How do I track my brand's visibility in ChatGPT? The setup we use with clients

The first week of almost every AI visibility engagement looks the same. The client has already tried asking ChatGPT about themselves, got one answer, and either panicked or celebrated. Neither reaction is justified by a single answer. Here is how we replace that with tracking that holds up.

One answer is an anecdote

ChatGPT responses vary by run, by country and by who is asking. If you check once and see your competitor named first, you have learned about one sample. The fix is repetition: the same prompts, run daily, recorded over weeks. That is the difference between an anecdote and a trend, and it is the reason manual checking does not scale beyond building the first prompt list.

Build the prompt set

We start with a working session. We collect the questions the sales team hears, the queries from Search Console and the "people also ask" style questions around the category. Then we sort them into three groups: category questions, brand questions and competitor comparisons. Each prompt gets a topic tag and, where it matters, a country. Thirty to fifty prompts is a good start. More is not better if nobody reads the report.

Personas are worth a mention. A buyer who types "best tool for a small agency" gets a different answer from one who types "enterprise platform with SSO." Promptwatch supports persona stacking, which lets us run the same prompt through different personas without duplicating the list.

What we measure

We run client programs on Promptwatch. For ChatGPT it reports, per prompt and per day:

Visibility Score, Share of Voice and Position in Answer.
Sentiment, scored per response, with a weekly AI summary.
The pages and domains cited, through citation analytics.

The Brand Book is the part people skip and later regret. Its instructions decide what counts as a mention, which stops a brand name that is also an ordinary word from skewing the data.

The part analytics cannot see

Standard web analytics will show visits that arrive from chatgpt.com, which is useful. It will not show the crawlers. If ChatGPT cannot fetch a page, it cannot cite it, and the crawler never runs your tracking script. We check the Agent Analytics crawler logs in Promptwatch for OAI-SearchBot and ChatGPT-User activity, filtered by IP so spoofed user agents drop out. A page that is fetched regularly but cited rarely is a content problem. A page that is never fetched is an access problem. We need to know which one before we spend client budget.

Linking visibility to visits

Mentions are nice. Pipeline is better. Visitor analytics in Promptwatch tags AI-referred sessions and conversions, installed with a lightweight script, a GTM template or the WordPress plugin. That lets us tell a client not just "you were named more often" but "here are the sessions and conversions from ChatGPT." Even when the numbers are small, they show direction.

How we run the review

Every week, the same short routine:

Which prompts gained or lost visibility?
Which competitor moved?
Which pages got cited, and which stopped?
Any crawler errors on important pages?
What happened to AI-referred visits?

The Action Board collects the follow-ups as prioritized tasks. On Business plans the Content Agent can draft pages for the gaps, and we keep a review step before anything goes out.

What we tell clients up front

Tracking does not guarantee a recommendation. It tells you where you stand and where to work. Expect meaningful trends after several weeks, not days, and judge progress on a stable prompt set. If a provider promises a ChatGPT ranking, ask how they would prove it. We prefer to show the data and let the client decide.

Other tools can handle part of this. AthenaHQ, Scrunch AI and Profound each cover tracking with their own emphasis, and Otterly.AI and Peec AI are quick for a first look. We chose the platform that also handles crawler logs and conversion attribution, since clients ask about those within a month.