Recommendation Campaigns

Recommendation campaigns rank catalog products for a site and placement. Their quality depends on catalog health, stable interaction events, and a configuration that has an explicit fallback.

Before you create a campaign

• The site is active and the browser SDK is installed.
• Active catalog products have stable IDs, titles, prices, currency, availability, and safe destination URLs.
• Product view, cart, purchase, recommendation impression, and recommendation click events use the same product IDs.
• Embeddings are ready for content-based or hybrid behavior. Product updates normally queue refresh work automatically.

Choose an algorithm

Algorithm	Use when	Important controls
Rule based	You need deterministic merchandising logic.	Same category/brand, price range, frequently bought together, recently viewed, popular; weighted rules.
Content based	Product attributes and embeddings are strong.	Similarity threshold, category weight, attribute weights.
Collaborative	Interaction volume is sufficient.	Model version, minimum confidence, popular fallback.
Hybrid	You want behavior plus content fallback.	Collaborative/content weights and popular, recent, or random fallback.

Long-lived model registry and background training hardening remain separate operational work. Do not promise collaborative lift without enough interaction history.

Configure placement

1. Choose the site and campaign name.
2. Select carousel, grid, or list.
3. Choose product page, cart, checkout, home, category, search results, or custom placement.
4. For a custom placement, provide a stable CSS selector owned by the storefront theme.
5. Set minimum and maximum product counts. The renderer should not display an underfilled result below your minimum.

Configure targeting

Recommendation campaigns can use page patterns, user segments, and device types, and can exclude purchased or out-of-stock products.

Product activation can additionally use consented profile boosts and intent boosts. Recently purchased IDs are excluded unless replenishment is allowed. The reach estimate shown for intent targeting is an identity-graph inventory estimate, not guaranteed changed output for every request.

Test and activate

1. Save as Draft.
2. Request recommendations from a matching page and site.
3. Verify safe product URLs, currency, availability, count, and fallback behavior.
4. Check identified and anonymous visitors separately.
5. Activate only after impression and click tracking are present.

For controlled comparisons, create an A/B test with 2–5 variants, exactly one control, traffic percentages totaling 100, confidence from 0.80–0.99, and a minimum sample size of at least 100. Test lifecycle states are draft, running, paused, and completed.

Measure

The SDK and custom surfaces use:

• recommendation_impression with recommendation_id and campaign_id;
• recommendation_click with those IDs plus product_id;
• downstream purchase or configured conversion events for attribution.

Analytics labels product results as product attributed unless a broader trusted conversion source is connected.

Troubleshooting

• No products: check site scope, active status, catalog status, placement context, minimum count, and campaign status.
• Irrelevant results: confirm product IDs and categories, embeddings, algorithm weights, and fallback strategy.
• Wrong visitor context: verify identify(), device ID behavior, consent, and activation fields allowed for recommendations.
• Quiz result failure: a quiz request with require_campaign=true fails for a missing, cross-site, paused, or archived recommendation campaign instead of silently falling back.