Multi-Modal Recommendation Case Studies
Evidence for computer vision, visual similarity, complete-the-look and recommendations that combine images, attributes and behavior.
When words are not enough, show the model the product.
Computer vision, product attributes and behavioral signals working together across visually similar, complete-the-look and related-product experiences.
Retail discovery becomes richer when the model can understand silhouette, color, composition and product context—not only keywords and co-purchase history.
How it works
- See: Represent visual form, style and product composition.
- Connect: Combine image similarity with attributes, behavior and category logic.
- Convert: Use the resulting relationships in grids, recommendations and complete-the-look sets.
Multi-modal discovery in practice
Specific implementations of visual similarity, style-led cross-sell and cold-start discovery.
- PVH Group (Tommy Hilfiger) — Mixed-basket merchandising at PVH. +403% increase in mixed baskets (shoes + accessories).
- Fashion retailer (anonymized) — See Similar turns a dead end into the next click. Shopper-led similarity buttons on listing, search and recommendation carousels — including a next-best-item route out of out-of-stock size variants.
- Multi-brand fashion retailer (anonymized) — Doubling the head of the range without buying more stock. Computer vision solved the cold-start problem on new arrivals, spreading demand across the catalog instead of concentrating it in a handful of SKUs.
Measured results
- Personalized recommendations across Life Pharmacy — Life Pharmacy. Controlled uplift: +23.2% Visitor conversion lift; +14.5% Revenue per visitor. +23.2% lift in visitor conversion rate; +14.5% revenue per visitor.
- Mixed-basket merchandising at PVH — PVH Group (Tommy Hilfiger). Controlled uplift: +403% Mixed-basket lift. +403% increase in mixed baskets (shoes + accessories).
- Discovery for a high-consideration bathroom retailer — The Blue Space. Compared with sitewide: +27.5% Post-click AOV vs sitewide; +17.6% Post-click units per transaction vs sitewide; 11.2% Click through rate; 21.2% Click conversion rate. Custom-filtered AI bundles delivered +27.5% post-click AOV and +17.6% units per transaction versus sitewide.
- See Similar turns a dead end into the next click — Fashion retailer (anonymized). Campaign performance: 60.1% Click through rate; 2.5% Click conversion rate. Shopper-led similarity buttons on listing, search and recommendation carousels — including a next-best-item route out of out-of-stock size variants.
- Limited drops, fanatical fans, bigger baskets — DTC boho fashion label (anonymized). Controlled uplift: +5.7% Average order value uplift; +12.3% Units per transaction uplift; 16.7% Click through rate. Multi-modal recommendations merged similar-item discovery with cross-sell and style buys on a direct-to-consumer own-brand range.
- Doubling the head of the range without buying more stock — Multi-brand fashion retailer (anonymized). Controlled uplift: +21.1% More SKUs in head of range; +14.7% Average order value; +11.9% Revenue per visitor. Computer vision solved the cold-start problem on new arrivals, spreading demand across the catalog instead of concentrating it in a handful of SKUs.