Idea
Generative recommender system increasing podcast discovery and engagement through semantic and contextual user modeling.
Research Paper
Core Innovation
This paper introduces GLIDE, a generative recommender that uses Semantic IDs to ground recommendations in a large podcast catalog. It integrates recent listening history, lightweight user context, and long-term user embeddings as soft prompts, enabling intent-aware, scalable podcast discovery with low latency suitable for production environments.
Why It Matters
Podcast listeners often rely on familiar shows but also seek new content as their interests evolve. GLIDE addresses this by balancing stable preferences with intent-aware exploration, improving user engagement and content discovery at scale. This enhances user satisfaction and retention while efficiently handling large catalogs under production constraints.
Market Size (TAM)
$10–20B TAM for digital audio streaming; $2–5B SAM from podcast platforms and advertisers. Driven by rising podcast consumption and demand for personalized discovery.
Potential Customers & Pain Points
- Streaming platforms – Need to improve content discovery and user engagement
- Podcast creators – Need better exposure to new audiences
- Advertisers – Need targeted reach to evolving listener interests
Business Model
Subscription and ad-supported streaming platforms licensing the generative recommendation technology to enhance user engagement and monetization.
Competitive Landscape
- Spotify Podcast Recommender
- Apple Podcasts
- Google Podcasts
- Stitcher
- Audible
Implementation Challenges
- Scaling generative models with low latency in production
- Accurately capturing evolving user intent and preferences
- Integrating semantic grounding with large
- dynamic catalogs
Validation Strategy
- Offline evaluation using retrieval metrics and human judgments
- LLM-based evaluation for semantic relevance
- Large-scale online A/B testing with millions of users measuring engagement and discovery metrics
Research Paper Overview
Deploying Semantic ID-based Generative Retrieval for Large-Scale Podcast Discovery at Spotify
Summary
GLIDE is a production-scale generative recommender system at Spotify that improves podcast discovery by combining semantic reasoning, user context, and long-term preferences. It increases non-habitual podcast streaming by up to 5.4% and new-show discovery by up to 14.3%, while meeting production cost and latency requirements.