Skip to main content
Use this checklist for a homepage feed, related-items rail, checkout upsell or email digest. Start with one surface so you can verify the full loop before adding others.

​
Prepare the data

  • Ingest the items the placement may show and store their returned integer IDs.
  • Include the metadata you need for eligibility and rendering: availability, category, URL and image, for example.
  • Confirm a known user event appears in Data → Events with the correct item and event-type IDs.
  • Decide the success measure and the event that represents it. A served list, a visible impression, a click and a purchase are different observations.
See Catalogue and Events for the integration contract.

​
Configure the context

  1. Open Relevance → Contexts and create a context from the closest template.
  2. Give it the placement’s name and record its integer ID.
  3. On Details, choose Automatic or the trained model you intend to use.
  4. Review Inputs, Ranking and Safety. Enable only sources for which you have data.
  5. Add rules for required eligibility and merchandising in Relevance → Rules Engine.
  6. Save the recipe, enable the context and review readiness.
Keep mandatory eligibility checks in every variant you test. For example, turning off all rules to measure a promotional boost would also turn off an availability rule in that stage.

​
Integrate and verify

Call your backend with a representative visitor and request the new context_id. For related items, provide an integer seed item_id; for a basket rail, provide the basket item IDs through a supported request path. Verify these cases before sending real traffic: Use Explainability for the served request and compare it with the current configuration. Recomputing an explanation after editing a context is not a replay of the earlier decision.

​
Release and measure

Start with a small, controlled audience. Keep the context, event mapping and fallback stable while collecting a baseline. If you compare two configurations, use Experiments and state the hypothesis before starting. Review performance, then increase traffic when the observed result supports it. To roll back, restore the previous context settings/model or stop routing the placement to the new context. Keep a note of the owner, launch time, hypothesis and rollback choice.