BaRT stands for Bandits for Recommendations as Treatments. It is the AI system that governs what appears on a Spotify user's Home screen, including the order of playlists, the songs within those playlists, and the explanatory text that accompanies recommendations.
The name comes from a machine learning technique called "multi-armed bandits," which helps the algorithm decide when to play it safe with familiar music versus when to introduce something new.
How BaRT organizes the Home screen
Spotify's Home screen is structured as rows of playlists called "shelves" (like "Made for You" or "Inspired by your recent listening"), with individual playlists inside those shelves called "cards."
BaRT has two jobs:
- Rank the cards within each shelf - deciding which playlists appear first
- Rank the shelves themselves - deciding which rows of content appear at the top of your screen
This ranking is personalized in real time based on your listening history, the time of day, and how you have responded to previous recommendations.
How Does BaRT Balance Exploration vs. Exploitation?
BaRT constantly balances two modes:
Exploitation mode recommends content the system is confident you will enjoy. It draws on your listening history, saved songs, skipped tracks, and playlist activity to predict what will keep you streaming.
Exploration mode recommends content the system is uncertain about. This serves two purposes: it helps Spotify learn more about your preferences, and it introduces you to music you might not have found otherwise.
The balance between these modes is managed by an "epsilon-greedy" strategy. Most of the time, BaRT exploits what it knows about you. Occasionally, it explores to gather new information.
For new users with little listening history, BaRT leans more heavily on exploration. For longtime users with established preferences, it leans more toward exploitation.
What Is the 30-Second Success Signal?
BaRT measures its own performance using a simple threshold: if a listener streams a recommended track for more than 30 seconds, the recommendation is counted as successful.
The longer someone listens to a recommended playlist or radio session, the more confidence BaRT gains in its predictions for that user. This is why early skips hurt your algorithmic reach - they teach BaRT that the recommendation failed.
What Are the Three Data Sources BaRT Uses?
BaRT does not work alone. It draws on three main data pipelines:
| Data source | What it captures | How it helps |
|---|---|---|
| Collaborative filtering | Patterns in what similar listeners enjoy | "Fans of Artist X also like Artist Y" |
| Audio analysis | Tempo, key, energy, timbre | Finds sonically similar tracks for Radio |
| Natural language processing | Lyrics, playlist titles, blog mentions | Understands mood and genre context |
These signals feed into BaRT, which then decides how to weight them for each individual user.
What this means for artists
BaRT is not a gatekeeper you can pitch. It is a prediction engine that learns from listener behavior.
High saves and low skips teach BaRT that your music satisfies the listeners it was shown to. This increases the likelihood of future recommendations.
High skips and low saves teach BaRT that the recommendation was a mismatch. The system becomes less likely to show your track to similar listeners.
The only way to influence BaRT is to send it positive signals through genuine listener engagement. This means optimizing for save rate, completion rate, and repeat listens rather than raw stream counts.
How Does BaRT Compare to Spotify's Other Systems?
BaRT specifically handles the Home screen and personalized shelf recommendations. Other algorithmic surfaces have their own logic:
- Discover Weekly refreshes every Monday using collaborative filtering
- Release Radar updates every Friday and prioritizes followed artists
- Radio and Autoplay use audio similarity and session-continuation signals
These systems share data, but they operate independently. A track that performs well in BaRT's Home recommendations may also get picked up by Radio, but there is no guaranteed crossover.
What Research Is Behind BaRT?
The foundational research was published by Spotify engineers in 2018 under the title "Explore, Exploit, Explain: Personalizing Explainable Recommendations with Bandits." The paper describes how BaRT learns to predict user satisfaction for any combination of item, explanation, and context.
Since then, Spotify has continued to refine the approach. A 2025 research paper describes using contextual bandits to calibrate content types (music, podcasts, audiobooks) within recommendation lists, adapting to users' evolving preferences in real time.