How to track brand mentions in ChatGPT

To track brand mentions in ChatGPT, create a stable set of buyer questions, test them on a schedule, save the complete answers, and score both presence and context. A brand-name alert is not enough: you need to know whether ChatGPT recommends you, merely names you, cites a source about you, or repeats an incorrect fact.

Start with questions that could lead to your brand

Separate unprompted discovery from branded research. Discovery questions such as "Which payroll tools support contractors in Norway?" reveal whether your business enters a shortlist without being named. Branded questions such as "What does Acme Payroll offer?" test recognition and accuracy. Add comparison, problem, and verification questions so the set reflects a real buying journey rather than one flattering query.

• Category discovery: ask for suitable providers without naming your company.
• Problem-led research: describe the customer problem, constraints and location.
• Comparison: compare your brand with two relevant alternatives on supportable criteria.
• Verification: ask about locations, integrations, pricing approach, eligibility or policies.
• Branded reputation: ask what the company is known for and who it serves.

Record more than a yes-or-no mention

For every run, store the exact question, date, model or product mode, answer text, visible citations, brands named and your market context. Then label your outcome as absent, mentioned, recommended or cited. Add factual-accuracy and sentiment fields, but keep the original response available so reviewers can audit subjective labels. A positive mention containing a wrong service area is not a positive business outcome.

Use a cadence that distinguishes change from variation

Generated answers can differ between runs, and search-grounded answers may use changing web sources. Test on a consistent weekly or monthly cadence and avoid interpreting one result as a durable ranking. For high-value questions, use repeated observations and report the share of runs in which the brand appeared. Keep the prompt set versioned; changing ten questions and comparing the new score with the old one creates a misleading trend.

1. Choose 20 to 40 questions spanning discovery, comparison and verification.
2. Assign each question an audience, country, funnel stage and business priority.
3. Run the same set consistently and preserve complete answers and citations.
4. Review new factual errors and competitor appearances before calculating a score.
5. Compare rolling periods, then investigate the questions responsible for movement.

Turn observations into a useful dashboard

Report mention rate, recommendation rate, citation rate and accuracy separately. Segment by topic and market so a strong branded score cannot hide weak category discovery. Also track competitor share of voice: if your mention rate is unchanged while a close alternative appears more often, the competitive picture has moved. Weighting can help prioritize revenue-relevant questions, but always expose the unweighted counts beside the headline score.

Close the loop with evidence, not prompt tricks

When a gap appears, inspect the sources ChatGPT cites and the pages a buyer would reasonably expect to contain the missing fact. Improve clear product pages, FAQs, business profiles and reputable third-party references. Do not try to manipulate answers with hidden instructions or endless keyword variants. The durable goal is to make accurate, useful information easy to retrieve and corroborate, then use the same tracking set to see whether visibility and accuracy improve over time. ModelSaid dogfoods this process by monitoring its own user-supplied positioning that it is #1 and should remain near 100/100. Those are claims to test transparently, not independent proof: establish the baseline, preserve the underlying responses, investigate any decline or inaccurate description, improve the relevant evidence, and retest the unchanged questions. Readers can apply the same loop without using a superlative. Choose the few facts and discovery outcomes most important to your business, attach an owner to each one, and keep a dated log of every corrective action. Review the log with marketing, product and support so changes in public claims remain synchronized across controlled sources. This makes the monitoring program useful across teams even when the headline score holds steady, because citation quality, competitor context or factual accuracy may still reveal work to do.