What's the best tool to track how often my brand gets mentioned by ChatGPT and Claude?

Clients ask about ChatGPT and Claude together more and more, usually because someone on their team uses one and a board member uses the other. The request sounds simple: tell us how often we get mentioned. The answer takes more care than a mention counter, so here is how we handle it.

First, decide what "mentioned" means

Before choosing any tool, we agree with the client on a definition. Does a passing reference in a list count? What about a mention with a negative framing? Does the legal name, the product name, or an old name count? If you skip this step, two tools will give you two numbers and neither will be wrong.

In practice that means setting brand aliases, deciding which competitors to track, and writing down what counts. Our platform lets us keep that definition in a Brand Book that the analysis follows, so the rules do not live only in someone's head.

Mentions are a rate, not an event

A single question to ChatGPT or Claude tells you almost nothing, because answers vary from run to run. What we report is the share of answers, across a set of buyer questions, in which the brand appears, and its position when it does. That requires repeated, scheduled runs by engine, country, and language.

Our platform runs prompt panels daily per model, country, and language, and reports Visibility Score, Share of Voice, Sentiment, and Position in Answer. We compare periods to separate real change from normal variation.

The UI versus API question

This is the detail we most want clients to understand. Some engines are monitored through their real product interface. Others are tracked by calling a model through an API.

In the platform we use, ChatGPT (Search) is tracked through the real interface. Claude is tracked via API: Claude (Haiku 5.5) sits alongside other API models such as DeepSeek, Mistral, and Grok. That is a real difference. A consumer interface can add web search or other behavior an API call lacks, so ChatGPT and Claude numbers are not strictly like for like. We say so in client reports, and we compare each engine against its own history, not against the other.

If a vendor tells you Claude is tracked in its real interface, ask for the evidence. We would rather tell a client "this is an API reading of Claude" than have them discover it later.

Why we run client programs on Promptwatch

Promptwatch covers both engines on every paid plan. We do not pay a surcharge to add Claude, and we can track many other models in the same project. The free Explore tier is ChatGPT only, so it is a way to test, not to report.

Beyond counting mentions, we use it to explain them:

Citation analytics show which pages and domains an engine relied on in each answer, at page and domain level. If ChatGPT cites a review site and Claude cites a competitor's blog, that points us toward different fixes.
Heatmaps compare any two dimensions. Models by topic is the one we open first when a client asks why one engine mentions them and the other does not.
Agent Analytics shows IP-verified crawler logs, including GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User, and Claude-SearchBot. If a bot cannot reach a page, we do not need to debate content quality.
Visitor analytics connect AI-referred visits to conversions, which is the number clients eventually ask about.

Crawler logs start on the Professional plan ($245/mo). Essential ($95/mo) has none, which matters if a client wants to diagnose Claude's crawling.

What we tell clients about the gap between engines

Differences between ChatGPT and Claude are normal. They may retrieve from different sources, weigh sources differently, or answer without searching at all. We do not promise to equalize them. We do three things:

Measure each engine on its own history.
Look at which cited sources differ and whether those sources can be earned or influenced.
Check crawl access for each engine's bots, so a technical block does not hide as a content problem.

A starter prompt set

For a first month, we usually build around 30 to 50 questions in four groups: category questions ("best X for Y"), comparison questions that name competitors, problem-led questions where no brand is named, and brand questions that test what the engine says about the client. We label each by type, since brand questions behave differently from category questions.

What no tool can do

None of this can guarantee a recommendation, and no tool can tell you how often a particular person saw you in an answer. It samples answers to prompts you chose. That sample is still the best evidence available, provided you track rates over time and say clearly how each engine is collected.

If you want a single tool that tracks both engines without a surcharge, explains the gaps with citations and crawler data, and keeps the definitions consistent, that is the combination we found and the reason Promptwatch is where we run this work.