Acquisition, Traffic acquisition, switch the primary dimension to Session source / medium, and confirm you see rows reading chatgpt.com / referral (older data) or sessions landing in the new AI Assistant channel (current data). If you see nothing, that can mean genuinely zero referral clicks, or it can mean your GA4 property predates the channel update and needs the definition refreshed, so don't read an empty report as proof of zero visibility without double-checking the setup.

Treat this number as a floor, not a ceiling. ChatGPT's mobile app and some paid-tier flows don't reliably pass a referrer, and when the UTM parameter is also missing, that session lands in Direct traffic in GA4 with no way to attribute it back to ChatGPT at all. The honest read on this data: it confirms mentions are driving some clicks, and it shows you which specific pages are getting picked, but a low or flat number doesn't mean mentions themselves are flat, because a real share of that traffic is uncountable by design.

If you're also tracking Google AI Overviews mentions, keep the two data sets separate rather than blending them into one "AI traffic" number. They come from different retrieval systems, update on different schedules, and a page that gets picked up constantly by one can be invisible to the other for reasons that have nothing to do with overall content quality.

The difference between tracking a mention and tracking a citation

Worth being precise about this before you report numbers to anyone upstream. A citation is a link ChatGPT attaches to a specific claim, pointing back to a page it actually retrieved. A mention is just your brand name appearing in the answer text, which can happen with a citation attached or with none at all, especially for established brands the model already has some baseline knowledge of from training. Your tracking sheet should have a column for each, because conflating them will make a "we got cited" report mean something different depending on which kind actually happened. The brand mentions piece goes deeper on why that distinction changes what you should actually do about it.

Tools that track ChatGPT mentions, and what they don't do

Manual testing is free and it's the only way to see the exact wording of what gets said about you, which none of the scaled tools show you verbatim at the volume you'd want. It doesn't scale past a handful of prompts checked a few times a month without eating real staff time, which is the gap the newer monitoring platforms are built to close.

We've covered the four tools worth knowing in this category in detail, with current pricing, in our guide to getting your brand mentioned in AI search: Semrush's AI Visibility Toolkit, Profound, Otterly.AI, and Peec AI. All four track ChatGPT specifically, not just Google AI Overviews, and all four run recurring prompt sets automatically so you're not relying on a person remembering to check. The thing to verify before buying any of them: ask for (or trial) the exact prompt list and cadence the tool uses for your category, because a platform that checks five generic prompts a week is not equivalent to one checking fifty, and the marketing pages don't always make that volume obvious up front. None of them currently expose the literal, word-for-word chat transcript the way your own manual spot-check does, which is why we'd keep doing both: the scaled tool for trend lines and competitive benchmarking, the manual sheet for knowing exactly what sentence you're actually being named in.

Once you know where you stand, the fix depends on the gap. If ChatGPT names competitors when people ask for recommendations, start with our SEO tactics to get recommended by ChatGPT. If you sell products and they never appear in shopping answers, the ChatGPT merchant setup guide is the place to start.

Tracking tells you where you stand; it doesn't move the numbers. When your prompt panel shows gaps, our LLM seeding framework lays out which off-site work to prioritize first.

Manual checks work for a small prompt set. Past a few dozen prompts, a dedicated tool saves hours; our roundup of ChatGPT rank tracker tools compares coverage and pricing.

Being mentioned isn't enough if the answer gets your pricing or features wrong. Our guide to AI brand sentiment covers how to audit accuracy and fix the sources behind it.

Not sure which prompts to track? Start with our method for estimating AI prompt volume. And if ChatGPT seems unaware of a recent launch, check the model's knowledge cutoff date.

FAQ

Does ChatGPT give the same answer every time for the same prompt?

No, and that's the main reason single-run testing is unreliable. Model responses are sampled, not fixed, so the same prompt in a fresh chat can return different brand names, a different order, or no mention at all on a different run. Testing the same prompt three or more times per session and logging each result separately is the only way to see the real pattern instead of one sample of it.

Should I test logged in or logged out?

Both, for different reasons. Logged out (or a temporary chat with memory off) shows a neutral baseline closest to what a new prospect with no account history would see. Logged in with memory intact shows the personalized experience, which matters if your buyers are existing ChatGPT users, but it's not a fair comparison point across time unless you keep the same account state consistent between test runs.

Can I see exactly which page ChatGPT cited when it mentioned my brand?

Sometimes. When ChatGPT's search feature fires and the answer includes an inline citation, you can usually see the linked source in the UI. When the mention comes with no citation attached, which happens often for well-known brands the model has prior knowledge of, there's no way to know from the chat alone which page or source it's drawing on.

Is GA4 referral traffic from chatgpt.com a reliable way to measure how often I'm mentioned?

It's a reliable way to measure clicks that resulted from a mention, not mentions themselves. A lot of ChatGPT-originated traffic lands in Direct because the referrer or UTM parameter doesn't always survive the trip, especially from the mobile app, so your GA4 number will understate real visibility. Use it alongside manual or tool-based prompt tracking, not instead of it.

How often should I actually run this check?

Weekly or every two weeks is enough to see a real trend without drowning in noise from normal answer variability. Daily checks mostly show you that the model is non-deterministic, which you already know by the second day.

Start with ten prompts, three runs each, logged out with memory off, and put the results in a sheet before you do anything else. You'll know within one session whether you're actually being named in the conversations that matter, or whether the one good screenshot from three months ago was the exception, not the rule.