AI Search

How to monitor your brand in ChatGPT: a question set

Branddi · Published on August 27, 2026 · Updated on October 17, 2026

To monitor your brand in ChatGPT, build a fixed question set covering purchase, reputation and comparison decisions. Run it periodically in new conversations. Record the platform, language, country, recommended brands, sources and evidence. The goal is to detect changes and decide what to improve in the company’s public information.

How can you find out whether AI products mention your company?

Ask questions a buyer would use before knowing your company. Record whether the answer mentions the brand, recommends its solution or includes a source from its domain. These are different observations. Asking what your company does helps assess its description, but does not show whether it would be discovered spontaneously.

Keep questions without brand names separate from those that include the company. A combined rate can look high simply because the questionnaire supplied the name. Check for namesakes too: confirm that the answer concerns the right company and country.

A single query captures one moment. A fixed set builds a history. Google explains that answers and links can vary between AI Mode and AI Overviews. The protocol below is a proposed way to compare observations, not an official ranking or a sample of every user's experience.

Which questions belong in the first version?

As an operational starting point, we suggest 15 to 25 questions across distinct intents. This is not a statistical rule: adapt the range to your capacity to collect and review answers. Draw them from commercial terms, sales conversations, recurring objections, and risks the brand needs to watch.

• Group — Example question — What to observe
• Category — “Which companies provide digital brand protection in Brazil?” — Presence, order, and recommendation rationale
• Problem — “How can I identify who advertises on my brand name?” — Whether the brand is linked to the right solution
• Purchase — “Which tool monitors unauthorized marketplace sellers?” — Recommendations and criteria used
• Comparison — “Branddi or [competitor]: which supports media and legal teams?” — Differences attributed to each company
• Industry — “How does a pharmaceutical company protect its brand online?” — Industry coverage and supporting evidence
• Reputation — “Is [brand] trustworthy?” — Factual errors, tone, and displayed sources

Do not build only brand-friendly questions. “Why is Branddi the best?” measures a prompted answer, not an actual market decision. Use the language a buyer would use before knowing the company. Marketing, legal, and e-commerce teams should each contribute questions when their needs differ.

How do you create a comparable protocol?

Write a few rules before the first run. They stop the report from treating different experiences as equivalent.

• Keep the core wording fixed. Treat a new wording as a new question.
• Record the context. Save platform, mode, date, time, language, country, device, and whether the conversation was new.
• Separate discovery from follow-up. Run the core question in a new chat, then use a planned follow-up if relevant.
• Document the competitor list. A list may change, but the change must be visible in the history.
• Keep evidence. Associate screenshots, answer text, links, sources, ad labels, and destinations with the observation.
• Respect platform terms. Scale does not justify collection methods that violate terms of use or obscure the method.

For a Brazilian brand, begin in Brazilian Portuguese. If the company operates internationally, create separate English and Spanish sets. Literal translation can change intent: “brand protection” may be understood as branding, trademark registration, or fraud monitoring depending on the market.

What should you record for each ChatGPT, AI Mode and Gemini observation?

Set a cadence the team can maintain. A weekly run can be a starting point; incidents or relevant changes may justify another frequency. Document changes to the schedule. Save the exact question and complete answer before classifying the result.

Copy this template into one record per answer or turn the fields into spreadsheet columns. Each repeated question should keep its question ID and receive a separate observation ID.

• Field — What to enter
• Question — ID, version and complete prompt
• Engine — Product, mode and model if the interface identifies it
• Date — Date, time and time zone
• Presence — Yes, no or not assessed; which entity was identified
• Recommendation — Yes or no, with the supporting passage
• Source — Cited URL; first-party, third-party or none visible
• Context — Language, country, new session, search and known settings
• Evidence — Complete answer, screenshot and internal file reference
• Collection quality — Valid, failed or incomplete, with a reason
• Next action — Owner and hypothesis to investigate

The official ChatGPT search documentation says results and citations appear when search is used. Record only the sources displayed: without a visible link, do not assume a page supplied the answer. OpenAI distinguishes OAI-SearchBot for search from GPTBot for training. Crawler access is not a recommendation metric.

Google announced generative AI reports in Search Console and states they became available worldwide on August 31, 2026. They complement observation within Google, but do not show ChatGPT answers. Sources in this section were accessed on October 17, 2026.

A collection failure is not brand absence. Keep the row and its failure reason, excluding it from the denominator of valid answers. For visits to the website, use the separate guide to AI traffic in GA4.

How do you classify answers without creating a misleading score?

Counting mentions alone exaggerates performance. A brand can be an irrelevant example, be cited without recommendation, or be connected to wrong information. Preserve the context with clear categories.

• Classification — Definition — Action
• Primary recommendation — The AI recommends the brand and explains its fit — Check that the rationale is accurate and supportable
• Secondary recommendation — The brand is an applicable alternative — Identify what is needed to compete for the first indication
• Neutral mention — The name appears without a buying recommendation — Evaluate whether the association is useful and accurate
• Absent — The brand does not appear in a strategically relevant question — Review intent, public coverage, and actual offer differences
• Risk — The answer contains an error, negative association, or confusion — Preserve evidence, check its origin, and prioritize a correction
• Sponsored element — The interface shows an ad or commercial label — Record it separately from organic recommendation

Add a short rationale to every classification. “Competitor X was recommended because it offers marketplace integration” is more useful than “we lost this question.” It lets the team test whether the gap is real, the company’s message is incomplete, or a public source is outdated.

Record risk and sponsored elements as additional flags: a mentioned brand can also be associated with an error. Do not add these categories as though they were separate answers.

How do you compare weeks without mixing mentions, sources and recommendations?

When questions are repeated, use valid answers as the unit. Count each answer once in each indicator. An answer containing three first-party links is still one answer with a first-party source, not three results. Calculate presence, recommendation and source rates separately over the same set, and report collection coverage.

Entirely fictional example: ten questions without brand names, repeated twice on the same engine each week, produce twenty valid answers per period. These are not Branddi or client results.

• Indicator — Week A — Week B
• Valid answers — 20 of 20 planned — 20 of 20 planned
• Brand presence — 4/20 = 20% — 8/20 = 40%
• Recommendation — 3/20 = 15% — 5/20 = 25%
• First-party source displayed — 2/20 = 10% — 3/20 = 15%

Presence rose by twenty percentage points, while first-party sources rose by five. This describes the sample; it does not establish that a content edit caused the change. A first-party link can also appear without an explicit brand mention. Preserve that combination in the record.

Compare by question, engine and intent first. If one engine has more answers than another, the combined figure will reflect that difference. Report counts, failures and methodological changes. Do not add new questions to the old series as though the set had remained unchanged.

Record which pages were updated and when. Closely timed publications, AI product changes and other sources limit causal attribution. Movement across a few answers is a reason to investigate and keep observing, not a guarantee of a trend or sales.

What should you do with findings, and when should you expand the operation?

For each relevant absence, check whether the company addresses the need and explains it publicly. For a factual error, preserve the passage and check the source before editing the website. The guide ChatGPT shows your brand incorrectly: how to fix it covers that investigation.

When the team can no longer repeat questions, review classifications or preserve evidence, assess operational support. Define coverage, cadence and ownership before expanding volume. For the business case, see how to demonstrate the ROI of AI Visibility. This article focuses on the protocol and record.

A decision should connect a question, evidence, hypothesis and action. Updating a page that already addresses the intent can be better than creating another on the same subject. Document the intervention and retain the questionnaire to track the outcome.

What questions come up when starting monitoring?

How many questions should we monitor?

We suggest 15 to 25 as an initial operational range, without automatic statistical validity. Prioritize varied intents and the capacity to maintain cadence and review.

Is a cited brand necessarily recommended?

No. Record the name in the answer, its recommendation as a solution and the presence of a first-party link separately. They can appear together or independently.

Can a failed observation count as absence?

No. Missing answers, errors and incomplete collection need their own status. Report the number of planned and valid observations before calculating rates.

Does Search Console replace this monitoring?

No. It tracks website performance within Google. The question set records the brand's role in specific answers and can cover other AI products.

To structure the routine, explore Branddi AI Visibility. Monitoring covers questions, context, mentions, sources and competitors within the scope agreed with the brand.

Back to blog