Reference

How to monitor brand mentions in ChatGPT

That single fact decides everything else on this page. If the data is sampled, the sampling design is the product, and two tools quoting different numbers for the same brand in the same week are probably both right about their own sample.

Can you monitor ChatGPT directly, or does everyone sample?

Everyone samples. What OpenAI publishes is a set of crawler controls and one referral signal. Its bots documentation names four agents: OAI-SearchBot, which surfaces sites in ChatGPT search; GPTBot, for model training; ChatGPT-User, for user-initiated fetches; and OAI-AdsBot. These are independent, which matters more than most guides say. You can allow OAI-SearchBot for visibility while disallowing GPTBot for training.

The only official measurement affordance is in OpenAI’s publishers and developers FAQ: ChatGPT appends utm_source=chatgpt.com to referral URLs. That tells you about clicks that arrived. It tells you nothing about the far larger number of answers where you were named, or not named, and nobody clicked.

So the honest framing of the whole category: you are buying a survey of a system, not a readout from it.

Why does ChatGPT tell me something different from what it tells my colleague?

Memory and personalization. OpenAI’s memory FAQ says ChatGPT remembers context from chats, files and connected apps to personalize your experience. Your logged-in answer is not a neutral reading.

Routing. ChatGPT is described in OpenAI’s GPT-5 announcement as a unified system with a fast model, a deeper reasoning model, and a real-time router deciding between them. Which one answers you is not something you set.

Whether it searched at all. Semrush analysed over a billion lines of US clickstream data and found ChatGPT enabled web search on 34.5% of queries in February 2026. Independently, an April 2026 variance study found ChatGPT returned zero citations on 57.8% of runs. Most ChatGPT answers about your brand come from model weights, not from your website.

Plain non-determinism. OpenAI documents that chat completions are non-deterministic by default. Thinking Machines Lab sent 1,000 identical requests at temperature 0 and got 80 different completions.

The API is not the same thing as ChatGPT

Most tools query the API, because it is scriptable and permitted. That is reasonable, but it is a different object. The API gives you a named model with no router, no memory and no logged-in state. Consumer ChatGPT applies all three. Ask any vendor which they use. Neither answer is wrong. Only one of them is what your buyer sees.

How much the answer moves week to week

The same April 2026 study measured repeated identical prompts across four engines. Day to day, the set of brands named overlapped at a Jaccard similarity of 0.45 to 0.59. Cited sources overlapped at 0.34 to 0.42. Practical reading: roughly half the brand list changes between two runs of the same question, and no dashboard in this category currently shows you which half.

Which brand mention monitoring in ChatGPT should a first-time buyer consider?

Ranked for a first-time buyer specifically, which means weighted toward low cost, low commitment and quick time to a first reading. Prices verified 2026-08-20. A tool ranked low here may be exactly right for a mature team.

If your budget is zero, start at row one. HubSpot’s grader is free, needs no account, and will tell you in two minutes whether you have a problem worth paying to measure. We would rather you spend nothing and find out than buy a subscription to discover you are already fine. If the question is genuinely only about ChatGPT, Profound’s $99 tier is ChatGPT-only by design. If you need repeatability more than breadth, Evertune is the only vendor of the nine that publishes how many times it samples.

What should a first-time buyer actually do in week one?

Write the questions your buyers ask, not the ones you want to win. “Best CRM for a 12-person agency” beats your product name. Ten to fifteen is enough to start.

Check robots.txt before anything else. If OAI-SearchBot is disallowed you will not appear in ChatGPT search answers, and no amount of monitoring changes that.

Get onto the sources ChatGPT already reads. Indig and Johnson’s analysis of roughly 35,000 ChatGPT citation URLs found 17.1% of cited domains were UGC platforms, against 11.1% review sites and 4.0% publishers. Reddit’s presence is not accidental: it signed data licensing deals with OpenAI in May 2024 and Google in February 2024.

Do not start with schema markup or llms.txt. Both have been tested. Ahrefs measured minus 4.6% on AI Overview citations after 1,885 pages added JSON-LD against matched controls, and found 97% of valid llms.txt files were never fetched.

Record your starting number and the date, and do not read week two as a trend. Given the variance above, a single-run measurement is a direction, not a delta. That is true of our score too, and we would rather write it here than defend it later.

What does a reasonable first 90 days look like?

A baseline in week one. A month of leaving it alone. Then a comparison that treats anything under a ten point move as unproven rather than as progress, unless your tool samples deeply enough to say otherwise.

If that sounds slow, it is worth knowing the scale of what is being measured. OpenAI reported 900 million weekly active users in February 2026, and G2’s survey found half of B2B software buyers now start a purchase inside a chatbot. The category is large enough to be worth measuring properly rather than quickly.