How to Track Brand Mentions in ChatGPT, by Hand or With a Tool

Here's how to track brand mentions in ChatGPT: ask it the questions your buyers ask, note whether it names your brand, and repeat that on a schedule. The method is short. Trusting the result is the hard part.

One answer proves little. Checking ChatGPT once is like checking the weather once and calling it the climate. It can reply differently to the same question next time, and your own account changes what it says.

There are three ways to do it: by hand, in a sheet or with a tool. Each one works until it doesn't, and the math below shows where.

Key takeaways

• Count "named" and "linked" separately. A link alone isn't a mention.
• Use counts, not percentages, until you track at least 20 prompts.
• A hand check is good for a first look, a sheet for a set you'll keep up, and a tool when you add more engines, more prompts or a history.
• Tracking shows what answers say. It doesn't show what buyers do next.

What counts as a ChatGPT mention?

A ChatGPT mention is your brand's name in the text of an answer. A link is a different thing: an answer is "linked" when its sources include your site. Together they give four outcomes: named and linked, named only, linked but not named, and absent.

Say you ask for the best CRM for a small retail team. If ChatGPT lists your brand and cites your site, you're named and linked. If it lists you but cites only a review site, you're named only.

If it cites your site without saying your name, you're linked but not named.

A name that appears only in a list of sources doesn't count as named, because the answer never said it. A link on its own is a citation, not a compliment. Keep the two counts apart from day one, or your number will drift up without you noticing.

How to track brand mentions in ChatGPT by hand

To track mentions by hand, ask ChatGPT ten to fifteen buyer questions, worded the same way each time, and write down which brands it names. Pick questions a buyer would type before choosing a vendor, not your own brand name. A brand query only tells you ChatGPT has heard of you, which is a low bar.

Write three kinds of question: research ("how do small retailers collect customer data"), compare ("Tool A or Tool B for a small retail team") and decide ("is Tool B right for a ten-store chain"). Each kind tests a different moment in the buying journey. A brand can be named in one and missing from another.

Ten to fifteen prompts is enough to learn the shape of the answers. It isn't enough to measure a share. The sampling section below explains why.

Your account changes the result more than most people expect. OpenAI says memory and custom instructions can personalize responses. A temporary chat set to unpersonalized removes those, but you're still signed in.

Treat it as a closer read, not a clean one.

Before each check:

• Use the same wording as last time.
• Run all prompts on the same day.
• Use a signed-out window or an unpersonalized temporary chat.
• Save the answer text, because screenshots are boring until someone asks why a recommendation vanished.

A hand check is good for a first look, a slide for a pitch or a sanity check on a tool's numbers. It's not good for spotting a trend.

How do you log mentions in sheets?

Log each check in a sheet. Use one row for each prompt, each engine and each run, so you can compare runs later. The sheet's main job is to stop you from remembering the good answers and forgetting the bad ones.

Set the schedule before you fill in rows. Weekly or monthly is plenty for a first set, and the same prompts should run every time. After each run, count how many answers named you, how many linked your site, and which rivals were named when you weren't.

Set the location too. When ChatGPT searches the web, it can guess where you are from your IP address.

OpenAI's help page gives an example of a query rewritten for the San Francisco area. A check from one city may not match a check from another.

Use counts until you have at least 20 prompts. A share on fewer prompts jumps around too much to read, which is why Firemark shows a percentage only from 20 prompts up. To see what a scored version looks like, open the demo report, which uses sample data.

How big is a hand-run sample?

Your sample is the number of answers you collected, and it's smaller than it feels. Ten checks on one engine is ten answers. It's like tasting one spoonful of soup and judging the whole pot.

Here's the math. Say 10 of 25 answers name your brand.

That's 40%. But the Wilson interval puts the 95% range at 23% to 59%. With 40 named in 100 answers, the range shrinks to 31% to 50%. These numbers are math on one engine's base, not data from a real brand.

A 5-point move inside a 36-point range is noise. Firemark's method page calls a movement only after three comparable runs and a change of 5 points or more.

You can see the wobble yourself. Ask one prompt three times in separate chats. If the list of brands changes, you've found variation, not movement.

By hand, 100 answers means 100 separate asks and 100 separate reads, every run.

What should a ChatGPT tracking tool do?

A ChatGPT tracking tool should ask a fixed set of prompts on a schedule, score each answer and keep the history. That way you read trends, not single answers. A sheet can't do this without a lot of upkeep.

A tool also keeps the answers, so you can say what ChatGPT said in March when someone asks why you vanished in April.

Before you buy any tool, check five things:

• Source: does it read the ChatGPT web app or the API? They're different products, and answers can differ.
• Session: is it signed out, or signed in with someone's history?
• Failures: what does it do with an answer that failed to load? A failed answer should never count as your brand being absent.
• Counts: does it show a count with a base, like "named in 40 of 100 answers"?
• Definitions: does it count named and linked separately?

Firemark reads the ChatGPT web app, signed out, from a US location in English. It doesn't use the API. Its ChatGPT tracking page explains how each prompt is asked and scored.

What do AI brand monitoring tools cover?

AI brand monitoring tools record what AI answers say about your brand across engines. They don't record what people do afterward. Some come as add-ons to an SEO suite, some are built just for tracking, and big platforms link answers to traffic and customer journeys.

If you need that link, a suite built for it fits better than Firemark, which doesn't count clicks, visits or sales. A tool you already pay for may also cover a first set, so check what it reads before adding another.

Which method fits you?

Pick a hand check for a first look, a sheet for a set you'll keep up, and a tool once the sheet costs more time than the reading.

These limits are a rule of thumb built on the 20-prompt floor, not a benchmark. Time your first run, then multiply it by the engines and weeks you plan to keep going. If the total makes you wince, you have your answer.

For scale, the Firemark pricing page, checked on October 22, 2026, lists Lite at $29 a month with ChatGPT, AI Overviews and AI Mode and 25 prompts. That's up to 75 answers per run, which is a lot of copy and paste before lunch.

What won't tracking tell you?

Tracking tells you what answers say, not what buyers do next. Clicks, visits and sales after an answer are outside the data. Pair your counts with signals you already trust, such as branded search in Search Console and referral sessions in your analytics.

Neither is complete, but together they tell a fuller story than either alone.

Counting mentions also doesn't tell you if what ChatGPT says about you is true. A brand can be named in every answer and described badly in all of them.

Start with ten buyer questions and a sheet. If the sheet gets too big, a tool can take over the same log. Firemark has a demo report, and a new account gets one free run of 10 prompts on 4 engines.