Braze product recommendations guide: combine customer behavior and product data to drive relevant cross-sell, upsell and discovery.

Build the product data foundation first

Recommendation quality depends on consistent item identifiers, useful product metadata and behavioral signals. Braze documents patterns where recommendation identifiers are stored on user profiles and product metadata is retrieved through Catalogs or Connected Content. Design that boundary so the recommendation engine can change without rebuilding message logic.

Implementation note: Keep product IDs stable, normalize category and price fields, and define what happens when a recommended item is missing or discontinued.

Choose where recommendation logic lives

A recommendation can be generated by a rules engine, a separate recommender, or Braze capabilities. The architecture decision should follow the required latency, explainability and data freshness. Do not force a large external response into message-time retrieval when a small set of user-specific product IDs is enough.

Implementation note: For complex recommenders, store compact recommendation outputs and keep heavy ranking calculations outside the message render path.

Activate across channels safely

Email, push, in-app and other channels can reuse the same product identifiers while rendering different creative. Keep channel-specific presentation separate from recommendation logic, and test inventory, price, currency, image and locale variations independently.

Implementation note: Use a fallback such as popular, recent or curated items when a personalized recommendation is unavailable, and make that fallback measurable.

Measure the recommendation decision

Track the recommendation exposure and compare downstream outcomes such as add-to-cart, purchase, repeat purchase or revenue per recipient. Separate recommendation quality from message quality when analyzing results.

Implementation note: Use holdouts or controlled comparisons when practical, and monitor category or margin guardrails so the optimizer does not chase clicks at the expense of business quality.

Production scenario: abandoned-cart recovery

A shopper leaves two products in a cart. The Canvas checks eligibility, retrieves the saved item IDs, enriches them from the catalog, and renders a small set of alternatives when a product is unavailable. The journey records whether the shopper clicked, viewed, purchased or ignored the recommendation so the recommendation system can be evaluated separately from send timing.