Product recommendations - “customers also bought”, “you may like” - are proven revenue when they are relevant and noise when they are not. Adobe Commerce’s Sensei-powered Product Recommendations service does the relevance work in Adobe’s cloud: your catalogue and shopper behaviour go up, personalised recommendation units come down. Here is how to deploy it so it actually lifts revenue.
How It Works
Two data feeds power the service: your catalogue (synced via the SaaS data export, shared with Live Search) and behavioural events collected by the storefront module (views, adds-to-cart, purchases). Sensei trains per-unit models on that behaviour and serves recommendations by API at render time. No model training on your side - the trade is that your data feeds must be healthy or the recommendations quietly degrade.
The Unit Types Worth Placing
Start with the placements with the strongest evidence:
- Product page - “Customers also viewed/bought”: the classic; consistently lifts items per order
- Homepage - “Recommended for you”: for returning visitors with history; falls back to trending for anonymous
- Cart page - “You may also like”: last chance cross-sell; keep it tight so it does not distract from checkout
- Category pages: trending-in-category units help discovery in large catalogues
Recommendation types differ: behavioural (viewed-this-viewed-that), content-similarity, trending, and recently-viewed. Match type to placement intent - similarity for PDP cross-sell, behavioural for cart, trending for cold traffic.
The Unsexy Prerequisite: Data Health
- Catalogue sync must be running: check the SaaS export status in admin. A stalled feed means recommendations stop reflecting new products and stock
- Events must fire: verify the storefront collector in network tools; a theme change that breaks event collection silently ages every model
- Attribution must work: Sensei reports view/click/revenue per unit only if the recommendation links carry their tracking - do not strip it in templates
Pitfalls We See
- Recommending the unavailable: stock sync lag shows out-of-stock products - verify availability filtering is on
- Too many units: three recommendation rails per page dilute each other and slow rendering. Two, placed well, outperform five
- No control for new products: cold-start items get no behavioural traction - use “similarity” units or merchandising boosts to give launches exposure
- Judging by eye: “those recommendations look wrong” is not measurement
Measuring Uplift
Sensei’s dashboard reports unit-level views, CTR and attributed revenue. Attributed revenue flatters - a shopper who clicked a recommendation was often going to buy anyway. For the true number, run a holdout: hide units from a slice of traffic for a few weeks and compare items-per-order and revenue per session. The delta is what you are actually paying for in data and complexity.
Deployed with healthy feeds and honest measurement, Sensei recommendations are one of Adobe Commerce’s more quietly valuable features - personalisation without a personalisation team.