Personalized Shopping Suggestions on Amazon: Product Design, Pipeline and Metrics

Amazon wants to suggest products its shoppers are likely to want. Lay out how you would build that recommendation system, then choose the metrics that tell you whether it is succeeding. ### Clarifying Questions - Which surface comes first: the home page, the product detail page, the cart, or follow-up emails? - Is the main goal more sales per visit, helping shoppers discover new categories, or long-term loyalty? - How deep should the design go: the shopper experience only, or also the data and model pipeline behind it? - Should sponsored product placements be part of this system, or kept separate? ### Part 1 — Design the recommendation system Describe who the recommendations serve, where they appear, and how the system decides which products to show. ```hint Intent changes by page A shopper comparing two blenders on a product page wants something different from one opening the app with no goal. Decide which moments you serve before choosing signals or models. ``` #### What This Part Should Cover - The users, surfaces and shopper intent the design serves, and which comes first - A pipeline a PM can explain: signals, candidate sources, ranking, filters and the feedback loop - Cold start for new shoppers and new products, and recommendations that should never appear ### Part 2 — Measure success Choose the key metrics you would use to decide whether the recommendations are working. ```hint Clicks can mislead Think about how click-through could rise while shoppers buy no more than before, and what measurement would catch that. ``` #### What This Part Should Cover - A primary metric that captures incremental value, not just engagement with the recommendation rows - Supporting funnel and quality metrics, and guardrails that catch harm - How the metrics are measured: controlled experiments, holdouts and attribution ### What a Strong Answer Covers - A clear choice of surface and shopper intent that both the design and the metrics follow - Trade-offs between relevance and discovery, and between short-term sales and long-term trust - Edge cases such as gifts, shared accounts, products already bought and sensitive categories - A launch and iteration plan that uses offline model metrics only as a filter before online tests ### Follow-up Questions - Click-through on recommendations rose 10% in a test, but revenue per visitor did not move. What happened, and what do you do next? - A shopper bought a baby gift once and now sees baby products everywhere. How should the system handle it? - How would you prove to leadership that recommendations create sales rather than take credit for purchases that would have happened anyway?