In this report, I analyze the recommendation system used on Amazon.com. The goal is to apply scenario design to understand how the system serves both customer needs and Amazon’s organizational objectives, to reverse engineer its likely mechanics, and to propose concrete improvements for its future development.
Describe the different user types: - Everyday shoppers - Browsers - Power users - Gift shoppers - Business users
Explain key goals: - Find suitable products quickly - Discover relevant alternatives and complements - Avoid low quality or irrelevant items - Reorder regular purchases easily - Get inspiration for related items
Summarize how the system supports those goals: - Personalized recommendations based on history - Complementary product suggestions - Reordering support - Simple controls to refine recommendations (in theory and as proposed)
List internal stakeholders: - Retail and category managers - Marketing and campaign teams - Advertising teams - Product and UX teams
Discuss goals: - Increase revenue and conversion - Increase basket size - Improve retention - Launch and promote new products - Balance monetization and trust
Explain: - Scalable recommendation infrastructure - Tunable business rules and ranking - Experimentation support - Analytics and reporting
Discuss: - Explicit data (purchases, ratings, reviews, lists) - Implicit data (views, clicks, cart actions, searches) - Contextual data (time, device, location) - Item metadata (category, price, brand, text, images)
Summarize item-to-item collaborative filtering and how it works at a high level. Explain that Amazon likely combines collaborative filtering with content-based models and personalized ranking models in production.
Describe the different slots and their objectives: - Frequently bought together - Customers who bought this item also bought - Inspired by your browsing history - Recommended for you on the homepage - Email and push notification recommendations
Connect the scenario design to the reverse engineering: - How user goals are addressed - Where tension exists between short-term monetization and long-term trust - Why placing user goals first is valuable for long-term success
Propose: - Not interested / already own this / do not show this brand controls - A page explaining why recommendations appear and letting users configure signals
Propose: - Balancing relevance, profitability, and seller diversity in ranking - Avoiding over-reliance on sponsored products at the top
Propose: - Detecting user mode (reorder, exploration, gift) - Adapting recommendation layout and ranking accordingly
Propose: - Using sequence-aware models to capture session behavior - Improving next-item predictions and recovery from dead ends
Propose: - Lightweight onboarding questions - Combining explicit and early implicit signals
Summarize the key findings: - Amazon’s recommender is deeply integrated into the user experience and business model. - Scenario design clarifies user goals and organizational goals and highlights potential tensions. - Several improvements, especially around transparency, control, multi-objective optimization, and mode awareness, could further enhance both customer experience and business outcomes.