Prompt Tracking

A brand can be recommended by ChatGPT on Monday and absent from the same answer a month later, without anything on its own website changing. The answer is assembled fresh each time from whatever the model retrieves, so visibility in AI answers is a moving quantity, not a setting.

Nothing in a normal analytics stack sees this. There is no click to count when a model names a brand and the user reads on, and no referrer when it does not. Prompt tracking is the instrument that closes the gap: it re-asks the buying questions on a schedule so the movement becomes visible while there is still time to act on it.

Track the questions that decide a purchase, not the ones that mention your brand. "Best CRM for a 10-person startup" is worth tracking; "what is Salesforce" is not, because the answer names you either way and teaches you nothing about whether you are recommended.

A workable set has four groups. Category questions ("best project management tool"), which are the head terms. Qualified questions carrying a real constraint ("free", "for agencies", "that integrates with Slack"), which is where smaller brands actually win. Competitor questions ("Salesforce alternatives", "X vs Y"), where you are being compared directly. And problem questions phrased the way a buyer would ("how do I stop losing leads"), which often surface a different set of brands entirely.

Size the set to what you will genuinely re-read. Twenty prompts reviewed at every close beats two hundred nobody opens. Phrase them the way buyers actually type, follow-ups and all, rather than as sterile benchmark strings, and then leave them alone: the value is in the comparison across closes, which a rewritten prompt destroys.

These three names are used interchangeably in practice, and mostly describe the same activity: asking a fixed prompt set on a schedule and recording what comes back. Where teams do draw a line, monitoring is the alerting half (tell me when something changes) and tracking is the measurement half (show me the trend and the evidence).

The distinction that actually matters is not the label but whether the answers are kept. A tool that reports a score without the response behind it cannot be audited, and a number nobody can check is not a measurement. Insist on the recorded answer, whatever the feature is called.

Four readings carry most of the signal. Presence: does the answer name you at all, which is binary and the one that matters most. Position: how early you appear, since an answer that names three brands has no fourth place. Share of voice: the fraction of recorded answers that name you, per model. And sentiment: whether the mention recommends you or merely lists you.

Read them per model, never blended into a single number. A brand can hold first place on one model and be absent from another, and the gap between a brand's best and worst model score is itself a finding, because it points at what one model is reading that another is not.

Then read the trend, not the snapshot. A single answer is a data point; movement across several closes is the measurement. That is also the honest test of any change you make: if a fix worked, the line moves.

That asking once tells you something. It tells you what one model said on one day. Models are non-deterministic and retrieval shifts under them, so a single answer is weak evidence in either direction.

That tracking your brand name is enough. Buyers do not search your name when they are choosing between options; they describe their problem. If you only track prompts that mention you, you cannot see the answers you were left out of, which is the whole point.

That a rank is a verdict. It is a reading of one close, from one prompt set, on the models that were asked. It becomes useful when it is repeated under the same conditions, which is why the prompt set has to stay fixed.