Similar products release (June 2026)

The recommendation engine's sixth algorithm (previously marked "coming soon") is live, along with a new way to seed recommendation placements from each shopper's own browsing.

E-commerce projects using the recommendation engine. If your behavior-based recipes leave new products, new stores, or low-traffic categories uncovered, or you want a "picked up where you left off" placement, this release is for you.

Similar products: recommendations with zero traffic

The new Similar products algorithm matches products by what your catalog says about them: each product's title and description are turned into embeddings (numerical fingerprints of meaning), and for every product the engine finds the others whose text reads most similarly. The results are built ahead of time during the run and written to a normal versioned output dataset; serving stays the same fast edge lookup as every other recipe.

Because it reads only your catalog, it produces full results from day one with zero shopper traffic: the cold-start answer for new stores and new products, and a natural fallback step for behavior-based recipes.

Three parameters: topK (items per seed, default 12, max 50), sameCategoryOnly, and excludeOutOfStock (driven by the catalog stock field). Turning a Similar products recipe on kicks off an embed-and-build run right away, so the feed exists before the next nightly run. Activating a new catalog version does not trigger a run by itself; the recipe picks up the new catalog on its next scheduled or manual run. The recipe page shows run progress ("Embedding catalog…", "Building similar-products feed…") plus new run stats: products embedded, skipped for having no text, and seeds with no similar matches.

The Recently viewed seed source

Recommendation tests, the dedicated experiment type available at the time (retired 12 July 2026, see Recommendation experiments release), gained a fourth seed source: Recently viewed. The snippet keeps the shopper's last 20 viewed products in their browser (product IDs only, no personal data) and seeds the placement from the 3 most recent, "picking up where you left off" on homepages, cart pages, and landing pages that have no single current product. Shoppers with no viewing history see nothing and count no impression, so first-time visitors do not dilute results.

This seed source did not go away with the experiment type. It lives on as the context.recentlyViewed option on avsb.recs.get(), so any experiment, or any manual call, can use it today, not only the retired experiment type.

Views are recorded automatically wherever a placement resolves a product seed; for pages without a placement, a new public API, avsb.recs.trackView(sku), records a view explicitly.