What Should an Agency White-Label Report Disclose About Model Names, Dates, and Repeated Tests?

Why does disclosure matter more for AI visibility than for SEO reports?

Rank tracking is fairly stable: a keyword, a location and a date define a result. AI answers vary between runs, differ by surface, and change when providers update models. A client who checks a question themselves and gets a different answer will doubt the whole report, unless the report explained what it measured and how. Disclosure also protects the agency at renewal time. A drop caused by a model update is easier to explain if the report already records which model and period each figure covers. Why clients often cannot reproduce agency results is covered in why an agency's GEO report shows citations the client cannot reproduce in Perplexity.

What should the report disclose?

How should model names be reported?

Model names are the hardest item, because many consumer apps do not show which model produced an answer, and providers change default models over time. Report what is actually known:

• Where the API reports the model, name it exactly.
• Where a provider has announced a default, state it with the source. Google, for example, announced in January 2026 that Gemini 3 had become the default model for AI Overviews globally.
• Where the model is unknown, write "consumer app, model not disclosed" rather than guessing.
• Where a known model change happened during the period, mark it on trend charts.

Never present a guessed model name as fact. A client who later learns the model was different will question everything else in the report.

How should repeated tests be reported?

State the number of runs, and report figures as shares or ranges, not single outcomes:

• "Named in 7 of 9 runs (3 engines × 3 runs)" is clear.
• "Mention rate 42%, based on 540 runs across 60 questions" gives the base.
• For small samples, show the counts, not only the percentage.
• When comparing periods, say whether the change is larger than normal run-to-run variation.

If a figure rests on a single run, label it as such, or leave it out.

What does a method box look like?

A short box on the first page, in plain language, covers most readers' questions:

How this report was produced. 60 buyer questions (48 unbranded, 12 branded) were run three times each in ChatGPT (consumer app, web search on), Gemini (consumer app), Perplexity (default mode) and Google AI Overviews. Runs used a clean, logged-out setup where possible, with the location set to Chicago, US, in English, between 1 and 28 September. Figures show the share of runs in which each brand was named. Twelve per cent of runs were checked by hand; two brand-detection errors were found and corrected. Changes since the last report: four questions replaced (listed in the appendix); ChatGPT's default model changed on 12 September (marked on charts).

Put the full question list, settings and any changes in an appendix.

What about the white-label vendor itself?

White-labelling is common, and agencies do not usually name their tool vendor in reports. Three things still need care:

• Data protection. If personal data, such as customer names or contact details, passes through the vendor, the vendor is a subprocessor, and your data processing agreement with the client may need to list it. Confidential prompts, such as questions about unreleased products, are covered by your confidentiality terms instead. Your data processing agreement with the client may need to list it, whatever the report says.
• Method. Clients should know answers were collected through a third-party tracking platform and how, even if the platform is not named.
• Contracts. If the client might take the programme in-house, agree whether raw data can be exported to them.

What should change notes cover?

Each report should list what changed since the previous one:

• questions added, removed or reworded;
• engines or surfaces added or dropped;
• known model or product changes at the providers;
• changes to locations, languages or run counts;
• changes to the tracking tool or collection method.

Changes like these explain many sudden movements. Recording them prevents both false alarms and false celebrations. The checks to run before a report goes out are covered in what an agency should check before sending a client an AI visibility report.

What might an illustrative case look like?

Consider an illustrative agency whose monthly report shows a client's mention rate in ChatGPT falling from 38% to 24%. The client's marketing director checks five questions themselves and sees the brand in three. They ask whether the report is wrong. Because the report's method box states the surface, the logged-out setup, three runs per question, and a known model change mid-month, the agency can show that the drop coincided with the change. It appears across the client's competitors too, and the director's personalised account produced different answers. The conversation moves to what to do about the change, rather than whether the data can be trusted.

What do people ask about white-label AI visibility reports?

Should we name the tracking tool in white-label reports?

Describe the method in every report. Whether you must name the vendor depends on your contracts: if the tool handles the client's personal data, check your data processing agreement and the law that applies before leaving it unnamed.

How much methodology detail do clients want?

Most read a short method box. Keep the full detail in an appendix for those who check.

What if the tool does not tell us the model?

Say so. "Model not disclosed by the provider" is accurate and credible.

Should reports include raw answers?

Include examples for key findings, and make full answers available on request or in an export.