Track your brand in ChatGPT

Resolve the brand name before the first question

Our fictional client is Northstar, an analytics product for small SaaS teams. Its approved owned domain is northstar.example; that reserved example address is not a live customer website.

Approve ambiguous cases with the client before calculating a percentage. The mentions-and-citations guide explains why the final row matters.

Ask questions that do not supply the answer

Begin with questions a buyer could ask before knowing the brand. These are illustrative candidates, not observed search demand:

A separate question such as “Does Northstar support cohort analysis?” checks brand facts because it supplies the brand name. Tag it as branded verification and report that cohort separately from discovery questions.

Expand the set with the buyer-prompt worksheet or the prompt generator.

Collect the answer and the context together

For a manual check, use ChatGPT’s Search mode when available, submit the approved question and save the readable answer. Open visible citations and capture their actual destinations. OpenAI documents the Search and Sources controls.

Attach the exact question/version, time, observed product surface, requested market, available location context and run outcome to the saved answer. If the location cannot be established, mark that limitation. A requested “US” question does not itself prove which location the product used.

For collection through a service, check the supported observation surface and preserve its provider/method label. A website search answer and a model API response are separate observations.

Review a single answer three ways

Consider this constructed Q04 example:

Northstar and Example C are options to compare for signup analysis. Review the retention criteria in this comparison [1].

The explicit reference [1] points to review.example/analytics-criteria. A source panel also lists northstar.example/docs as another relevant page.

The classifications are:

• Target mention: yes, because the answer text names the fictional target.
• Owned citation: no, because the explicit reference goes to a third party.
• Retrieved owned page: present in the source panel; it does not establish an owned citation.

If the citation list was only partly extracted, the owned-citation result would instead be unknown. Use the saved answer as the review artifact rather than classifying from an aggregate dashboard number.

Calculate the example without hiding missing evidence

Our complete constructed worksheet plans twelve runs. Ten yield valid answers, six of those mention Northstar, eight have complete citation evidence and three explicitly cite an owned URL.

Q11 failed and Q12 is pending. Q09 and Q10 have readable answer text but incomplete citation evidence. Those are four different states to resolve, not four answers to label “brand absent.”

The full calculation guide provides all twelve rows.

Turn the next review into a specific decision

Repeat the approved question versions under comparable conditions. If a question, brand alias or observation surface changes, label the change before comparing the results.

Choose the follow-up from the evidence: a wrong product fact needs a factual correction; an incomplete citation list needs another evidence review; an appropriate owned page absent from a complete list needs an investigation of that question and source context.

The scoped answer checker explains its current collection boundary, and AI visibility tracking connects saved answers to later reviews. Neither a single answer nor a favorable second run establishes a market-wide position or proves that a particular edit caused the change.