AI Search
How to monitor your brand in ChatGPT: a question set
Branddi · Published on
To monitor your brand in ChatGPT, build a fixed question set that represents real purchase, reputation, and comparison decisions. Run it periodically in new conversations and record the platform, language, country, recommended brands, sources, and evidence. The purpose is not to force an answer; it is to spot change and decide what to improve in the company’s public information.
What is an AI question set?
An AI question set is a small, stable collection of prompts that a company repeats to observe how its brand appears in answer engines. It turns an isolated impression — “ChatGPT did not mention us today” — into comparable observations with the same intent, language, evaluation rules, and evidence trail.
It does not measure an official ranking. ChatGPT, Google AI Mode, and Gemini may answer the same question differently, evolve over time, and draw on different sources. Google explains that AI Mode and AI Overviews can display different answers and links, and that meeting technical requirements does not guarantee a page will appear. The sound interpretation is therefore specific: in this question, context, and date, the brand played this role.
Direct observation should sit alongside traffic data. On June 3, 2026, Google announced dedicated Search Console reporting for generative AI features, initially for a subset of sites. It can show impressions, pages, countries, devices, and change over time in Google Search; it does not replace examining the answer a person actually receives.
Which questions belong in the first version?
Start with 15 to 25 questions. Fewer often reflect only one intent; more make the routine difficult to sustain. 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.
How should the collection run in ChatGPT, Google AI Mode, and Gemini?
Choose a cadence that matches the risk. A weekly run is enough for an initial diagnosis; campaigns, launches, crises, or major product changes merit more frequent checks. Consistency matters more than asking the same tool dozens of times in a day.
For every run, record the exact prompt and capture the answer before continuing. A spreadsheet is enough at first if it includes: question ID, platform, date, brand presence, presence type, highlighted competitor, cited source, tone, identified risk, evidence, and owner of the next action.
In ChatGPT, note whether the answer included search, links, or sources. OpenAI’s publisher guidance says public sites can appear in search and that allowing OAI-SearchBot helps discovery and citation; it does not promise inclusion in any particular answer. That is why outcomes must be measured rather than inferred from crawl access.
In Google AI Mode, also use Search Console when generative-feature reporting is available to your property. In Gemini, treat each answer as evidence that needs review: Google warns that Gemini responses can be inaccurate. The rule applies everywhere: an incorrect response is a signal to investigate, not a fact that should drive a campaign.
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.
Which metrics help leadership make decisions?
There is no single reliable ranking across all AI products. A compact dashboard should combine coverage, quality, and change between periods.
- Presence rate: questions where the brand appeared divided by all monitored questions.
- Recommendation rate: questions where the brand was recommended divided by all questions.
- Share of voice by intent: the brand’s share of recommendations or mentions within each question group.
- Citation rate: answers that displayed a first-party link or source when the brand appeared.
- Risk incidence: answers with error, inappropriate association, or concerning tone.
- Change: questions that gained or lost presence since the previous run.
Compare periods with the same set. Changing ten questions, adding competitors, and changing country mid-month makes movement ambiguous. Do not turn a mention into a sale either: read the AI panel alongside leads, referral traffic, conversion data, and Search Console.
What should you do when the brand disappears or is described incorrectly?
Start with a verifiable cause. Does the question describe a need the company actually solves? Does a public page explain that solution clearly? Is there a case study, FAQ, documentation, or product page that supports the claim? Is the source surfaced by the AI accurate and current?
The usual actions are within the company’s control: update solution content, publish specific evidence, correct official data, clarify geographic coverage, review business profiles, or create a page for an intent that has no public answer. There is no legitimate “hack” that can force an AI to recommend a company. Branddi’s guide on how AI search engines choose which brand to cite explains what makes a source useful and discoverable.
For competitive analysis, pair the set with how to find out whether a competitor is recommended before your brand. Monitoring becomes action when every finding has a hypothesis, owner, and due date, without promising that one edit will change every answer.
Frequently asked questions
How many questions should we monitor?
Start with 15 to 25 well-chosen questions. Expand only when the team can sustain frequency and classification quality.
Can we use a single AI Visibility score?
An internal score can be useful as a summary if it does not hide the evidence. Also show presence, recommendation, risk, and the questions that explain movement.
Is a cited brand necessarily recommended?
No. Citation, mention, and recommendation are different signals. The classification should state the brand’s role in the answer and the rationale offered.
Does Search Console replace this monitoring?
No. It measures site performance in Google and may report generative features when available, but it does not compare brand roles in each answer or cover the other platforms.
A well-maintained question set shows where a brand gains relevance, where competitors take space, and where misinformation creates risk. Branddi AI Visibility turns that monitoring into comparable evidence for marketing, media, legal, and leadership.
Sources consulted
- Google Search Central — AI Features and Your Website, updated December 10, 2025.
- Google Search Central — generative AI performance reports, June 3, 2026.
- OpenAI — Publishers and Developers FAQ, accessed August 27, 2026.
- Google Gemini Apps Help — Learn about responses from Gemini Apps, accessed August 27, 2026.