Spotify tests a contextual bandit for its Home content mix

What happened

In September 2025, Spotify-associated researchers published a contextual-bandit method for choosing the music and podcast mix on Home. They described offline evaluations followed by an online A/B test against a historical-preference baseline. The paper and companion company blog report deployment in March 2025, after improvements in engagement measures.

The psychological idea

The model learns from feedback about which content mixtures lead to streaming in a particular context. Exploration sometimes tries an alternative; exploitation uses the currently preferred action. This is a computational learning procedure. Treating its reward signal as equivalent to a person receiving psychological reinforcement would confuse the system’s optimization target with the listener’s experience.

In practice

The system predicted content-type distributions using features including past consumption, temporal signals, and device context. It built recommendation shelves around that distribution. The company blog reports higher impression-to-stream ratios, Home activity, and total consumption in its A/B test. It also says the reward formulation emphasized short-term engagement, with longer-term satisfaction proposed as future work.

Why it mattered

The case makes personalization specific enough to inspect: the algorithm chooses a content mix, receives a defined response, and is compared with an alternative ranking method. It supports discussion of how commercial platforms experiment on recommendation presentation without attributing hidden motives to every feature or claiming that all Spotify recommendations use this particular system.

Evidence behind the claims

• The September 2025 paper reports offline evaluation and an online A/B test of music and podcast calibration on Spotify Home, against a baseline using 90 days of historical preferences, and deployment in March 2025.

  Company self-reportChecked at page levelWhere: Sections 3.1 and 3.2[1]

• The contextual-bandit model uses user consumption, temporal and device features; its action is a content-type distribution and its binary reward indicates whether a user engages with a shelf.

  Company self-reportChecked at page levelWhere: Sections 2.2 and 2.2.1[1]

• The company blog explains epsilon-greedy exploration and reports increases in impression-to-stream ratios, Home activity, and overall minutes streamed in the online A/B test.

  Company self-reportChecked at page levelWhere: Contextual bandit approach; Empirical evaluation, Table 2[2]

• The blog says the current reward emphasizes short-term engagement and proposes rewards incorporating long-term satisfaction as future work.

  Company self-reportChecked at page levelWhere: Lessons learned and looking ahead[2]

• The inspected online evaluation does not state its sample size or duration; its engagement measures do not establish listener wellbeing, addiction, or a psychological purpose for Wrapped.

  InterpretationChecked at page levelWhere: Paper section 3.2 and conclusion; blog empirical evaluation and future work[1][2]

People involved

Course connections

• OpenStax Psychology 2e, section 2.2: Approaches to ResearchMethod. Nearest related section; it does not discuss this topic by name.

  Related research-methods context only; this course passage was not inspected and does not document the Spotify experiment.

• OpenStax Psychology 2e, section 6.3: Operant ConditioningContext. Nearest related section; it does not discuss this topic by name.

  Nearest related learning section only; a machine-learning reward is not confirmation of operant conditioning in listeners.

Course relevance is not evidence for the historical or corporate claims.

Sources and inspection

Inspection statements below preserve the original reading scope. A source marked as not verified supports no claim by itself.

1. [1]
   Feijer, D., Abdollahpouri, H., Gupta, S., Clare, A., Wen, Y., Wasson, T., Dimakopoulou, M., Nazari, Z., Kretschman, K., & Lalmas, M. (2025a). Calibrated recommendations with contextual bandits. arXiv. https://arxiv.org/html/2509.05460v1Feijer, D.; Abdollahpouri, H.; Gupta, S.; Clare, A.; Wen, Y.; Wasson, T.; Dimakopoulou, M.; Nazari, Z.; Kretschman, K.; Lalmas, M.. arXiv. 2025. Preprint. Primary.

   Relevant text checked. Opened arXiv HTML version 1 on October 21, 2026; inspected author affiliations, sections 2.2 through 3.2, Tables 1 and 2, and conclusion. Not just the abstract. Paper metadata identifies September 5, 2025 submission.

2. [2]
   Feijer, D., Abdollahpouri, H., Gupta, S., Clare, A., Wen, Y., Wasson, T., Dimakopoulou, M., Nazari, Z., Kretschman, K., & Lalmas, M. (2025b). Calibrated recommendations with contextual bandits on Spotify homepage. Spotify Research. https://research.atspotify.com/2025/9/calibrated-recommendations-with-contextual-bandits-on-spotify-homepageFeijer, D.; Abdollahpouri, H.; Gupta, S.; Clare, A.; Wen, Y.; Wasson, T.; Dimakopoulou, M.; Nazari, Z.; Kretschman, K.; Lalmas, M.. Spotify Research. 2025. Company document. Company document.

   Company document read. Opened original Spotify Research blog on October 21, 2026; read method, offline and online evaluation, Table 2, March deployment statement, future work and September 18, 2025 byline. No independent audit performed.

Access note: Company-associated experiment reported in a full paper and companion research blog, both opened on October 21, 2026. No Wrapped intention claim. Independent editorial review completed October 21, 2026.

Related cases and psychology

• Netflix describes its recommender system and its business value
• Netflix chooses the artwork each member sees
• TikTok explains how the For You feed recommends videos