← All posts
AIbrand trackingmarket research

How to track what ChatGPT and other AI assistants say about your brand

Sajin Kabeer · October 22, 2026

Brands can track what AI assistants say by running a fixed panel of buyer questions through ChatGPT, Gemini, Perplexity, Claude and Google's AI features on a schedule. Log mentions, framing, accuracy, competitors and cited sources. Each answer is one observation, so report rates across prompts, models and relevant languages over time, not a single screenshot.

Here, AI brand perception tracking means measuring how assistants describe and recommend a brand, not what consumers think of AI companies. Decide whose questions matter, define the sample and test the AI narrative against people.

Disclosure: Meaningful publishes this guide and is one of the platforms discussed. The research findings cited below come from external preprints.

Why can't you just ask ChatGPT about your brand once?

One answer tells you what appeared in that session, not how often your brand appears for buyers. Wording, model, retrieval, language, context and time can change the response. A screenshot is useful for diagnosing an error; it is not a visibility trend.

The preprint Don't Measure Once treats AI-search visibility as a distribution. Another preprint on 12,933 brand-related answers finds variation across runs, systems and languages, but measures sentiment polarity, not mention rate in every consumer interface. Neither establishes a universal run count.

Define two metrics before you collect anything:

Keep branded checks separate. “What does the assistant say about Acme?” almost forces a mention and will inflate a competitive rate if mixed with unbranded buyer questions. A citation to Acme's website is also not the same as recommending Acme.

What should you record from each AI answer?

Save the full answer and its collection conditions, not just a positive or negative flag. Two researchers should be able to code the same output the same way and revisit a disputed claim months later.

FieldWhat to save or codeWhy it matters
Exact prompt and panel versionWording, prompt type and any personaA paraphrase can change the answer
Date and collection pathTime, interface or API, signed-in state, web accessDifferent paths may not show the same experience
Model, market and languageProduct or model label, location setting and query languageDo not pool unlike conditions by default
Full answer and outcomeText, refusal, missing feature or access failureKeep the denominator honest
Brand mentions and positionTarget and competitors; first, later or absentSeparates appearance from prominence
Framing and accuracyRecommended, neutral, discouraged; checkable claim and correctionA mention can be harmful or simply wrong
Cited sourcesURLs or source labels actually shownA citation is evidence of a surfaced source, not proof of persuasion

Code from the complete response, not a cropped screenshot. Count names in answer text separately from citation labels, and retain ambiguous cases rather than forcing a sentiment score.

How do you build a prompt panel that reflects real buyers?

Build the panel from the questions buyers ask at each decision point, then freeze a core set for comparison across waves. Use interviews, sales calls, support questions and search queries as candidate sources, but do not pretend those inputs reveal the full distribution of private AI conversations.

For a fictional trail-shoe brand, TrailNest, a small starting panel could include:

Prompt typeIllustrative questionInclude in competitive rate?
CategoryWhich trail shoes work on wet rock?Yes
Problem-ledWhat should a new trail runner look for in grip and fit?Yes
Buyer segmentWhich brands make trail shoes for wide feet?Yes
ComparisonHow do cushioned and low-profile trail shoes compare on wet trails?Yes
Branded accuracyIs TrailNest suitable for wet conditions?No; audit separately
Branded reputationWhat complaints do people have about TrailNest?No; audit separately

Add relevant markets and languages. Translate the buyer intent, keep the local wording and group prompts by use case. Version panel changes rather than drawing a continuous trend across incompatible questions.

Around 30 related prompts per topic is a vendor planning heuristic, not a validated minimum. Start with a defensible panel and expand where pilot variation warrants it. Apply the same market definitions used in cross-country research.

How many times should you run each prompt?

Repeat enough to see whether a result is stable, but do not spend the entire budget rerunning one wording in one model. Spread observations across buyer intents, relevant markets, languages and assistants, then compare rates over a consistent schedule.

Common adviceWhat the evidence supportsPractical decision
“Run each prompt three to five times”Repeats expose variation, but five is not a universal precision thresholdPilot several runs; keep run count fixed within a wave
“More repeats will make the metric reliable”The 12,933-answer preprint found diminishing returns after several repeats for its sentiment outcomeAdd relevant prompts, models or languages when those dimensions are missing; validate mention-rate precision separately
“A small weekly increase is progress”Differences can arise from sampling and changing product conditionsShow numerator, denominator and an empirical range from repeated baseline waves

In that sentiment study, query language accounted for 26.5% of the modelled variation and brand identity 1.5% of a single-response result. Those are properties of its sentiment measure and sample, not a claim that language always drives a quarter of mention-rate variation. The study's small gain from a sixth identical run likewise cannot tell every brand to stop at five. The separate visibility preprint supports the broader point: one observation is unstable.

Run priority prompts weekly and others monthly if feasible. Build a normal range before setting alerts; LLMPulse suggests eight weeks as a vendor heuristic, not a universal threshold. Treat tiny movements as uncertain until repeated data separate them from fluctuation.

Do AI tracking tools see what your customers see?

Not necessarily. A tool may query an API while a buyer uses a consumer chat interface with different model settings, search access, memory or presentation modules. Ask what the vendor actually ran before calling its dashboard a view of customer experience.

An industry explanation of collection methods by Metehan Yeşilyurt makes this distinction; its author works at Peec AI, so treat it as vendor-informed guidance. Official OpenAI documentation also shows that API web search is a configurable tool and responses can carry source citations. That documents an API path, not that any particular tracker reproduces a signed-out ChatGPT session.

Should you use a spreadsheet or a tracking tool?

Start with a spreadsheet when the panel is small and the team needs to establish definitions. Buy automation when repeated collection, exports and review across markets or clients become the bottleneck, while retaining raw answers for audit.

RouteExamplesStrongest useQuestion to ask
Manual logSpreadsheet and saved answer exportsA transparent pilot and codebookCan two coders reproduce the same mention and framing labels?
AI visibility platformProfound, Peec AI, Otterly, Ahrefs Brand Radar, SemrushRepeated prompt monitoring and competitor/source dashboardsWhich interfaces, models, languages, runs and prompt sets produced the metric?
Research-led workflowMeaningful, Listen LabsPut AI descriptions beside customer evidenceCan a researcher trace an AI claim to its answer and test it against a survey or interview?

Request the exact prompts, collection path, runs by model and market, raw answers and uncertainty method. Meaningful puts AI perception beside surveys and social research; Listen Labs uses interviews to test an AI narrative. Neither is simply an SEO dashboard.

Free signals measure something narrower. Search Console's generative AI Search report shows URL impressions in AI Overviews and AI Mode, not every brand mention; a separate Discover report exists. A property may not show data without enough impressions. Referral sessions reveal some clicks, not recommendations without clicks.

Does what AI says match what your customers think?

Treat an assistant's description as a hypothesis about the information it presents, not a measurement of customer belief. The right response to “competitor X is better value” is to check the cited sources and ask buyers what value means in the category.

In a hypothetical study, an AI assistant might call a rival a better value because its product page states a lower unit price. A survey could test whether target buyers actually perceive the rival that way and which attributes they compare; interviews could probe the reasoning. Social listening can reveal unprompted questions about pack size or durability, but it cannot estimate prevalence in the buyer population. Combining social listening with surveys and interviews keeps those methods in their proper roles.

If the assistant repeats an outdated factual claim, correct the owned source and document any relevant third-party source. If buyers share the perception, the client may have a proposition or product problem, not merely an AI visibility problem. If buyers disagree, show that tension in the report rather than forcing one source to explain the other.

What can an insight agency offer clients here?

Offer a recurring AI-perception wave within brand tracking: an agreed buyer-prompt panel, controlled collection, raw-answer log, reproducible rates and a comparison with customer evidence. An agency's value is the method and interpretation, not a promise to control an assistant's next answer.

The readout should show the brands and descriptions that appeared, the rates with denominators, the conditions under which they appeared, the normal wave-to-wave range and the claims worth testing. Add a separate accuracy register for errors and cited sources. Keep recommendations modest: correct verifiable misinformation, investigate a customer-perception gap and rerun the panel on the same terms.

For multi-country clients, produce market-language cuts rather than one pooled global score. An AI answer may be right for one market and wrong for another, just as a survey finding may differ by audience. This is an extension of an agency's existing research discipline, not a new universal ranking.

Frequently asked questions

AI brand perception tracking is repeated sampling of defined questions, not access to a census of private conversations. These answers separate what a brand can observe directly from what must be inferred cautiously.

Is it possible to track brand mentions in ChatGPT?

Yes, by running and saving answers to a fixed panel of relevant prompts on a schedule. Your panel estimates what appears under those conditions; it does not show every answer real users received.

How many prompts do I need?

There is no validated universal minimum. Group enough buyer questions to cover distinct intents and markets; around 30 per topic is one vendor's planning suggestion, not a guarantee of representativeness.

How often should I check?

Weekly checks can suit priority questions and monthly waves less urgent topics if you hold the method steady. Establish a repeated baseline before treating small changes as a trend or setting an alert.

Why do different tools show different numbers for my brand?

They may use different prompts, competitors, models, markets, repeat counts and collection paths. Ask for raw answers and denominators before comparing two dashboards' percentages.

What's the difference between AI brand monitoring and social listening?

AI monitoring records what sampled assistants say in response to prompts. Social listening studies public conversation people chose to post; neither substitutes for a designed consumer sample.

Can I see AI traffic in analytics?

You may see some visits referred by AI services, and Google Search Console can show impressions of your pages in Google's generative AI Search features. Neither tells you how often an assistant named your brand without sending a visitor to your site.

How do AI assistants decide which brands to mention?

The result depends on the question, context, model and whether it retrieves current sources. A displayed citation identifies a source used or surfaced in that answer, not a complete explanation of why a brand was selected.

Should we change our website to influence AI answers?

Fix verifiable errors on your own pages and investigate incorrect third-party sources, then rerun the same panel. Treat content optimisation as a separate project; an improved mention rate in your sample does not prove what all users will see.

Next step

See how assistants describe a client's brand next to what its customers say. Run a fixed panel, retain the raw answers and test one surprising claim against real research. If you want those sources in one project, try Meaningful for three days and judge the evidence trail as carefully as the dashboard.