How to Track Brand Mentions in ChatGPT (Tested, 2026)
Track brand mentions in ChatGPT with fixed buyer questions, web search on and a weekly rate with its range. Why 10 answers mislead and what to log.
- chatgpt
- brand mentions
- ai visibility
- measurement
To track brand mentions in ChatGPT reliably, ask a fixed set of unbranded buyer questions through the API with web search on, set to your buyers' country, every week, and report a rate with its range (see why scores jump around). One check proves little: 3 mentions in 10 answers still leaves a 95 % range from 11 to 60 %.
Can you check by hand if ChatGPT mentions your brand?
Checking by hand shows what ChatGPT said once, to one person, not how often it names your brand to buyers. A small example shows why.
A marketing lead at a small invoicing startup types "best invoicing tool for freelancers" into her ChatGPT. Her brand is the second name in the list. She sends the screenshot to the founder, who forwards it to the board. A freelancer in another city asks the same thing an hour later and gets five names, none of them hers. (Invented example; the mechanism is documented below.)
Both answers are real. Neither is a measurement. Three things separate her screen from her buyers' screens:
- Every answer is a new draw. SparkToro had 600 volunteers run 12 prompts a combined 2,961 times and reports a "<1 in 100 chance that ChatGPT or Google's AI, if asked 100X, will give you the same list of brands in any two responses" (published 27 January 2026; the study was run with Gumshoe.ai, an AI-tracking startup, a possible conflict of interest SparkToro names itself).
- Search is optional. OpenAI's web search documentation says "the model can choose to search the web or not based on the content of the input prompt", so one answer may read fresh pages and the next may not.
- Her app knows her. A signed-in app has her history and location. Buyers ask from their own accounts, cities and languages.
So the screenshot answers "did ChatGPT name us once, for me?". The board wants a different answer: "how often does ChatGPT name us for buyers?" That is a rate, and a rate needs enough answers. Nodpick is the tool we build to measure that rate for your own brand, every week, with a range; the rest of this post explains what the setup has to get right.
How many answers do you need to track brand mentions in ChatGPT? Our test
A ChatGPT mention rate from 10 answers is too vague to act on; around 50 answers per question start to separate a weak brand from a strong one. We computed 95 % Wilson intervals, the standard range for a share, to show how wide the honest range is after a given number of answers.
| Named in | Rate | 95 % range | Width |
|---|---|---|---|
| 1 of 3 answers | 33 % | 6 to 79 % | 73 points |
| 3 of 10 | 30 % | 11 to 60 % | 50 points |
| 6 of 20 | 30 % | 15 to 52 % | 37 points |
| 15 of 50 | 30 % | 19 to 44 % | 25 points |
| 30 of 100 | 30 % | 22 to 40 % | 18 points |
| 60 of 200 | 30 % | 24 to 37 % | 13 points |
| 5 of 10 | 50 % | 24 to 76 % | 53 points |
| 1 of 10 | 10 % | 2 to 40 % | 39 points |
Can you tell a brand named in 30 % of answers from one named in 60 %? With 40 answers each, the ranges still touch (18 to 45 % against 45 to 74 %). Only at 50 answers do they separate (19 to 44 % against 46 to 72 %).
Method. Python standard library, 95 % Wilson score interval (z = 1.96), counts as shown; computed 31 October 2026. Limits. The interval assumes independent answers to the same question. It tells you how sure you can be about one question, not about your brand across all the questions buyers ask.
That last limit is the bigger problem, and it changes how you should spend your answers.
Why is one question asked 50 times still the wrong setup?
Rerunning one question only measures that question. Sistrix, analysing 10,826 daily measurement series across four AI models (published 27 October 2026), found that differences between prompts are on average 2.4 times as large as the day-to-day swing of one prompt. Which questions you ask moves the result more than how often you ask them.
Sistrix's figures make the trade-off concrete:
- One prompt: a typical deviation of ±34 points from the true value.
- 20 prompts: ±9.5 points, and never below ±7.6 however often they are repeated.
- 100 prompts: ±4.2 points.
- For ±5 points, "rund 70 Prompts"; for ±3 points, "rund 200".
Real buyers also phrase the same need in very different ways. SparkToro measured an average semantic similarity of 0.081 across human-written prompts with the same intent: almost no two people ask alike. A set of 20 to 70 varied, realistic questions, each asked once a week, beats one question asked every hour.
So what should each of those weekly answers look like?
What setup makes ChatGPT mention tracking reliable?
Reliable ChatGPT mention tracking holds every setting fixed except time: the same unbranded questions, the same access route with web search on, the same country and language, the same weekly rhythm, and the same fields recorded per answer. Change any one of them and week-over-week comparisons break.
| Setting | Do this | Why |
|---|---|---|
| Questions | Unbranded buyer questions, wording frozen | A question that names you measures recall, not discovery |
| Access | API with web search enabled, model id logged | Removes your personal history; the model id explains sudden shifts |
| Location | Set country and language of your buyers | OpenAI's API accepts "an approximate user location using country, city, region, and/or timezone" |
| Cadence | Weekly, plus an extra run when a mention flips | Breadth beats frequency; a flip is the moment noise is most likely |
| Record | Named or not, position, other brands, cited domains | Who replaces you and which sources they come from is the fix list |
| Report | A rate with its range and n | A bare percentage invites over-reading |
| Compare | Each question with itself, week over week | Removes the effect of which questions happened to change |
The manual tracking guide shows the spreadsheet version, and the glossary defines prompt tracking. It works, with patience: 20 questions on ChatGPT alone is 20 logged answers a week, and buyers do not only ask ChatGPT.
How do you track brand mentions in AI search, not just ChatGPT?
Tracking brand mentions in AI search means covering every assistant buyers use, because ChatGPT is only one of several places a buyer meets your brand before visiting your site, and the other places behave differently. Google shows AI summaries for a share of searches: Pew Research found that 18 % of the Google searches in its March 2025 browsing panel produced one, and users clicked a traditional result in 8 % of those visits against 15 % without a summary.
That gap is the case for measuring mentions at all. A buyer who reads your name in an answer and never clicks is invisible in analytics (here is how to track the clicks you do get), yet the answer shaped the shortlist. Tracking ChatGPT alone misses the same buyer when they ask Google, Perplexity, Gemini or Claude, where the cited sources and named brands can differ. For how assistants pick names in the first place, see how AI assistants choose brands.
The remaining question is whether you run all of this yourself.
What does Nodpick do that a spreadsheet cannot?
Nodpick asks ChatGPT and five other engines the same buyer questions every week and turns the answers into a rate you can trust: a score with its 95 % range and sample size, a confirmation run when your mention flips, and an alert only when a change passes a written noise rule. Every answer stays readable, with the sources it cites.
What that means in practice, from Nodpick's published method:
- ChatGPT is asked with web search on, the same way every week, and every answer is stored with the sources it cites.
- Scored questions never name your brand, so every mention counted is one the engine made on its own. One extra question asks about your brand by name for the "what AI knows about you" view and never enters the score.
- Each assistant is told your buyers' country and language.
- Every score comes with a 95 % range, and a change only counts as real when it passes a written rule.
- Every answer that names you is sorted into recommended, just mentioned or negative, and whether it cites your site is recorded separately. A mention rate alone cannot tell a pick from a passing name.
- Every plan asks the same 6 core engines: ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity and Claude.
The usual objection is "this is not what my buyers see in the app". It is true, and Nodpick says so plainly: someone signed in, with chat history, can see a different answer. No tracker can reproduce every buyer's account. What it can do is ask the same way every week, so a change in the rate is a change in ChatGPT, not in your setup.
A screenshot tells you about one answer. A weekly rate across many questions tells you whether buyers hear your name.
Check your own brand
Start with the free AI visibility checker: it asks ChatGPT, Perplexity, Gemini and Claude five buyer questions written from your website and shows whether they name you and who they name instead. It opens on launch day, and you can join the waitlist on that page to get one email when Nodpick opens. If you want the same questions asked every week with a range, see the plans.
Common follow-up questions
Can I track brand mentions in ChatGPT for free?
Yes, with a fixed question list, a spreadsheet and a weekly hour. Record whether you were named, your position, the other brands named and the domains cited, and keep the wording frozen. The limit is volume. Nodpick's free check asks 5 scored buyer questions plus 1 about your brand, on ChatGPT, Perplexity, Gemini and Claude, with no card; read it as a first look, since 5 questions give wide ranges.
Should my tracking questions include my brand name?
Not the ones that feed your mention rate. A question that names you tests whether ChatGPT knows you, not whether it suggests you to a buyer who has never heard of you. Keep one branded question separately to check what ChatGPT says about you, and keep it out of the score.
How often should I check ChatGPT for mentions?
Weekly is enough for most brands, provided the question set is broad. Sistrix advises querying "lieber 100 Prompts wöchentlich abfragen als 15 Prompts täglich" (100 prompts weekly rather than 15 daily), because the choice of prompts drives more of the uncertainty than day-to-day randomness.
Does it matter how a tracker asks ChatGPT?
Yes. Whether web search is on, which country is set and whether the wording stays frozen all change the answers. If you track by hand, log those settings and the date with every answer, so a sudden shift can be checked against a change on OpenAI's side rather than in your setup. Nodpick keeps them fixed for every weekly run, with web search on.
Sources
- SparkToro, "NEW Research: AIs are highly inconsistent when recommending brands or products" (27 January 2026), https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/, Last checked: 31 October 2026
- SparkToro, "Why everyone prompts AI differently", https://sparktoro.com/blog/why-everyone-prompts-ai-differently-even-when-they-want-the-same-thing/, Last checked: 1 November 2026
- Sistrix, "Breite schlägt Wiederholung: Wie man AI-Sichtbarkeit verlässlich misst" (27 October 2026), https://www.sistrix.de/news/breite-schlaegt-wiederholung-wie-man-ai-sichtbarkeit-verlaesslich-misst/, Last checked: 1 November 2026
- OpenAI API documentation, Web search, https://developers.openai.com/api/docs/guides/tools-web-search, Last checked: 1 November 2026
- Pew Research Center, "Google users are less likely to click on links when an AI summary appears in the results" (22 July 2025), https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/, Last checked: 1 November 2026
- Nodpick, How we measure AI visibility, https://nodpick.com/method, Last checked: 1 November 2026
- Wilson interval table (95 % Wilson score interval), computed 31 October 2026
Checked 1 November 2026.