Spotify deploys machine learning to automate user acquisition campaigns

Spotify’s engineering workforce published particulars on November 7, 2023, describing an automatic content material advertising and marketing system that generates hundreds of advert variations day by day for person acquisition campaigns. The streaming service developed a machine studying mannequin utilizing XGBoost to pick out which advert creatives to deploy throughout Fb, Google UAC, TikTok, and different digital platforms.

The technical weblog put up from Spotify Engineering outlined how the corporate mixed automated artistic technology with predictive algorithms to handle campaigns throughout tens of hundreds of advertisements globally. Bryan Maloney, Christopher Tang, Deepak Bhat, and Ryan Kim authored the detailed account of constructing what they described as an end-to-end automated system for efficiency advertising and marketing.

Excessive-cardinality artistic variants created deployment challenges

Spotify’s advertising and marketing workforce developed a artistic manufacturing pipeline that would generate and deploy advert creatives primarily based on listening habits in geographic areas. The system produced creatives from a high-cardinality dataset, the place the variety of variations being uploaded to advertising and marketing channels overwhelmed these channels’ means to optimize advertisements successfully.

The issue stemmed from Spotify’s huge content material catalog. With thousands and thousands of artists and quite a few template designs, the mix prospects reached a whole lot of hundreds of potential advert creatives. Digital advert platforms like Fb are designed to optimize throughout 4 to eight advert variations, not orders of magnitude extra choices.

Previous to 2019, Spotify performed guide checks of content material advertising and marketing that confirmed various success ranges. A guide take a look at within the first half of 2019 demonstrated potential for bringing incremental customers by content material advertisements. The engineering problem grew to become constructing a system that mechanically generated content-based advertisements, loaded them to digital advertising and marketing channels, and noticed efficiency to make changes repeatedly.

XGBoost mannequin outperformed heuristic method throughout metrics

The engineering workforce conceived a five-stage loop for system habits: ingest, rank, deploy, study, and repeat. The technical method required automating every step sequentially, beginning with content material technology.

Spotify initially used primary templating with a Java-based backend service that retrieved content material parts from metadata providers. This service generated static photographs however restricted the workforce to simplistic templates with out animation capabilities.

The workforce evaluated a number of options together with Lottie and Blender earlier than deciding on Adobe After Results for movement graphics. After Results supplied artistic freedom for designers whereas supporting template technology throughout dozens of facet ratios and sizes. The open supply challenge nexrender prolonged aerender right into a batchable system, enabling the workforce to script file actions and handle headless render nodes effectively.

For content material rating, Spotify leveraged machine studying mixed with knowledge sources to rank content material day by day. The system collected knowledge factors on advert efficiency metrics together with clicks, impressions, app installs, registrations, and subscriptions from completely different advert platform APIs and cellular measurement companion APIs.

The primary heuristic mannequin used three calculations: recognition, share of registrations, and variety. The mannequin used recognition to construct a set of eight artists for campaigns, noticed efficiency utilizing share-of-registrations metrics, and evaluated optimization primarily based on each recognition and differentiation from different artists and advertisements.

The machine studying method remodeled these fastened elements right into a supervised studying drawback. Options of every artist predicted share of registrations, with recognition added as one studying function. The mannequin included marketing campaign metadata like market, advert artistic dimension, working system, and template theme.

Spotify used the XGBoost library by way of the corporate’s Kubeflow managed service to implement the algorithm. The mannequin educated day by day with historic knowledge over a lookback window and predicted two foremost goal variables for free-tier advertisements: reg_percentage (proportion of registrations the ranked artist would contribute) and relative_cpr_ratio (ratio of the ranked artist in total CPR).

For premium-tier advertisements, the mannequin predicted sub_percentage (proportion of premium subscriptions) and relative_cps_ratio (share in total price per subscription).

The workforce selected relative metrics as an alternative of absolute metrics as a result of uncooked metrics rely on exterior market elements which are tough to mannequin.

A/B testing validated efficiency enhancements

An A/B take a look at ran for 3 weeks in two areas with the heuristic mannequin as management and the ML mannequin as therapy. Outcomes confirmed the ML mannequin reaching 4% and 14% cheaper price per registration than the heuristic mannequin within the two areas.

The advertisements generated from ML mannequin rankings had 11% to 12% increased click-through charges than the heuristic mannequin. This enchancment got here from the ML mannequin being educated with richer knowledge containing the next variety of options.

Following take a look at outcomes, the pure alternative was productionizing the ML mannequin for content material rating throughout all energetic areas operating advertising and marketing campaigns.

The top-to-end structure included metadata providers, mosaic rendering, color-picker instruments, an orchestrator, oseary-drakoulias coordination system, BigQuery for content material rankings per nation, Belafonte processing, BigQuery knowledge cleanup by way of Klaus, and integration with digital advert platforms. The system operated by scheduled runs utilizing Styx and Dataflow.

iOS privateness modifications didn’t influence mannequin efficiency

A big technical consideration emerged when Apple unveiled Identifier for Advertisers (IDFA) restrictions in summer time 2021 for iOS model 14.5 onward. The change remodeled the advert tech business panorama concerning which knowledge factors advertisers might gather and use.

Spotify would now not depend on getting user-level or log-level advert efficiency knowledge to optimize campaigns. Nonetheless, as a result of the ML mannequin educated on aggregated knowledge over a lookback window, the change didn’t have an effect on it adversely.

The second Spotify grew to become conscious that IDFA would activate for iOS model 14.5 onward, the workforce evaluated the ML mannequin output by offline analyses for any detrimental influence on mannequin efficiency. The evaluation confirmed the mannequin efficiency wouldn’t be impacted negatively.

Eric Seufert, cellular advertising and marketing analyst, mentioned the Spotify implementation in an October 4, 2025, put up on X. “Spotify’s ML pre-ranking mannequin outperformed a easy heuristic mannequin, with 4%-14% decrease CPRs and 11%-12% increased CTRs,” Seufert wrote. “The ML mannequin utilized a wealthy set of options to foretell sub_percentage (the proportion of contributed subscriptions from the artist) and relative_cps_ratio (the share of the artist’s price per subscription within the advertising and marketing marketing campaign) for premium subscriptions, whereas the heuristic mannequin used three fastened options.”

Seufert famous that though this was deployed earlier than Apple’s App Monitoring Transparency framework, the workforce discovered that ATT did not influence efficiency, as coaching relied on aggregated knowledge. “This clearly stays a related challenge as advertisers scale the quantity of their artistic manufacturing by generative instruments,” Seufert noticed.

Technical challenges included exterior dependencies and migration

All through the challenge, the workforce encountered a number of technical challenges. Transferring from Java templates to After Results turned asset technology from one thing that could possibly be finished inline in an API name to one thing that wanted rendering asynchronously. Scaling render staff up and down in response to asset quantity proved difficult.

Dependency on advert platform APIs for ingesting advert efficiency metrics created vulnerability. To feed the ML content material rating mannequin with high quality coaching knowledge, knowledge pipelines needed to fetch advert efficiency metrics from platform APIs day by day. Fb API outages in a number of situations brought on disruptions in knowledge pipelines, which resulted within the ML mannequin not with the ability to practice and produce content material rankings till Fb’s advertising and marketing API returned to regular.

As a workaround, the workforce determined to proceed preserving artist rankings from yesterday when the ML mannequin lacked perception into the most recent day’s advert efficiency.

Spotify additionally migrated from Modify to Department for its most well-liked deep linking and attribution companion. This required updating all knowledge pipelines and the ML mannequin to eat Department-powered advert metrics, then calibrating the system for content material rating. The workforce verified that consuming advert efficiency attribution knowledge from Department as an alternative of Modify didn’t lead to efficiency implications for the ML rating mannequin.

Automation addresses rising artistic manufacturing scale

The implementation demonstrates how firms deal with rising artistic manufacturing volumes. Spotify has expanded its advertising infrastructure all through 2025, together with the April 3 announcement of Spotify Advert Alternate for programmatic shopping for and generative AI integration for audio advert creation.

The streaming service integrated with Smartly on September 3, 2025, enabling advertisers to handle audio, show, and video campaigns by AI-powered platforms. That partnership mixed Spotify’s listener base with automated artistic manufacturing and cross-channel measurement capabilities.

IAB Europe’s September 18, 2025, report discovered 85% of firms already deploy AI-based instruments for advertising and marketing functions, with concentrating on and content material technology main adoption patterns at 64% and 61% respectively.

The technical structure Spotify constructed displays broader business motion towards automated marketing campaign administration. Amazon DSP added Spotify’s global audio and video inventory by a programmatic partnership introduced October 1, 2025, combining Amazon’s purchasing indicators with Spotify’s 696 million month-to-month customers throughout 9 markets.

Artistic automation methods like these Spotify developed deal with manufacturing bottlenecks by enabling speedy technology of a number of artistic variants tailor-made to completely different audiences, codecs, and platforms. Neatly’s purchasers produce over 150 million artistic variants month-to-month by automated instruments, demonstrating scalability potential.

For advertising and marketing professionals, this improvement carries a number of implications. The pre-ranking algorithm method permits firms to generate artistic variations at scale with out overwhelming platform optimization algorithms. The usage of aggregated moderately than user-level knowledge for coaching supplied resilience towards privateness framework modifications.

The Spotify workforce emphasised that solely a handful of tech firms have absolutely automated the efficiency advertising and marketing cycle globally. The product required cross-functional efforts between engineering and advertising and marketing to resolve issues on the intersection of content material catalog knowledge, artistic manufacturing, and efficiency optimization.

Machine studying methods for promoting proceed evolving as platforms stability automation with advertiser management. Google introduced text guidelines for AI-powered campaigns on September 10, 2025, enabling time period exclusions and pure language restrictions in Efficiency Max and AI Max campaigns.

The technical particulars revealed by Spotify present perception into implementation challenges for automated advertising and marketing methods. The stability between producing artistic variations and sustaining platform optimization capabilities represents a constraint that pre-ranking algorithms can deal with by predictive modeling and feature-rich coaching knowledge.

Spotify’s method of constructing a supervised studying mannequin that predicts relative metrics moderately than absolute values demonstrates one technique for dealing with exterior market elements that have an effect on marketing campaign efficiency. The day by day retraining cycle primarily based on outlined lookback home windows permits the system to adapt to altering circumstances whereas sustaining prediction high quality.

Timeline

Abstract

Who: Spotify’s engineering workforce, together with Bryan Maloney (Senior Engineering Supervisor), Christopher Tang (Senior ML Engineer), Deepak Bhat (Senior Knowledge Engineer), and Ryan Kim (Senior Product Supervisor), developed the automated content material advertising and marketing system.

What: An automatic content material advertising and marketing system combining artistic technology, machine learning-based content material rating utilizing XGBoost, and deployment to digital promoting platforms. The ML mannequin achieved 4%-14% decrease price per registration and 11%-12% increased click-through charges in comparison with heuristic approaches whereas managing tens of hundreds of advert variations globally.

When: Spotify started creating the system after profitable guide checks in H1 2019, deployed the ML mannequin earlier than 2021, and revealed technical particulars on November 7, 2023. The system stays operational as of October 2025.

The place: The system operates throughout Spotify’s international person acquisition campaigns on platforms together with Fb, Google UAC, TikTok, and different digital promoting channels. Implementation spans a number of markets the place Spotify runs advertising and marketing campaigns.

Why: The system addresses the problem of optimizing person acquisition at scale when artistic manufacturing generates variations that overwhelm promoting platforms’ native optimization capabilities. By implementing pre-ranking algorithms, Spotify can leverage its intensive content material catalog whereas sustaining promoting platform efficiency, in the end enhancing return on advert spend for advertising and marketing budgets.