How to track brand mentions in ChatGPT, step by step
- Write 10 to 30 questions your buyers really ask, in their words, without your brand name in them.
- Ask each one in ChatGPT with web search on, in a Temporary Chat so your own history and memory do not shape the answer.
- Record, per answer: named or not, cited or not, your position among the brands named, the competitors named, and every source URL.
- Repeat the same questions on the same day each week, in the same order and wording.
- Ask each question at least twice per run; ChatGPT gives different answers to the same question minutes apart.
- Run the same list on Perplexity, Gemini and Google's AI answers if your buyers use them.
- Read the trend after four or more weeks, with the sample size next to every rate.
The rest of this guide explains each step, the three ways to run it, and how to tell a real change from noise.
Build the question set first
The most common mistake is tracking flattering questions ("what is <your brand>?") instead of buying questions ("best <category> for <use case>"). The first kind you almost always win; the second kind is where customers actually are. Build a list of 10–30 questions your buyers genuinely ask, in their words:
- "Best X for Y" comparisons in your category — the money queries.
- Problem-phrased questions ("how do I stop AI traffic showing as direct") where your product is one honest answer.
- Alternative-seeking questions ("alternatives to <big competitor>") where challengers get discovered.
- Localized variants if you sell in multiple markets — AI assistants answer differently per language and country.
| Question group | Example question | What it shows | Counts toward the rate |
|---|---|---|---|
| Category recommendation | "best <category> in <city or country>" | Whether you are found without your name being asked for | Yes |
| Competitor alternative | "alternatives to <competitor>" | Whether you are named to someone looking for your competitor | Yes |
| Budget or constraint | "<product or service> under <budget>" | Whether your price and features are being read | Yes |
| Comparison | "<your brand> vs <competitor>" | How ChatGPT describes the two of you | No, branded |
| Brand facts | "what is <your brand> and what does it do?" | Whether what it says about you is correct | No, branded |
Keep your brand name out of the questions. A question that names you gets your name back in the answer because you asked for it, which measures nothing. How many questions a stable rate needs is set out in how many prompts to track.
To set your numbers beside named rivals on the same questions, see benchmarking your brand in AI answers.
Record more than yes/no
"Mentioned" alone hides most of the signal. For each question and AI assistant, worth recording:
- Mentioned vs cited: named in the text is good; your URL linked as a source is better — they move independently.
- Position: named first among three brands is a different outcome than named last with a caveat.
- Who won instead: the competitors and sources in the answers you lose are your roadmap.
- Sentiment and accuracy: AI assistants sometimes describe products wrongly; catching a wrong claim early matters as much as a missing mention.
- Which domains the AI assistant cited: that source list tells you where answers come from — and where you need presence.
| Date | AI assistant | Question | Run | Named | Cited | Position | Others named | Source URLs |
|---|---|---|---|---|---|---|---|---|
| 3 Nov | ChatGPT | best accounting software for freelancers | 1 | Yes | No | 2 of 4 | Brand A, Brand B, Brand C | review site, forum thread, vendor page |
| 3 Nov | ChatGPT | best accounting software for freelancers | 2 | No | No | — | Brand A, Brand C | review site, comparison article |
The two example rows are the same question asked twice on the same day, and they disagree. That is normal, and it is why the second run matters.
1. Track ChatGPT mentions by hand
By hand works for about ten questions on one or two assistants. Open a Temporary Chat, which neither reads nor writes your memory, so your own past conversations do not tilt the answer towards brands you have researched. Make sure the answer searched the web: the sources it used appear as citations beside the text, and those links are what goes in the Source URLs column.
- Paste each question exactly as written in your list; rephrasing between weeks breaks the trend.
- Ask it twice, in two separate temporary chats, and record both answers.
- Copy the cited links, not just the brand names. The cited pages are the ones you can influence: a review, a comparison, a forum thread.
- Note when ChatGPT answered without searching. An answer from the model's training data alone changes slowly and says little about this week.
The limit is arithmetic. 20 questions, asked twice, on five assistants, every week is 200 answers to read and type up. Past that point the spreadsheet stops being the cheap option.
2. Track ChatGPT mentions through the OpenAI API
If you can write a short script, OpenAI's API answers the same questions with its web search tool switched on and returns the cited URLs as structured annotations next to the text, so nothing has to be copied by hand. A scheduled job can ask each question twice a week and append the rows to a sheet or a database.
- An API answer carries no account memory or personalisation, so it is closer to a fresh user's answer than to yours, and it can still differ from what a signed-in user sees in the app.
- Pin the model name in the script and write it into every row. When the model changes, the trend has a step in it that is not about you.
- Count the cost before scheduling: every question, run and AI assistant is one paid request, plus the search tool's own charge.
- Do the same for the other assistants your buyers use; Perplexity, Gemini and Claude have their own APIs with search, and the rows join on question and date.
Seeing which sources ChatGPT cites goes through the cited-URL part of this in more detail, manually and through the API.
3. Track ChatGPT mentions with an AI visibility tool
A tool runs the loop for you: it asks your list on a schedule, stores every answer and cited URL, and computes the rates. Check two things before paying: that ChatGPT is included in the plan you would buy rather than sold as an add-on, and on what schedule ChatGPT is checked. The entry plans below include ChatGPT; the figures and read dates are the ones in our comparison of sixteen AI visibility tools.
| Tool and entry plan | Price a month | Prompts | ChatGPT cadence | Read |
|---|---|---|---|---|
| Otterly.ai Lite | $29 | 15 | Daily | 22 Oct |
| Ansvisor Starter | $49 | 50 | Daily (with Microsoft Copilot only) | 22 Oct |
| Promvia Starter (ours) | €49 | 25 | Weekly, plus checks you start yourself | 15 Oct |
| Peec AI Starter | €85 | 50 | Daily, if ChatGPT is one of your 3 models | 7 Oct |
| SE Visible Basic | $99 | 200 | Daily | 19 Oct |
| Semrush AI Visibility Toolkit | $99 | 25 | Daily | 19 Oct |
| Analyze AI Growth | $99 | 25 | Daily | 19 Oct |
Promvia's Free plan asks five questions on every default AI assistant once, then tracks them on Perplexity only, automatically once a month, so ongoing ChatGPT tracking starts on Starter. Whatever you use, hold it to the same bars as a hand-run sheet: more than one sample per question, mentioned and cited kept apart, results per assistant, and no promise of placement. No tool, ours included, can make ChatGPT name you.
How often to check, and how to read the trend
Weekly is enough for most categories. The part that changes most from week to week is the list of sources, not the brands. In our study of six weekly runs of the same 40 buyer questions, fewer than half of the sources an AI assistant cited for a question were still cited for it one run later:
So read rates, not single answers, and keep the sample size next to every rate. On 20 questions, a 20% mention rate has a 95% interval of 8.1% to 41.6%: a move from 20% to 25% is inside the noise. Promvia runs this loop across five default AI assistants and prints that interval next to every rate — that's the product — but hold any vendor, us included, to the same rules.
Google's AI answers need their own check; how to track Google AI Overviews and AI Mode covers both.
Check that ChatGPT can read your pages
A mention tracker shows the answers; your server logs show whether ChatGPT ever opened your pages to write them. OpenAI fetches pages with two agents worth telling apart: OAI-SearchBot, which crawls pages for ChatGPT's search, and ChatGPT-User, which opens a page while answering a user's question. A page ChatGPT-User never opens is a page ChatGPT is not citing from.
- Filter your access logs for the two user agents, and confirm the requests come from OpenAI's published IP ranges; the user-agent string alone can be faked. AI crawler user agents, verified lists the strings and where each operator publishes its ranges.
- Check robots.txt does not block them by accident; robots.txt for AI crawlers shows the lines.
- Compare the pages ChatGPT-User opens with the questions you track. The pages it opens are your candidates; the ones it never opens need links and a clearer answer to the question.
Close the loop with money
Mentions are a means. The loop closes when you connect them to visits and revenue: track which AI assistants send traffic (including the share hiding in "Direct" — see our guide), attribute conversions to the assistant that drove them, and compare against your mention trend. When mentions rise and attributed revenue follows, you have the paired trend that makes the case for the channel — to yourself, your team, or an acquirer.
Why AI traffic shows as direct explains the hidden share, and seeing AI traffic in Google Analytics 4 sets up the visits half.
Frequently asked questions
Does ChatGPT report my brand's mentions to me?
No: ChatGPT gives a business no report of the answers that name it. You measure it by sampling: ask a fixed list of buyer questions on a schedule and count the answers that name you, by hand, through the OpenAI API, or with an AI visibility tool.
How do I find out whether ChatGPT mentions my brand?
In a Temporary Chat with web search on, ask an unbranded question your buyer would type ("best <category> for <need>"), repeat the same question a few times, and record separately for each answer whether your name appears, whether your site is among the sources and who appears instead of you. Do not count questions that contain your brand: ChatGPT repeats your name back from the question. For the visit side, look at traffic from chatgpt.com in your analytics.
ChatGPT names my brand for me but not for my customers. What is going on?
Memory and history. A signed-in ChatGPT remembers what you have asked before, and you have asked about your own brand a lot. Check in a Temporary Chat or through the API, which answer without your history.
How often should I check?
Weekly for competitive categories, monthly minimum otherwise. AI assistants update continuously; quarterly checks mostly measure noise between two distant points.
Why do I get different answers than my colleague for the same question?
Sampling variance, personalization, location, and phrasing all shift answers. That's exactly why trend-over-repeated-samples is the honest metric and single screenshots are not evidence.
Which AI assistants are worth tracking?
Start where your buyers are: ChatGPT and Perplexity almost always, Google's AI answers for search-heavy categories, Gemini, Claude, and regional favorites as relevant. Per-assistant results differ enough that one AI assistant never represents the rest.