Recommendations
MXP RecSys — personalized, ML-driven product recommendations across web, mobile, and offline channels
MXP RecSys delivers personalized product recommendations to shoppers across the Website UI, Mobile App, and offline prediction channels. Recommendations are powered by custom ML models trained on real user behavior — views, clicks, and purchases. Merchandisers control which models are active and how they are configured entirely through MXP UI, with no code changes required.
What it solves
Algorithmic search ranking answers the query use case, but it doesn't serve adjacent moments: what to show on a product detail page when there's no active query, what to suggest alongside the cart, or how to present tiered price options within a category. Recommendations fill those gaps with context-aware, personalized product lists tuned to each placement.
Recommendation models
MXP RecSys currently supports five model types, each aimed at a different shopper moment:
| Model | What it recommends | Best used for |
|---|---|---|
| Customers Also Viewed (CAV) | Items other users browsed after viewing this product | Product detail pages — "explore similar" |
| Frequently Bought Together (FBT) | Items commonly purchased alongside this product | Cart page — cross-sell |
| Similar Items | Products similar to the one being viewed | Product detail pages — alternatives |
| Good / Better / Best (GBB) | Three price-tiered options within the same category | Category pages |
| GetBundle | Complementary products from related categories, as a bundle | Product detail pages — complete-the-setup |
Each model type is covered in more depth on the Recommendation Models page, and its internal architecture in the Model Architecture reference.
How it works
At a high level, MXP RecSys is split into two layers:
- Serving layer — handles real-time requests. When a shopper opens a page, the PredictApp service reads the active configuration, calls the relevant model(s), applies personalization and post-processing, and returns a ranked list of products. Recent session activity and brand preferences personalize the result.
- Config & Training layer — handles the model lifecycle. ML engineers train models on Vertex AI Pipelines; experiments and versions are tracked in MLflow. Merchandisers then deploy and configure those models through MXP UI, which talks to the EditApp service. All placement configuration is stored in GCS and picked up by PredictApp on the next request — no restart needed.
A handful of offline jobs precompute data the serving layer reads at request time (product embeddings, bundle plans, and per-user brand affinity). For the full picture — services, data stores, and data flows — see the Technical Architecture.
Placements are configured, not coded. A Recommendation Container maps a page placement to one or more models. Swapping the model behind a placement is a configuration change in MXP UI — the storefront keeps calling the same container.
Related pages
- Recommendation Models — the five model types, and how to deploy them
- Recommendation Containers — map placements to models
- Recommendation Rules — boost, bury, and filter recommendation output
- Technical Architecture — services, data stores, and flows
- Recommendations API — REST endpoints for inference and configuration