I’m designing a multimodal application that uses public web data as one input source, and I’m trying to understand how other teams evaluate regional variation.
The same page, image, advertisement, or search result can change by country or city. For a Gemini-based workflow, should the evaluation set treat these as separate slices, or should the application normalize them before sending them to the model?
The checks I’m considering are:
- requested versus observed location
- language and local formatting
- image-text consistency
- freshness of the source
- answer quality across regional variants
Would appreciate examples of evaluation designs or production lessons. The collection layer is being built around permitted public sources, with location-aware network access where the task requires it.