Recommendations engine release (June 2026)
A vs B now computes product recommendations from your store's own event data, without a third-party recommendation service. The engine reads orders and product-view events you are already collecting, and runs a nightly ranking job. It writes results into a standard A vs B dataset: versioned, rollback-able, and readable from the snippet on any page.
E-commerce projects using the Shopify integration or purchase tracking. Any project that collects orders or product-view events can use the engine. The Shopify pixel provides product-view data automatically; other platforms need a product_view custom event metric on product pages.
Recipes
A recommendation rule is called a recipe. You create recipes at Commerce → Recommendations, pick one of five algorithms, set parameters, link a catalog dataset, and enable. A vs B runs the rest.
Read more: Concepts, Setup Guide.
Five algorithms
- Bestsellers: products ranked by units sold, over a rolling window. Optional per-category lists.
- Trending: products ranked by the ratio of recent sales to baseline rate. Surfaces products gaining momentum.
- New arrivals: products sorted by
createdAtfrom your catalog dataset. No order data needed. - Viewed together: products most often viewed alongside a given seed product, from product-view events.
- Bought together: products most often purchased alongside a given seed product, from order data.
A sixth algorithm, Similar products, appears as "coming soon" and ships in a later phase.
Outputs are datasets
Each recipe writes its results into a system-managed FEED dataset. Every nightly run creates a new dataset version and activates it atomically. Rollback works like any other dataset version: re-activate an older version from the Datasets page. A failed run never replaces the previous version.
Read results with avsb.dataset(slug).get(key) in the snippet or client.datasets.get(slug, key) in the Node SDK. See Serving & Reading Datasets.
Catalog enrichment
Linking a TABLE dataset keyed by SKU enriches output items with product titles, images, links, and prices. Without a catalog, output contains only { sku, score }. SKUs missing from the catalog are dropped from results and counted in run stats, so there are no broken cards.
New arrivals requires a catalog with a createdAt field.
Fallback chains
Each recipe can have an ordered fallback chain (up to four steps of other recipes or datasets). When the primary recipe has no row for a requested key, the chain is walked in order. Serve-time fallback resolution and the drop-in recommendation ribbon ship in the next phase.
Run schedule
Enabled recipes run nightly at 03:00 UTC. A "Run now" button on the recipe page triggers an immediate run, rate-limited to once per 10 minutes.
What's next
- Drop-in recommendation ribbon: a pre-built UI component that reads a recipe's output dataset and renders a carousel, with built-in fallback chain resolution.
- Recommendation experiments: A/B test two recipes against each other by assigning different output dataset versions per variation.
- Similar products: vector-embedding-based similarity using product attributes.
What became of each, as of today:
- Similar products shipped as described. Read Similar products release.
- Drop-in recommendation ribbon and Recommendation experiments shipped together in June 2026, as a dedicated Recommendation test experiment type.
- A vs B retired that type on 12 July 2026. A simpler mechanism replaced it. Any experiment is measured the same way now, once its code calls the recs API. Read Recommendation experiments release for the full history.
Released June 2026.