Fashion search and recommendations

Help shoppers narrow a broad style request to products they can actually buy. Keep size, colour, variant availability and product similarity distinct in both the data model and the evaluation.

## Treat variants as part of the decision Record the relationship between a product family and its sellable variants. Decide how to show a style when the requested size is unavailable, and test that filters and recommendations agree with the product page.

## Describe similarity precisely A substitute may share silhouette, material or visual characteristics. A complementary item serves a different purpose. Write separate examples for “show me an alternative” and “complete this outfit” so one placement is not judged against conflicting goals.

## Keep seasonal controls accountable Document which campaigns or merchandising rules affect a result, who owns them and when they expire. A seasonal boost should be observable and removable without obscuring whether the underlying retrieval works.

## Evaluate the range, not only bestsellers Include new products, sparse descriptions, unusual sizes and products outside the current sales leaders. Measure relevance separately from the distribution of exposure and sales across the range. A change in product mix alone does not establish incremental revenue.