Recommendation algorithm test
Compares three ways of picking recommended products (also bought, recently viewed, trending) on cross-sell revenue.
Authored by Sid Chaudhary
Founder & CEO at Intempt
Agentic workflow
Coming soon1 step Blu runs end to end. It drafts a plan and nothing runs until you approve it.
- 1
Set up the algorithm test
Coming soonBlu builds a step
About this recipe
When to use it
Collaborative filtering, session-based, and popularity-weighted recommendations all sound reasonable in theory, and they perform differently in practice depending on your catalog. This tests all three on real revenue.
What it delivers
Runs a server-side experiment testing which recommendation engine - collaborative filtering, session-based, or popularity-weighted - drives more cross-sell revenue.
How you prompt this
Describe the outcome in your own words and Blu Agent maps it to this recipe, or run it by name. It drafts a plan first and nothing runs until you approve it.
Tell Blu Agent what you want
“Compares three ways of picking recommended products (also bought, recently viewed, trending) on cross-sell revenue.”
recipe.md preview
Built from the published catalog. Blu Agent's full step prompts load in your workspace when you run it.
Inputs & outputs
What the recipe needs before it runs, and what it leaves in your workspace.
Inputs
- Nothing declared beyond an Intempt workspace. Blu asks for anything else it needs as it goes.
Outputs
How the recipe runs
Blu follows the steps in order and shows you each one before anything goes live.
- 01Set up the algorithm testCreate a SERVER EXPERIMENT on /experiences titled "Recommendation Algorithm Test". ═══ PATH 1: Top-level configuration ═══ Experience type: server_experiment Flag Key: recommendation_algorithm Variants: - Control (33%): collaborative filtering: "Customers also bought" - Variant B (33%): session-based: "Recently viewed by you" - Variant C (34%): popularity-based: "Trending now" Targeting: - Pages: any page that renders the recommendation block (typically PDPs and category pages: page URL contains "/products/" OR "/category/") - Devices: any (server experiments are device-agnostic; the recommendation API is called server-side regardless of client device) - Audience: visitors with >1 View page in the current session (excludes one-and-done bouncers: recommendations are most useful for engaged shoppers) - Display frequency: always (every page render that includes the recommendation block calls getFlag and serves a variant) Primary metric: goal_completed_in_experience where experience_id = <this> AND value > 0 (revenue from Placed order where the order's items include a clicked recommendation) Secondary metrics: - Click on where target_id starts with "rec-" (recommendation click-through rate) - Placed order.total_price for orders following a recommendation click (revenue from recommended products) - Recommendation click position (which slot (1, 2, 3, etc.) the user clicked) - Add-to-cart-from-rec rate: Cart created where the added product_id matches a recently-clicked recommendation Guardrail: PDP View page to Cart created conversion rate must not drop >5% (avoid recommendations distracting from primary purchase decision) Schedule: 14 days ═══ PATH 2: Variant JSON payload ═══ Variant: Control (collaborative filtering) JSON payload: { "algorithm": "collaborative_filtering", "endpoint": "/api/recs/collaborative", "header_text": "Customers also bought", "max_items": 4, "fallback_algorithm": "popularity" } Variant: B (session-based) { "algorithm": "session_based", "endpoint": "/api/recs/session", "header_text": "Recently viewed by you", "max_items": 4, "fallback_algorithm": "popularity" } Variant: C (popularity-based) { "algorithm": "popularity", "endpoint": "/api/recs/trending", "header_text": "Trending now", "max_items": 4, "fallback_algorithm": null } SDK integration: const recsConfig = await intempt.getFlag('recommendation_algorithm', userId); const recs = await fetch(recsConfig.endpoint + '?productId=' + currentProductId); // render recs.items in the PDP recommendation block, using recsConfig.header_text Each rendered recommendation should have target_id = "rec-{position}" (e.g., "rec-1", "rec-2") so Click on per-position aggregation works. Taxonomy notes: - The recommendation endpoints (/api/recs/*) are application-side and not part of the canonical taxonomy; this recipe assumes the merchant has these implemented. - Attributing Placed order revenue to a recommendation requires the order to record which items came from a clicked recommendation: typically via session tracking joining Click on (target_id starts with "rec-") to subsequent Placed order.items. - "Display frequency: always" for server experiments means every getFlag call returns the same variant for the same userId (sticky assignment); the variant is computed once and cached for that user for the experiment's duration.
Install Intempt and run it
- 1
Sign up
Create a free Intempt workspace.
- 2
Connect your data
Add the sources this recipe reads from.
- 3
Ask Blu
Type /product-recommendation-algorithm in Blu once this recipe is live there.
- 4
Review and approve
Blu builds each step and waits for your OK before anything goes live.
Related recipes
Sid ChaudharyDiscovery call qualification fieldsGenerates an AI summary for Discovery calls from meeting data using the qualification framework in the summary recipe.View recipe Sid ChaudharyPost demo nurtureKeeps a demo warm for 90 days with a playbook, a case study, an ROI calculator and a customer story, and pulls the AE in early if they bite.View recipe Sid ChaudharyReturn rate by categoryShows which product categories get returned most, and which ones have got worse since last quarter.View recipe 