Product · Recommendations
The next item worth showing — from the first visit.
Recommendations blend catalogue similarity, what visitors open together and each visitor's own history. A cold-start ladder keeps the row useful on day one, and diversification keeps it from repeating itself.
How “bought together” is learned
Edges between products, built from your visitors' visits. Thicker lines mean stronger evidence; bestsellers are weighted down so they don't pair with everything.
Rebuilt automatically from the last 90 days. An edge needs at least two visits. Manual “never suggest” links always win.
Models
Similar items
Item-to-item similarity over embeddings, without re-embedding the seed on every call.
Clicked together
Items opened in the same session as the seed — the cheapest behavioural signal, no long-term identity needed.
Personal
A visitor's decayed history of clicks and queries reshapes ranking; the weight ramps up as history grows.
Trending and popular
Index-wide popularity and this week's trending queries as cold-start fallbacks.
Quality
Explainable
Every item comes with a reason and a score breakdown — similarity, affinity, popularity — for debugging and trust.
Coming next: the product graph
Designed in detail; see the graph page.
Complements, not lookalikes
Learns what is bought together in one visit — the charger with the laptop. Ask for it with strategy: complement.
Toward the purchase
Will rank by how likely an item is to lead to a purchase, not only by similarity.
Specifications
- Items per request
- 1–50
- Excluded ids per request
- up to 200
- Strategies
complement,next,co_view, similar, personal, popular, trending- Visit window
- 30 minutes of inactivity ends a visit
- Product graph rebuild
- automatic, from the last 90 days of events
- Minimum evidence for an edge
- seen in 2 or more visits
- Personal signals
- used only with consent when the index requires it
Before you start
- Click and view events (the widget sends them automatically)
- Add-to-cart and purchase events from your backend for bought-together
- Product embeddings for similar items — or plain text fields as a fallback
How to start
- 1
Send events
The widget records clicks and views automatically; from your backend, post add-to-cart and conversion events with an order value.
POST /api/v1/indexes/products/events { "type": "conversion", "document_id": "sku-123", "value_cents": 459900 } - 2
Ask for recommendations
For a product page, a cart or a returning visitor.
POST /api/v1/recommendations { "index": "products", "seed_document_id": "sku-123", "limit": 8 }