A reference guide to the technical terms behind Spotify's recommendation engine. Covers BaRT, collaborative filtering, audio features, and engagement metrics.
This glossary defines the technical terms used to describe how Spotify's recommendation system works. Use it as a reference when reading about algorithmic playlists, engagement optimization, or platform strategy.
Core systems
BaRT (Bandits for Recommendations as Treatments)
The AI system that governs Spotify's Home screen recommendations. BaRT uses a multi-armed bandit approach to balance exploitation (showing content it knows you like) with exploration (testing new content to learn more about your preferences). Full explainer →
Collaborative filtering
A recommendation technique that identifies patterns in what similar listeners enjoy. If listeners who saved your track also save Track B, the algorithm is more likely to recommend Track B to new listeners of your music. This powers "Fans Also Like" associations and much of Discover Weekly. Full explainer →
Audio analysis
The process of extracting measurable features from a track's raw audio waveform. Spotify uses convolutional neural networks (CNNs) to analyze spectrograms and detect characteristics like tempo, key, energy, and mood. This enables sonic similarity recommendations for Radio and Autoplay. Full explainer →
Natural language processing (NLP)
The algorithm's ability to understand text context. Spotify's NLP component analyzes playlist titles, song lyrics, blog mentions, and social media discussions to understand cultural context around music.
Engagement metrics
Save rate
The percentage of listeners who save your track to their library. Calculated as saves ÷ unique listeners. A high save rate signals strong listener intent and is one of the most important predictors of algorithmic success. Full explainer →
Skip rate
The percentage of streams where the listener skips before a certain threshold, often the 30-second mark. High early skip rates teach the algorithm that your track is a poor fit for the audience it was shown to.
Completion rate
The percentage of listeners who play a track from start to finish. High completion rates indicate strong retention and contribute positively to algorithmic scoring.
Repeat listen rate
How often individual listeners return to play the same track multiple times. High repeat rates signal deep engagement and can trigger "super listener" classification.
Stream-to-listener ratio
The average number of streams per unique listener over a time period. A ratio above 1.0 indicates repeat listening. Higher ratios suggest dedicated fans rather than one-time casual plays.
Audio features
Tempo
Listener segments
Super listener
A fan in your top 2% of listeners based on streaming frequency and intent. Super listeners drive 18%+ of streams and 50% of ticket purchases through Spotify. Full explainer →
Active listener
A listener who intentionally seeks out your music from active sources (artist profile, release pages, personal playlists, search) within the past 28 days.
Monthly active listener
The subset of monthly listeners who have intentionally streamed from active sources, not just passive playlisting or Radio exposure.
Programmed listener
A listener who has heard your music through algorithmic or editorial playlists but has not actively sought you out.
Platform concepts
30-second rule
A stream only counts toward royalties and engagement metrics if the listener plays the track for 30 seconds or more. Skips before 30 seconds generate no revenue and send negative signals. Full explainer →
1,000-stream threshold
Definition
•
Updated
Spotify Algorithm Glossary
A reference guide to the technical terms behind Spotify's recommendation engine. Covers BaRT, collaborative filtering, audio features, and engagement metrics.
This glossary defines the technical terms used to describe how Spotify's recommendation system works. Use it as a reference when reading about algorithmic playlists, engagement optimization, or platform strategy.
Core systems
BaRT (Bandits for Recommendations as Treatments)
The AI system that governs Spotify's Home screen recommendations. BaRT uses a multi-armed bandit approach to balance exploitation (showing content it knows you like) with exploration (testing new content to learn more about your preferences). Full explainer →
Collaborative filtering
A recommendation technique that identifies patterns in what similar listeners enjoy. If listeners who saved your track also save Track B, the algorithm is more likely to recommend Track B to new listeners of your music. This powers "Fans Also Like" associations and much of Discover Weekly. Full explainer →
Audio analysis
The process of extracting measurable features from a track's raw audio waveform. Spotify uses convolutional neural networks (CNNs) to analyze spectrograms and detect characteristics like tempo, key, energy, and mood. This enables sonic similarity recommendations for Radio and Autoplay. Full explainer →
Natural language processing (NLP)
The algorithm's ability to understand text context. Spotify's NLP component analyzes playlist titles, song lyrics, blog mentions, and social media discussions to understand cultural context around music.
Engagement metrics
Save rate
The percentage of listeners who save your track to their library. Calculated as saves ÷ unique listeners. A high save rate signals strong listener intent and is one of the most important predictors of algorithmic success. Full explainer →
Skip rate
The percentage of streams where the listener skips before a certain threshold, often the 30-second mark. High early skip rates teach the algorithm that your track is a poor fit for the audience it was shown to.
Completion rate
The percentage of listeners who play a track from start to finish. High completion rates indicate strong retention and contribute positively to algorithmic scoring.
Repeat listen rate
How often individual listeners return to play the same track multiple times. High repeat rates signal deep engagement and can trigger "super listener" classification.
Stream-to-listener ratio
The average number of streams per unique listener over a time period. A ratio above 1.0 indicates repeat listening. Higher ratios suggest dedicated fans rather than one-time casual plays.
Audio features
Tempo
Listener segments
Super listener
A fan in your top 2% of listeners based on streaming frequency and intent. Super listeners drive 18%+ of streams and 50% of ticket purchases through Spotify. Full explainer →
Active listener
A listener who intentionally seeks out your music from active sources (artist profile, release pages, personal playlists, search) within the past 28 days.
Monthly active listener
The subset of monthly listeners who have intentionally streamed from active sources, not just passive playlisting or Radio exposure.
Programmed listener
A listener who has heard your music through algorithmic or editorial playlists but has not actively sought you out.
Platform concepts
30-second rule
A stream only counts toward royalties and engagement metrics if the listener plays the track for 30 seconds or more. Skips before 30 seconds generate no revenue and send negative signals. Full explainer →
1,000-stream threshold
The speed of a track measured in beats per minute (BPM). Ranges from 0-250 BPM.
Energy
A 0.0-1.0 measure of intensity and activity. Combines dynamic range, perceived loudness, timbre, onset rate, and entropy. Death metal scores high; ambient music scores low.
Valence
A 0.0-1.0 measure of musical positiveness. High valence (0.8+) sounds happy or euphoric. Low valence (0.2 or below) sounds sad, melancholic, or angry.
Danceability
A 0.0-1.0 score of how suitable a track is for dancing based on tempo, rhythm stability, beat strength, and regularity.
Acousticness
A 0.0-1.0 confidence measure of whether a track is acoustic. A value of 1.0 indicates high confidence the track contains no electronic or amplified instruments.
Instrumentalness
A 0.0-1.0 prediction of whether a track contains vocals. Values above 0.5 suggest instrumental tracks. "Ooh" and "aah" sounds are treated as instrumental.
Speechiness
A 0.0-1.0 measure of spoken word presence. Podcast-style content scores high; pure instrumental music scores low.
Algorithmic surfaces
Discover Weekly
A personalized playlist of 30 tracks the listener has not heard before, refreshed every Monday. Primarily powered by collaborative filtering. Full explainer →
Release Radar
A personalized playlist of new releases from followed artists and similar acts, refreshed every Friday. Requires pitching through Spotify for Artists at least 7 days before release.
Radio
An auto-generated queue of tracks similar to a seed song or artist. Uses audio analysis for sonic similarity and collaborative filtering for audience overlap.
Autoplay
The feature that continues playing music when a playlist or album ends. Powered by the same signals as Radio, extending the listening session indefinitely.
Daylist
A dynamic playlist that updates six times per day based on time-of-day listening patterns. Uses microgenre labels like "indie sad girl rainy evening." Full explainer →
Daily Mix
A set of 4-6 playlists grouping your saved music and similar tracks by genre or mood. Updates daily.
AI DJ
A personalized radio experience with synthesized voice commentary. Uses the same personalization engine as other surfaces but adds spoken context about artists and tracks. Full explainer →
Promotional tools
Discovery Mode
A royalty-sharing arrangement where artists accept a 30% commission cut on streams from Radio, Autoplay, and Mixes in exchange for increased recommendation likelihood. Full explainer →
Marquee
A full-screen pop-up ad for new releases (within 21 days), shown when targeted listeners open the mobile app. Part of Campaign Kit; pay-per-click pricing.
Showcase
A banner ad on Spotify's Home feed for any release. Part of Campaign Kit; pay-per-click pricing. Full explainer →
Canvas
A looping 3-8 second video that plays behind your track on mobile. Not directly algorithmic, but can improve engagement metrics by encouraging listeners to watch and save.
Technical concepts
Exploration vs exploitation
The trade-off BaRT manages between showing you content it knows you like (exploitation) and testing new content to learn your preferences (exploration). New users get more exploration; established users get more exploitation.
Cold start problem
The challenge of recommending music for new artists or new users who have no listening history. Audio analysis helps solve this by enabling sonic similarity recommendations without behavioral data.
Epsilon-greedy strategy
The specific algorithm BaRT uses to balance exploration and exploitation. Most of the time it exploits (shows high-confidence recommendations); occasionally it explores (shows uncertain recommendations to gather data).
Contextual bandit
A machine learning framework that learns optimal actions based on context. BaRT is a contextual bandit system that considers user context (time of day, device, recent activity) when making recommendations.
Spectrogram
A visual representation of sound frequencies over time. Spotify's CNNs analyze spectrograms to extract audio features from raw waveforms.
Microgenre
A granular sub-genre classification used for personalization. Spotify maintains a taxonomy of 6,000+ microgenres based on listening patterns and audio characteristics.
As of 2024, tracks must accumulate 1,000 streams within a rolling 12-month period to generate royalties. Streams below this threshold do not pay out.
Popularity index
A 0-100 score Spotify assigns to each track and artist based on recent streaming velocity. Higher scores indicate faster growth relative to the platform average. Full explainer →
Taste profile
The algorithmic model Spotify builds for each user based on their listening history, saves, skips, and playlist activity. Powers all personalized recommendations.
The speed of a track measured in beats per minute (BPM). Ranges from 0-250 BPM.
Energy
A 0.0-1.0 measure of intensity and activity. Combines dynamic range, perceived loudness, timbre, onset rate, and entropy. Death metal scores high; ambient music scores low.
Valence
A 0.0-1.0 measure of musical positiveness. High valence (0.8+) sounds happy or euphoric. Low valence (0.2 or below) sounds sad, melancholic, or angry.
Danceability
A 0.0-1.0 score of how suitable a track is for dancing based on tempo, rhythm stability, beat strength, and regularity.
Acousticness
A 0.0-1.0 confidence measure of whether a track is acoustic. A value of 1.0 indicates high confidence the track contains no electronic or amplified instruments.
Instrumentalness
A 0.0-1.0 prediction of whether a track contains vocals. Values above 0.5 suggest instrumental tracks. "Ooh" and "aah" sounds are treated as instrumental.
Speechiness
A 0.0-1.0 measure of spoken word presence. Podcast-style content scores high; pure instrumental music scores low.
Algorithmic surfaces
Discover Weekly
A personalized playlist of 30 tracks the listener has not heard before, refreshed every Monday. Primarily powered by collaborative filtering. Full explainer →
Release Radar
A personalized playlist of new releases from followed artists and similar acts, refreshed every Friday. Requires pitching through Spotify for Artists at least 7 days before release.
Radio
An auto-generated queue of tracks similar to a seed song or artist. Uses audio analysis for sonic similarity and collaborative filtering for audience overlap.
Autoplay
The feature that continues playing music when a playlist or album ends. Powered by the same signals as Radio, extending the listening session indefinitely.
Daylist
A dynamic playlist that updates six times per day based on time-of-day listening patterns. Uses microgenre labels like "indie sad girl rainy evening." Full explainer →
Daily Mix
A set of 4-6 playlists grouping your saved music and similar tracks by genre or mood. Updates daily.
AI DJ
A personalized radio experience with synthesized voice commentary. Uses the same personalization engine as other surfaces but adds spoken context about artists and tracks. Full explainer →
Promotional tools
Discovery Mode
A royalty-sharing arrangement where artists accept a 30% commission cut on streams from Radio, Autoplay, and Mixes in exchange for increased recommendation likelihood. Full explainer →
Marquee
A full-screen pop-up ad for new releases (within 21 days), shown when targeted listeners open the mobile app. Part of Campaign Kit; pay-per-click pricing.
Showcase
A banner ad on Spotify's Home feed for any release. Part of Campaign Kit; pay-per-click pricing. Full explainer →
Canvas
A looping 3-8 second video that plays behind your track on mobile. Not directly algorithmic, but can improve engagement metrics by encouraging listeners to watch and save.
Technical concepts
Exploration vs exploitation
The trade-off BaRT manages between showing you content it knows you like (exploitation) and testing new content to learn your preferences (exploration). New users get more exploration; established users get more exploitation.
Cold start problem
The challenge of recommending music for new artists or new users who have no listening history. Audio analysis helps solve this by enabling sonic similarity recommendations without behavioral data.
Epsilon-greedy strategy
The specific algorithm BaRT uses to balance exploration and exploitation. Most of the time it exploits (shows high-confidence recommendations); occasionally it explores (shows uncertain recommendations to gather data).
Contextual bandit
A machine learning framework that learns optimal actions based on context. BaRT is a contextual bandit system that considers user context (time of day, device, recent activity) when making recommendations.
Spectrogram
A visual representation of sound frequencies over time. Spotify's CNNs analyze spectrograms to extract audio features from raw waveforms.
Microgenre
A granular sub-genre classification used for personalization. Spotify maintains a taxonomy of 6,000+ microgenres based on listening patterns and audio characteristics.
As of 2024, tracks must accumulate 1,000 streams within a rolling 12-month period to generate royalties. Streams below this threshold do not pay out.
Popularity index
A 0-100 score Spotify assigns to each track and artist based on recent streaming velocity. Higher scores indicate faster growth relative to the platform average. Full explainer →
Taste profile
The algorithmic model Spotify builds for each user based on their listening history, saves, skips, and playlist activity. Powers all personalized recommendations.
