Treat a model change as a measurement event. Preserve the old series and decide whether the new configuration deserves a separate baseline before describing movement as a change in brand visibility. This applies to SaaS marketing teams and agencies maintaining trend reports.
Record what changed
Create a ledger entry with the provider, API or consumer surface, requested model identifier, returned identifier when available, collection time, prompt-panel version, search configuration and scoring version. Distinguish a change you made from a change you only suspect.
Consult the provider's model catalog for current model identities. A name in a marketing report is not enough to establish which model actually served a historical request. Avoid reconstructing missing identifiers from today's defaults.
Run an overlap comparison
If both configurations remain available and you have an approved collection budget, run the same frozen questions on both during a short overlap window. Keep those observations separate from the normal historical series.
Illustrative example: an unchanged question mentions a brand in two of three old-configuration observations and one of three new-configuration observations. That is a small observed difference. It is not proof of a market decline or proof that the model change caused it.
Review answer validity and source patterns alongside mentions. A configuration that begins using web search can change the evidence available to the answer, making a single combined trend difficult to interpret.
Choose the reporting treatment
- Continue with an annotated event only when the measurement contract remains compatible and you can explain why.
- Start a new series when model or retrieval changes materially alter the observation surface.
- Publish parallel series during an evaluation period when compatibility is uncertain.
- Mark the event unverified if the provider does not expose enough information to identify the serving configuration.
Do not splice two series simply because their percentages look similar. Agreement on one period does not establish equivalence for every question.
Update the report footnote
A useful note says what configuration changed, when it changed, which comparisons remain valid and what evidence is unavailable. Keep the previous definition available so a reviewer can reproduce the interpretation.
The existing change-monitoring guide covers prompt-level comparisons. Next, review the provider and model information associated with an observation in the AI visibility workflow. If the information is incomplete, label the comparison uncertain and begin recording the missing provenance prospectively.
Fact-checked: 17 October 2026. Examples are illustrative; no customer results are reported.