DoorDash Rebuilds Homepage with LLMs
DoorDash replaced 300 static carousels on its homepage with dynamically generated, AI-curated ones powered by Large Language Models. The engineering team used embeddings, ranking, and continuous A/B testing to boost relevance. The company noted that deep iteration was required to avoid "superficial" personalization that looked smart but didn't improve metrics.
- The new system led to a double-digit improvement in click-through rates and higher conversion rates in A/B tests conducted in San Francisco and Manhattan. - To manage costs and avoid latency, the LLM-powered carousel themes are pre-computed in batches rather than being generated in real-time when a user opens the app. - DoorDash's engineering team improved the precision of homepage recommendations from 68% to 85%, meaning 8 to 9 out of 10 stores in a given carousel now closely match the generated theme. - A content moderation system using a "jury" of three distinct LLMs reviews every generated carousel title to ensure safety and appropriateness, achieving a 95% recall rate in catching problematic content. - The new approach moves beyond simple categorization like "Vietnamese food" to generate nuanced themes like "Late-Night Noodle Cravings" by analyzing a user's past orders, browsing history, and dish-level preferences. - Other consumer tech companies are leveraging LLMs for personalization; Spotify, for instance, uses them to add contextual narratives to recommendations, boosting user engagement, while Netflix is building a unified foundation model to better understand long-term taste. - The system also enhances merchant discovery by exposing customers to a wider variety of cuisines and new merchants, which in turn drives more business to small and mid-sized restaurants. - Beyond the homepage, DoorDash is also using AI to generate descriptive and enticing menu item descriptions for restaurants, which undergo A/B testing to measure their impact on customer engagement and conversion.
