The Value of Silence: The Effect of UMG’s Licensing Dispute with TikTok on Music Demand

1. Introduction

On February 1, 2024, TikTok users like @mariona.roma woke up to a grim reality. Effective that date, Universal Music Group (UMG), the label representing numerous artists—like Taylor Swift, Adele, and Drake—had pulled the soundtracks of all of its artists from TikTok’s music library. Consequently, all videos that had used these tracks, including several by @mariona.roma, went silent. The silence disrupted the platform’s dynamics; dancers moved without the usual beat drops, and creators lip-synced to muted sound, highlighting the indispensable role of music in social media content creation and the potential implications of such disputes on the creative community. The situation sparked widespread disappointment and consternation among creators and followers alike (Kuo 2024).

UMG’s dramatic action followed months of efforts to negotiate a new licensing agreement with TikTok. At the heart of the dispute between the parties was UMG’s claim that the previous agreement was, in its own words, “unfair” (Universal Music Group 2024b) because it failed to adequately compensate the label as well as its artists and songwriters for the use and consumption of tracks on the social media platform. In particular, although TikTok would pay the label a certain amount each time a creator incorporated a UMG track into their video, there was no further compensation. Thus, although the videos that included these music tracks were often viewed thousands and sometimes millions of times by TikTok users, UMG and its artists/songwriters did not benefit from this consumption. This was particularly concerning given that more than 85% of TikTok videos contain music—a significantly higher proportion than on other platforms (Crawford 2024). Moreover, although TikTok videos are typically short and tend to include only part of a track, they often feature the most memorable or catchy parts, such as the chorus or hook (Deliver My Tune 2024). Repeated exposure to these music segments may reduce listeners’ desire to hear the full song again on other streaming platforms, such as Spotify and YouTube. If this is the case, then UMG and its talent may be negatively affected by this cross-effect on demand. Notably, music streaming platforms usually compensate labels and artists every time a track is streamed (i.e., listened to). Given that streaming represents more than 84% of music labels’ revenues (RIAA 2024), any potential cannibalization can thus have significant financial implications for these firms.

For its part, TikTok believed that the compensation it had been paying was, in fact, “fair.” The core of its argument was that being featured on TikTok was a boon for music labels, artists, and songwriters because such exposure could help promote the tracks and hence, assist the artists in getting discovered (TikTok News 2024a), which would in turn spur greater demand on other channels. In a sense, TikTok intimated the potential for a positive cross-effect on music demand on other platforms. In its 2025 Music Impact Report, TikTok indeed highlighted a broader link to the music streaming industry. According to the report, U.S. music listeners who use TikTok are 68% more likely to subscribe to paid music streaming services than the general U.S. population. Moreover, TikTok users in the United States spend 46% more money on music each month compared with the average U.S. music listener. Together, these figures suggest that TikTok particularly appealed to music lovers and possibly contributed to music consumption beyond its own platform (TikTok News 2025).

In this paper, we examine empirically the conflicting arguments made by each party and, more broadly, address the issue of how major content owners like UMG—which holds a large number of music copyrights and trademarks—should consider the impact of social media consumption on the demand for their content on paid streaming platforms.

Identifying the causal impact of social media consumption on demand on other outlets is typically challenging because of potential endogeneity issues often associated with online user behavior (Godes and Mayzlin 2004). To overcome such endogeneity concerns, we leverage the dispute between UMG and TikTok as a unique natural quasi-experiment. Specifically, given that the other two major labels that constitute the so-called “Big Three” (Rys 2024), Sony Music Entertainment (SME) and Warner Music Group (WMG), did not remove their music from TikTok during this time frame, we can use their tracks as a control group to causally examine how withdrawing UMG tracks from TikTok affects the demand for UMG music on streaming platforms, such as Spotify and YouTube. This context allows us to conduct a Difference-in-Differences (DiD) analysis to evaluate the effect of the treatment (i.e., removal of UMG music from TikTok) on our outcome of interest (i.e., demand for UMG music on streaming platforms). If the removal of UMG music from TikTok results in a relatively positive impact on Spotify streams and YouTube views, compared with continued availability, then one can infer a substitution (or cannibalistic) effect of TikTok on streaming demand, thus supporting UMG’s concerns. Conversely, if there is a relatively negative impact, then it implies a complementarity effect, which aligns with TikTok’s reasoning regarding the promotion and discovery role of its social media platform.

We compiled a data set by drawing on multiple sources, including the websites of the Big Three record labels, as well as two music information aggregators, Soundcharts and Chartmetric (Chartmetric 2024, Soundcharts 2024). We start with a comprehensive list of the artists affiliated with the Big Three labels, which is based on information from the labels’ official websites and Wikipedia. We then use Soundcharts and Chartmetric to obtain a listing of all the music tracks for each artist, which gives us a total of 235,741 tracks (across the three labels). Next, for each track, we obtain a set of relevant track-specific and artist-specific information, for example, the career stage of the artists, track release date, and availability on TikTok in the pre-dispute period. Finally, for each track, we also collect data on its daily streaming demand—Spotify streams and YouTube views—over a six-month period from October 10, 2023, to April 7, 2024.

In our main study, we conduct a log Difference-in-Differences analysis, leveraging the removal of UMG music from TikTok on February 1, 2024, as a quasi-natural experiment. We find that, on average, the silencing of UMG’s music on TikTok did not have a significant impact on the overall demand for UMG tracks on Spotify and YouTube compared with the counterfactual scenario where UMG tracks are not removed from TikTok. However, this null effect masks considerable heterogeneity. In particular, we focus on one dimension that is likely to impact the estimated treatment effects—the availability versus the nonavailability of tracks on TikTok prior to the dispute. The former group represents music that TikTok creators could incorporate into their posted videos and hence, that users could be exposed to repeatedly on the social media platform. The latter group represents music tracks that could not be incorporated into posted content on TikTok but that could still be impacted indirectly by tracks that were leveraged on the platform. For example, if exposure to an artist’s available tracks on TikTok resulted in users seeking alternative outlets to listen to the same artist’s other music that was not made available on the platform. Moreover, descriptively, these two groups are systematically different; tracks on TikTok (pre-dispute) tend to be more popular and performed by more successful artists, whereas tracks not on TikTok tend to be less popular and less likely to be performed by big-name artists. For example, in our data, the median number of daily Spotify streams is 2,353 for tracks that were available on TikTok compared with just 82 for those that were not available on TikTok.

For tracks that were available on TikTok prior to the dispute, removal from TikTok led to a 2%–3% increase in consumption on Spotify and YouTube. This suggests a substitution effect for tracks that had been available for creators to incorporate into their posted content on TikTok, and indicates that TikTok may cannibalize the consumption of songs that would otherwise occur on revenue-sharing platforms like Spotify and YouTube. This supports UMG’s concerns that TikTok had not been adequately compensating its artists and songwriters. In contrast, tracks that were not available on TikTok prior to the dispute experienced a roughly 2% decrease in their consumption on Spotify and YouTube as a result of UMG’s action. This points to a complementarity effect, indicating that UMG tracks not previously available on TikTok could potentially benefit from the label placing other music on the platform. Further analysis suggests that this complementarity effect is likely due to the promotion and discovery role that TikTok could be playing, because it is driven mainly by artists with a partial presence on TikTok; that is, some of their songs are available for use by TikTok creators, whereas others are not. In such cases, social media users may first discover the artist via one (or more) of their songs embedded in TikTok videos and then seek out additional music by the same artist that is not available on TikTok yet is available on streaming platforms like Spotify. This explanation would be consistent with the negative effect that these tracks (i.e., by artists with partial TikTok availability) exhibit in the posttreatment period. Taken together, our findings support the arguments of both UMG and TikTok, albeit for different groups.

We present an extensive set of robustness checks (in Section 5.3) for the results obtained in our main analysis. These include examining the validity of the parallel trends assumptions across all estimated models, considering alternative specifications, assessing sensitivity to outliers, varying the pre- and posttreatment time windows, and controlling for time-varying factors. We largely obtain consistent support for our empirical approach and findings. In particular, when we split the tracks by availability on TikTok prior to the dispute, the heterogeneous results are always consistent across all model specifications and samples and form the basis of our economic impact calculations and policy conclusions discussed below.

Lastly, we quantify the economic implications of our findings by conducting a simple back-of-the-envelope calculation. We incorporate the results indicating that (1) for tracks available on TikTok, there is a substitution effect on Spotify, which implies a revenue gain if UMG’s music is removed from TikTok, and (2) for tracks not available on TikTok, there is a complementarity effect, which implies a revenue loss if UMG’s music is removed from TikTok. Assuming that TikTok maintains its original compensation structure to artists, we calculate the expected annual revenue gain from the former group and the expected annual revenue loss from the latter group and sum them together to obtain an estimate of the overall revenue impact on UMG. We find that if UMG’s music is removed from TikTok, then the potential revenue gains from the tracks already on TikTok would easily outweigh the losses from the tracks not on TikTok. Specifically, we estimate that such a silencing move would result in an expected annual net revenue gain of approximately $317 million from Spotify streaming alone. This estimation is much higher than the roughly $111 million that TikTok was paying UMG prior to the dispute (Universal Music Group 2024b). In the “aftermath” of the dispute, and consistent with our findings, on May 1, 2024, UMG and TikTok announced a new licensing agreement that promises to “improve remuneration for UMG’s songwriters and artists” (Universal Music Group 2024c).¹

Our paper makes several contributions to the literature on cross-platform digital content consumption and the economic implications of social media platforms for the monetization of copyrighted content. First, substantively, we show that there are both substitution and complementarity effects in cross-platform consumption of digital content and that social media firms like TikTok can both help and hurt demand in other channels. In particular, we find that streaming demand for content available on the social media platform can directly suffer (i.e., a substitution effect), whereas streaming demand for content not available on the social media platform can indirectly benefit (i.e., a complementarity effect) provided that some of the artists’ tracks receive exposure on the platform. Second, from a managerial and economic perspective, we show that these two opposing cross-effects imply that content owners need to make an informed decision based on which of them dominates in their setting, taking into account the full portfolio and the lifecycle stage of their content. In our case, we find that the potential streaming revenue gains from removing UMG’s music from TikTok outweigh the compensation from TikTok at the time of the licensing dispute. Taken together, these findings provide guidance to content owners and social media platforms for evaluating and setting pricing and licensing terms.

The rest of the paper is organized as follows. In the next section, we discuss the related literature, and in Section 3, we delineate our research context by providing more detailed information on the main players involved and the nature of the licensing dispute that transpired between UMG and TikTok. In Section 4, we describe our data collection process and summarize the data features. Section 5 lays out our empirical framework, describes the main findings, and provides details of the robustness checks. In Section 6, we offer an assessment of the economic implications of our findings for UMG and TikTok. Finally, in Section 7, we summarize the paper’s findings and suggest managerial implications for the various stakeholders involved.

2. Related Literature

Our work is related to the small but growing literature in marketing and economics that examines the impact of digital platforms on music consumption and revenue generation across platforms. This work focuses largely on the impact of YouTube on music sales, and so far, the findings are mixed. Hiller (2016) analyzed the temporary removal and subsequent reinstatement of Warner Music content on YouTube in 2009 and found that the availability of popular albums on YouTube displaced Warner album sales. In contrast, Kretschmer and Peukert (2020) found that restricting access to online videos can decrease recorded music sales while enabling access tends to increase sales, as evidenced by two natural experiments in Germany—the 2009 blocking of all music videos on YouTube because of a legal dispute and the subsequent introduction of Vevo, which provided access to a large catalog of music videos. More specific to music streaming, Wlömert et al. (2024) showed that although the availability of user-generated content incorporating a specific track generally increases demand across other streaming platforms, it can cannibalize sales for new and hit releases, thereby negatively impacting overall revenue.

Our work both speaks and contributes to this debate by considering the impact of a different platform on music demand, TikTok, which has increasingly become a game changer in the music industry (Whateley 2023). There are a few important differences between TikTok and YouTube that can affect the substantive findings. On the one hand, unlike on YouTube where users typically engage with entire songs/tracks on TikTok, the music is typically embedded in user-generated videos as a backdrop, and only a portion of the full song is featured.² Thus, it is not ex ante clear whether TikTok can serve as a relevant channel for music consumption when compared with standard streaming services. On the other hand, TikTok is deeply rooted in music, and its scale is unprecedented. As indicated by UMG in its open letters (Universal Music Group 2024b), “music is at the heart of the TikTok experience,” and “TikTok is trying to build a music-based business.” Furthermore, there are more than 34 million videos posted daily on TikTok, and 85% of these feature music. Thus, TikTok surpasses all other social media platforms on this measure (i.e., content that features music), including YouTube, Instagram, and Facebook (Whateley 2023, Smith 2024, Taylor 2024). As such, even small consumption changes in the platform may have a substantial impact on demand outside the platform. In this paper, we investigate empirically whether and how music availability on TikTok impacts streaming demand on Spotify and YouTube.

Research on this specific phenomenon, that is, the effect of TikTok-like short-form video platforms on streaming demand, remains limited. Perhaps the works most relevant to our research are the following three papers. First, Yang et al. (2024) examined an exogenous boycott event in April 2021 that forced Douyin (the Chinese version of TikTok) to more proactively remove condensed TV series clips from the platform. They found that the removal of these clips reduced the demand for corresponding full-length original works on a major video streaming platform by approximately 3%, suggesting positive spillover effects from Douyin, consistent with a promotional effect. However, our findings indicate that TikTok does not influence the demand for music streaming in the same way. A notable distinction between TV and music streaming on short-form video platforms like TikTok concerns licensing aspects. Specifically, unlike TV streaming, where content is often edited into condensed clips by users without obtaining copyright permissions from the original TV series provider, the use of music on TikTok is governed by enforced licensing agreements.³ This practice necessitates music labels to consider the “fair value” of licensing their music to TikTok, given the potential for millions of social media users to consume it for free.

Second, a concurrent working paper by Winkler et al. (2024) used weekly music streaming data from a different source over a nine-week period and found that the removal of UMG tracks had a positive effect on music streams. Although we cannot definitively pinpoint the reasons for the differences in the treatment effects, it is possible that they stem from a combination of data and modeling choices; for example, our sample spans over 25-week period, we use global Spotify streams and YouTube views, and our unit of analysis is at the daily level (which gives significant power). In contrast, Winkler et al. (2024) had a larger number of tracks, but they spanned a shorter time horizon of nine weeks and were limited to only a subset of geographies (as opposed to global data), and they employed a weekly analysis. They also did not focus on the sources of heterogeneity that we do, for example, availability on TikTok. Furthermore, we employ a standard log DiD and bound our treatment effects using the Honest Difference-in-Differences approach suggested by Rambachan and Roth (2023), which provides robustness relative to their synthetic log DiD approach; see Section 5.3.1 for more detailed discussions.

Finally, in a working paper subsequent to ours, Bairathi et al. (2024) examined the same empirical context using weekly streaming data from one of our data sources over a 10-week period, focusing on a much smaller data set of 60,000 tracks and only considering tracks available on TikTok before the dispute. They reported that the removal of UMG tracks had a negative effect on music streams previously available on TikTok. They employed a synthetic levels DiD specification and argued that the null/positive effects we find are likely due to our use of the log DiD specification, which can flip the sign of the estimated treatment effect under certain conditions (McConnell 2024). We conduct numerous robustness checks to examine whether this could be the case (including running levels DiD, rescaled DiD, and testing for the condition for sign flipping across specifications; see Section 5.3.3 for details) and find that the overall main treatment effects (across all tracks) are always null/positive, and the treatment effects on the subset of tracks available on TikTok (the specific subset of music that is the focus of Bairathi et al. 2024) are consistently positive irrespective of the specification used. That is, we find no evidence of negative treatment effects for tracks available on TikTok prior to the dispute. We further point out that our findings are consistent with the new agreement later reached between the parties, wherein TikTok increased the compensation to UMG and its artists (TikTok News 2024b; Universal Music Group 2024b, c). Notably, if treatment effects were indeed negative on the whole (which suggests an overall complementarity effect), then it is highly unlikely that there would be an economic rationale for TikTok to increase its compensation to UMG.

3. Research Context

We now describe our research context, including the main players and the licensing dispute, which is the focus of this study.

3.1. Main Players

We start by describing the three main players, their sources of revenue, and their incentives.

• TikTok: TikTok is a short-form video-hosting service and one of the largest social media platforms, with more than 1 billion active monthly users in more than 140 countries (Woodward 2024). The platform is powered by user-generated content, where users create/post, share, and consume short videos. An interesting aspect of these videos is that they often use soundtracks from music labels as their backdrop (Novecore Blog 2023). This has sparked a unique video creation phenomenon on TikTok—when a video or a meme gains popularity, other creators on TikTok often jump on the bandwagon and adapt the original video to create new content, typically using the same sound as in the original post. A prime example of this phenomenon is Fleetwood Mac’s resurgence in popularity. The band’s 1977 album Rumours reentered the charts after an obscure TikTok user posted a laid-back clip of himself skateboarding and sipping Ocean Spray cranberry juice, all while grooving to the band’s hit song “Dreams” (TikTok 2020). This sound clip inspired millions to create similar videos, cementing the song’s iconic status on the platform (TikTok Newsroom 2019). As a result, many now view TikTok as a channel for users to (re)discover, share, and enjoy music.

  TikTok’s primary source of revenue is advertising. Hence, the more users spend time on and engage with the platform, the better off it is (Iqbal 2024). As such, the creation, sharing, and consumption of engaging content that draws and keeps users in the system positively impacts TikTok’s relevance to advertisers and hence, its revenues. Music has often been cited as a major component of such engaging content on TikTok (TikTok 2021).

• Music Labels: As described earlier, much of the sound used in TikTok videos comes from music licensed to the platform by the labels. There are three record labels that dominate the global music industry, also referred to as The Big Three Record Labels: Universal Music Group (UMG), Sony Music Entertainment (SME), and Warner Music Group (WMG). UMG leads with a market share of 33.90%, followed by SME at 26.91% and WMG at 15.98% (Rys 2024). Each of these labels represents a variety of well-known recording studios and artists. For example, UMG includes labels such as Interscope Records, Republic Records, Capitol Music Group, Abbey Road Studios, and prominent artists such as Drake, Billie Eilish, and The Weeknd (Universal Music Group 2024a). Similarly, SME’s portfolio includes Columbia Records, RCA Records, Arista Records, and Epic Records, which represent legendary figures such as Michael Jackson, Celine Dion, and Mariah Carey (Sony Music 2024a). Meanwhile, WMG operates labels that include Atlantic Records, Warner Records, and Parlophone Label Group, with top artists such as Ed Sheeran, Madonna, and Fleetwood Mac (Warner Recorded Music 2024).

  Music labels generate revenue from three primary sources: streaming services (such as YouTube, Spotify, and Apple Music), music sales, and licensing and synchronization fees (where the label allows partners such as social media platforms, movies, and video games to use their music; see Callaghan 2024). Streaming dominates the other two sources and accounts for more than 84% of revenues (RIAA 2024). Conceptually, these revenue streams can act as both complements and substitutes for each other. For example, if a consumer learns about a track on TikTok, he or she may stream it on Spotify; alternatively, if a consumer mostly uses TikTok to consume music, he or she may be less inclined to purchase or stream music on other channels. Thus, record labels need to have a good understanding of how each of these revenue sources affects the others in order to make informed pricing decisions for each of them.

• Streaming Services: Music streaming services are platforms where users can watch and listen to music. Spotify and YouTube are two of the largest such platforms. Spotify offers more than 100 million tracks and has more than 615 million users across more than 180 markets (Spotify 2024a). Similarly, YouTube, has a vast array of video content, which often contains music, and attracts more than 2 billion visitors monthly (YouTube News 2023). Streaming platforms license tracks from music labels and monetize this by serving ads to listeners as well as through subscription packages (for ad-free listening).

3.2. The UMG vs. TikTok Licensing Dispute

Because social media platforms like TikTok have become a primary venue for exposure to and consumption of music, they represent a double-edged sword for music studios and labels. On the one hand, they can serve as a channel for discovery and promotion, which may lead to increased demand on streaming platforms, thereby boosting the revenues of music labels. On the other hand, if users who consume music through TikTok substitute away from streaming the music elsewhere, for example, because of the fact that viewing content on TikTok is free and users’ repeated exposure to the same music may lead to “wear-out” (Pechmann and Stewart 1988), then the effect of TikTok on music labels’ revenues can be negative. With younger users spending more and more time on TikTok (Duarte 2024), labels may indeed harbor such apprehensions.

Fueled by these concerns, in early 2024, the largest music label, UMG, alleged that TikTok did not fairly compensate UMG’s artists and songwriters for using their music in the existing agreement (Curto 2024). It noted that despite TikTok’s massive user base, rapidly increasing advertising revenue, and growing reliance on music-based content, TikTok contributed only about 1% to UMG’s total revenue in 2023 (Universal Music Group 2024b). As a result, after unsuccessful negotiations, on January 30, 2024, UMG announced the termination of its licensing agreement with TikTok (Universal Music Group 2024b). This breakdown in negotiations meant that starting on February 1, 2024, TikTok users could no longer access UMG’s catalog of music previously available on the platform. There were several immediate consequences of this termination, including the removal of UMG artists’ music videos from TikTok, the removal of the tracks’ music page, and the blocking of TikTok users from leveraging this music in new video creations. Moreover, existing TikTok videos featuring these UMG songs were muted, rendering them silent. The dispute caused varied reactions among artists, content creators, and TikTok users; see Online Appendix A for a detailed description of these reactions. Yet based on an extensive review of media coverage and public relations statements, there was no apparent evidence of systematic actions taken by artists and content creators after UMG pulled its music from TikTok.

This dispute lasted till May 1, 2024, when UMG and TikTok successfully renegotiated their licensing agreement (Universal Music Group 2024c). As a part of the new agreement, TikTok agreed to deliver improved remuneration for UMG’s songwriters and artists (Aswad 2024).

4. Data

Our data for the analysis comes from multiple sources, including the websites of the Big Three record labels, as well as Chartmetric and Soundcharts (Chartmetric 2024, Soundcharts 2024). Chartmetric and Soundcharts are platforms that provide historical and real-time data and analytics for music tracks across streaming services and social media. Soundcharts integrates data from a wide array of sources and offers track metadata, streaming information, label details, etc. Chartmetric provides insights into track usage on social media platforms like TikTok such as Video Creation Numbers, and also offers proprietary artist-level metrics such as Artist Career Stage. We describe the data collection process in detail below.

First, we compiled a list of all the artists who have worked with the Big Three record labels from a combination of the labels’ official websites and their Wikipedia pages (Universal 2021; Wikipedia 2023a, b, 2024; Warner Music Store 2024; Warner Records 2024; Sony Music 2024b). This gives us a total of 2,862 artists who have worked with at least one of these labels.⁴ Next, we use Chartmetric to obtain information on each artist’s characteristics (e.g., their career stage, how many music tracks they have produced so far) and Soundcharts to obtain a complete list of all the music tracks recorded by the artist over their career. Furthermore, for each track, we collect additional information, including its label (e.g., UMG, Sony, Warner, or some other label), its release date, and a global track identifier (i.e., ISRC).⁵ Overall, this process gives us 235,741 tracks across the three main music labels.

In addition, we use Chartmetric to ascertain whether each of the tracks in our sample has a corresponding music URL on TikTok and to monitor the number of videos posted on TikTok that feature each track. These data are crucial for understanding the differential impact of the licensing dispute on UMG tracks that were available on TikTok versus those that were not available on TikTok at the start of the dispute (February 1, 2024). The former tracks were removed from TikTok’s music library because of the dispute, which resulted in the muting of videos that leveraged these tracks. The latter tracks were not on TikTok prior to the dispute; that is, they were not licensed to be incorporated in video posts by TikTok users. In contrast, tracks from SME and WMG remained unaffected; that is, their tracks already on TikTok were still available.

Finally, we use Soundcharts to collect data on the performance of all the 235,741 tracks belonging to the Big Three record labels on the two main streaming platforms, Spotify and YouTube. Our data collection covers a 180-day period—from October 10, 2023, to April 7, 2024—about a four-month period prior to the start of the dispute and a two-month period after the dispute commenced.

In Section 4.1, we summarize the time-invariant track-level data, and in Section 4.2, we describe the time-varying data used in the empirical analysis, including the track-level daily music consumption on Spotify and YouTube.

4.1. Time Invariant Track Information

We now describe the time-invariant attributes of the tracks in our data:

• ${\mathit{\text{TrackName}}}_i$: The name of track i.

• ${\mathit{\text{ISRC}}}_i$: The unique global identifier for track i, which we use to map tracks across different data sources.

• ${\mathit{\text{Label}}}_i$: Categorical variable denoting track i’s label (i.e., UMG, Sony, or Warner). Of the 235,741 tracks, 113,808 are from UMG, 53,157 are from Sony, and 70,247 are from Warner. Because some tracks change their affiliated music labels over time, a given track may be associated with more than one label in our dataset.

• ${\mathit{\text{PrevOnTikTok}}}_i$: Categorical variable denoting whether track i has a music URL on TikTok or not prior to the dispute. Of the 235,741 tracks, 28.59% were available on TikTok prior to the dispute, and the rest were not.

• ${\mathit{\text{ReleaseDate}}}_i$: The release date of track i.

• ${\mathit{\text{ArtistName}}}_i$: The artist name of track i.

• ${\mathit{\text{ArtistPopularity}}}_i$: The Spotify popularity score of track i’s artist, capturing the popularity of the artist based upon all of his or her music on Spotify. For more details on how this variable is computed see Spotify for Developers (2022).

• ${\mathit{\text{CareerStage}}}_i$: The career stage of track i’s artist. It consists of six levels: undiscovered (1.32%), developing (18.50%), midlevel (19.15%), mainstream (36.94%), superstar (17.86%), and legendary (6.23%). More detailed definitions for the career stages are available at Chartmetric (2022).

4.2. Daily Track Consumption on Spotify and YouTube

There are two main metrics of demand from a music label’s perspective: streaming and sales. We focus on streaming demand because it accounts for more than 84% of the revenue for music labels and continues to grow (RIAA 2024). In contrast, whereas music sales used to be a significant source of revenue for labels in the past, this is no longer the case; digital music sales account for only 4% of revenues, and physical sales (e.g., CDs and LPs) account for only 11%. Streaming is thus the main source of revenue for music labels and accounted for $47.7 billion globally in 2023 (Curry 2023). From consumers’ perspective, streaming has grown to be a key channel for music consumption. Collectively, Americans streamed around 4.1 trillion songs in 2023 (Luminate 2023).

Among streaming services, Spotify is the market leader, with more than 30% market share and more than 615 million monthly active users (Duarte 2024, Spotify 2024a). The next four contenders consist of YouTube, Tencent, Apple Music, and Amazon Music—all with market shares between 12% and 15% (Curry 2023). In this study, we utilize the daily demand data for Spotify and YouTube, which together represent about 46% of global streaming demand.⁶

For the time period of the analysis, for each track in our data, we gather the following demand information: (1) the number of daily streams on Spotify, which is counted as the number of times the track was listened to for 30 seconds or more on a given day (Spotify 2024b); and (2) the number of views on YouTube, which is counted as the number of times a video was watched for at least 30 seconds on a given day (Tuberanker 2022).⁷

The summary statistics of these two demand metrics for the entire observation period (across all tracks and periods) are shown in Table 1. We find that these distributions are quite skewed with long tails; that is, some tracks (on some days) get extremely high demand, running into billions of streams/views, but the bulk of the daily streams/views is much smaller. The median demand for a track on Spotify is 228 daily streams, whereas the median demand on YouTube is about 281 views per day. We also calculate the average daily demand by track and present these track-level summary statistics in Table 2. Additionally, we show the summary statistics of daily music consumption on Spotify and YouTube for the pretreatment period in Tables A.1 and A.2 in Online Appendix B.

Table 1. Summary Statistics of Daily Music Consumption

Table 2. Summary Statistics of Average Daily Music Consumption by Track

5. Empirical Analysis

We now present our empirical analysis and findings.

5.1. Main Effect on Music Demand

The breakdown of the licensing agreement between UMG and TikTok provides a quasi-natural experiment for our study. In this context, the treatment is the removal of UMG music from TikTok. Consequently, UMG tracks that were previously available on TikTok were withdrawn, and videos leveraging those tracks were muted. As such, UMG tracks form our treatment group, whereas tracks from Sony and Warner, none of which were removed during the licensing dispute, serve as the control group.

We use a Difference-in-Differences (DiD) specification, which is a widely applied strategy for evaluating the effect of an intervention or treatment (e.g., the removal of UMG music from TikTok) on an outcome variable of interest (e.g., Spotify streams and YouTube views). A DiD analysis estimates the treatment effect by comparing the difference in the changes in the outcome variable between the two groups (e.g., treatment and control). Our estimation relies on the following DiD specification:

where our dependent variable ${\mathit{\text{Demand}}}_{it}$ is the music consumption of track i on day t on the streaming service, that is, the number of streams on Spotify or the number of views on YouTube. We logged our dependent variable ${\mathit{\text{Demand}}}_{it}$ to address skewness and because we are interested primarily in understanding the outcome in percentage terms (Fouka 2020). ${\mathit{\text{UMG}}}_i$ is an indicator equal to 1 if track i belongs to UMG and 0 otherwise. ${\mathit{\text{Post}}}_t$ equals 1 if date t is after Jan. 31, 2024, and 0 otherwise. The coefficient of interest is $\beta$, which represents the effect of track i belonging to the treatment group (i.e., removal of UMG music from TikTok) on its streaming demand relative to the counterfactual scenario where UMG music is not removed from TikTok. ${\mathit{\text{Track}}}_i$ captures track fixed effects, and ${\mathit{\text{Date}}}_t$ captures date fixed effects.

Table 3 reports the results from the estimation of the DiD model shown in Equation (1). Column (1) features the results for Spotify demand and column (2) for YouTube demand. As we can see, the main effect for both streaming platforms is insignificant. In Section 5.3.3, we run a series of robustness checks on this analysis using the demand directly as the dependent variable (instead of the log of the demand variable) and find largely consistent results, albeit under some specifications the main effect can be slightly positive, although its magnitude is sensitive to outliers and subject to model fit issues. This could be because some alternative specifications pick up the actual level of change, which may be negligible in percentage terms under the log-specification (see more details in Section 5.3.3). Overall, this suggests that the main effect is either null or slightly positive.

Table 3. Main Effect of Removing UMG Music from TikTok on Music Demand (Log-Specification)

5.2. Heterogeneity in Treatment Effects

We now explore whether the overall null result from estimating Equation (1) on the full data set masks any significant heterogeneity effects across different subgroups. In particular, one important dimension of interest for UMG is the strategic decision of whether only some, none, or all of an artist’s tracks should be made available on TikTok. Tracks that are available on the platform can be incorporated by creators into their posts and subsequently consumed by TikTok users, whereas tracks that are not available cannot be incorporated. Furthermore, because any potential cross-platform effects require the presence of at least some UMG tracks on TikTok, we use the variable ${\mathit{\text{PrevOnTikTok}}}_i$, described in Section 4.1, to capture whether a track was available on TikTok before the dispute. For brevity, we present the main analysis with the log DiD specification estimated over tracks with different TikTok presence. In addition, we perform a series of exhaustive robustness checks on the heterogeneity analysis and show that the results are valid and consistent elsewhere in Section 5.3.

We start by examining whether the following two groups—tracks that were previously available on TikTok prior to the dispute and tracks that were not available on TikTok prior to the dispute—differ from each other and, if so, how.

First, we look at the pretreatment demand level for each group. Table 4 shows the track-level average daily streams on Spotify, and Table 5 shows the track-level average daily views on YouTube for each group. As we can see, the two groups are systematically different—tracks that had been available on TikTok tend to be more popular than those that had not. For example, the median number of daily Spotify streams is 2,353 for tracks that were available on TikTok compared with just 82 for those that were not available on TikTok. Next, we examine the distribution of artists by career stage across the two groups; see Table 6. One interesting observation is that tracks that were available on TikTok before the dispute are almost three times more likely to be from superstar artists. Overall, it seems that tracks available on TikTok (pre-dispute) tend to be more popular and from more successful artists compared with tracks that were not available on TikTok. Finally, we note that it is highly unlikely that a track released before the dispute but unavailable on TikTok would have become available during the treatment period in a counterfactual world without the dispute. This is because the availability patterns prior to the dispute (i.e., up to January 31, 2024) and historical trends show that TikTok availability was remarkably stable; 99.88% of tracks that were available remained so throughout the pre-dispute period, whereas only a small fraction—0.12%—transitioned from being unavailable to available (as indicated by the consistent reporting of video creation numbers for each track).

Table 4. Summary Statistics of the Pre-Period Daily Spotify Streams: Tracks Previously on vs. Not on TikTok

Table 5. Summary Statistics of the Pre-Period Daily YouTube Views: Tracks on vs. Not on TikTok

Table 6. Artist Career Stage Distribution for Tracks Available on vs. Not Available on TikTok Prior to the Dispute

We now estimate the DiD specification in Equation (1) separately for these two groups of tracks and present the results in Table 7. Columns (1) and (2) display the regression results, whereby the treatment group consists of UMG tracks that were available on TikTok prior to the dispute, whereas the control group comprises Sony and Warner tracks that were also available on TikTok prior to the dispute. Columns (3) and (4) display the regression results whereby the treatment group consists of UMG tracks that were not available on TikTok prior to the dispute, whereas the control group comprises Sony and Warner tracks that were not available on TikTok prior to the dispute. We discuss both sets of results in turn.

Table 7. Heterogeneous Effects of Removing UMG Music from TikTok on Music Demand: Tracks on vs. Not on TikTok

The dispute had a differential implication for the various music tracks in our data. As we can see from columns (1) and (2) in Table 7, the estimated coefficient of the treatment status indicator, 1.UMG#1.post, is positive and significant for streams on Spotify (b = 0.0229, p < 0.001) and views on YouTube (b = 0.0216, p < 0.1). This indicates that the removal of UMG tracks that had been available on TikTok prior to the dispute led to a 2.32% (=$e^{0.0229}-1$) increase in the demand for these specific tracks on Spotify and a 2.18% (=$e^{0.0216}-1$) increase in their demand on YouTube compared with the counterfactual scenario where these UMG tracks had continued to be available on TikTok. This suggests a substitution effect for tracks that were available on TikTok prior to the dispute, indicating that these tracks received greater streaming demand after being withdrawn from TikTok’s music library.

This substitution effect likely stems from the following reasons. Music plays a central role in the TikTok experience, often serving as the backbone of video creation and virality (Crawford 2024). Popular tracks are frequently reused across numerous videos, particularly in the context of trending challenges or hashtag-driven trends, and are then typically viewed by many users. However, this repeated exposure to the same songs can lead to content fatigue, reducing users’ desire to consume those tracks elsewhere—a phenomenon supported by psychological research showing that repeated exposure, even to partial content, can lead to a wear-out effect (Margulis 2013, Zulli and Zulli 2022). Notably, in an experiment conducted by TikTok in Australia and New Zealand, the platform temporarily restricted access to popular songs for some users. The result was a significant decline in user time spent on the platform (D’Souza 2024), highlighting how central music is to the platform’s appeal.⁸ Furthermore, for some users, TikTok may serve as a key channel for music consumption. This role has been amplified by the platform’s support for longer video formats, which are increasingly popular among audiences. In fact, TikTok videos that are longer than one minute receive 63.8% more watch time than videos between 30 and 60 seconds (Duffy 2023, Hutchinson 2025). Therefore, in the absence of music-driven engagement on TikTok, users may turn to alternative streaming platforms, such as Spotify or YouTube, to satisfy their music consumption needs.

These findings echo the concerns of music labels and provide evidence in support of the cannibalizing impact of TikTok on tracks and artists. At the core of the dispute was UMG’s allegation that TikTok did not adequately compensate its artists and songwriters. For example, Music Business Worldwide (MBW) used data from Chartmetric to analyze the Top 1,000 most popular TikTok videos featuring Kate Bush’s “Running Up That Hill” and found that these videos collectively garnered almost 5 billion views/plays on TikTok. However, unlike streaming platforms such as Spotify, where musicians get paid based on the total number of streams, TikTok’s royalty payment system is based on the number of video creations that use a song (Hypebot 2023), which is typically orders of magnitude lower than the number of views. Notably, whereas Kate Bush’s track garnered nearly 5 billion unpaid views on TikTok, it was streamed only 400 million times on Spotify. Although this is just one anecdotal observation, examples like this abound and suggest that TikTok likely has a cannibalization/substitution effect on songs that would otherwise be consumed more heavily on revenue-sharing streaming platforms like Spotify (Ingham 2022).

Next, columns (3) and (4) in Table 7 show the estimation results for those tracks that were not available on TikTok prior to the dispute. Here, the treatment status indicator, 1.UMG#1.post, is negative and significant for both streams on Spotify (b = −0.0142, p < 0.001) and views on YouTube (b = −0.0266, p < 0.01). This implies that UMG tracks that were not available on TikTok prior to the dispute experienced a 1.41% (=$1-e^{-0.0142}$) decrease in Spotify streams and a 2.62% (=$1-e^{-0.0266}$) decrease in YouTube views in the period after the dispute. This suggests a complementarity effect for tracks that were not available on TikTok prior to the dispute, with these tracks indirectly affected by UMG’s decision to remove its music from TikTok.

This finding can be viewed as supporting the promotional and discovery role of TikTok, especially for content not already on the platform, which tends to be less popular and originate from less renowned artists. By not allowing TikTok users to incorporate any UMG music into their videos, the label likely hindered opportunities for artist discovery. In particular, TikTok users may discover artists who are new or previously unknown to them through their available tracks on TikTok. And after becoming familiar with these artists, users may search for them on music streaming platforms, such as Spotify, and discover other tracks by these artists that are not available on TikTok. For instance, music producer L Dre witnessed a remarkable rise in his Spotify monthly listeners after his track “Steven Universe” was incorporated in more than 10 million TikTok video creations, prompting fans to explore his other music on Spotify (Cirrkus 2022). Note, however, that this “promotional and discovery” explanation for the negative effect of the dispute on tracks not available on TikTok requires that the artists performing them have at least some other tracks that are available on the social media platform.

We next test TikTok’s afore-described promotional and discovery role for artists with partial track availability on the platform. We do so by further segmenting the tracks not available on TikTok in the pre-dispute period into two subgroups (for both the treatment and control groups): tracks from artists who had no tracks on TikTok prior to the dispute and tracks from artists who had some of their tracks available on TikTok prior to the dispute. We then examined how the removal of UMG music from TikTok affected the streaming demand for these two subgroups of tracks. As before, we estimate the DiD model in Equation (1) on these two subgroups separately and present the results in Table 8. Columns (1) and (2) display the results for tracks from artists with no TikTok presence prior to the dispute, whereas columns (3) and (4) show the results for tracks by artists with partial TikTok coverage.

Table 8. Heterogeneous Effects of Removing UMG Music from TikTok on Music Demand for Tracks Not Available on TikTok Prior to the Dispute: Artists with No vs. Partial TikTok Availability

First, for artists who had no tracks available on TikTok, we do not observe any significant impact; see columns (1) and (2) in Table 8, where the estimated coefficient of the treatment status indicator, 1.UMG#1.post, is insignificant. That is, for artists with no TikTok presence, the removal of UMG music from TikTok has no significant impact on their Spotify and YouTube demand. In contrast, for artists who had partial TikTok availability (i.e., who had other tracks that were available on TikTok), there is a significant negative impact of the dispute; see the estimated coefficient of the treatment status indicator, 1.UMG#1.post in columns (3) and (4) in Table 8. Together, these results suggest that the negative, or complementarity, effects estimated in columns (3) and (4) in Table 7 are driven mainly by tracks from artists who had some (but not full) presence on TikTok before the dispute.⁹ This supports the hypothesis that TikTok can serve as a promotional and discovery channel for artists (TikTok News 2024b); when they gain some traction on TikTok through certain tracks, their other tracks (which are not available on TikTok) tend to be discovered and streamed elsewhere. Although this analysis is not a formal test, it provides some evidence for the idea that TikTok can possibly play a complementary, promotional, and discovery role for artists.

Finally, we note that TikTok’s discovery role for tracks unavailable on the platform reflects a potential violation of the SUTVA. As discussed above, the discovery role for tracks not available on TikTok likely arises from spillovers generated by an artist’s other tracks that had previously been available on TikTok, implying that the “no interference between units” assumption does not hold. Therefore, it is important to note that our estimated treatment effects are valid only when “all” of UMG’s music goes silent on TikTok, and we cannot comment on what the effect would look like for a specific track if it is the only one going silent.

In summary, we find that TikTok has a differential effect on music tracks, depending on whether they had versus had not been available on the platform prior to the dispute. For the former type of tracks, there is a substitution effect; that is, they received greater streaming demand after being removed from TikTok’s music library. These tracks tend to be more popular and come from more well-known artists. However, for the latter type of tracks, there is a complementarity effect; that is, they were adversely affected by UMG’s decision to remove music from TikTok. These tracks tend to be less popular and less likely to be recorded by superstar artists. We further find that the complementarity effect is potentially driven by the promotion and discovery role that TikTok can play for artists with a partial presence on TikTok; once their tracks that had previously been available on TikTok were removed, this had a negative impact on the streaming of their other tracks not previously available on the social media platform, suggesting an indirect impact. Together, these findings support the arguments of both UMG and TikTok, albeit for different subgroups.

We further investigate whether the overall null effect observed in the full data set conceals meaningful heterogeneity along dimensions beyond track availability on TikTok. In particular, we examine whether the impact of the UMG music removal varies by track popularity and by artist characteristics, such as career stage and popularity level. Our analysis reveals systematic variation along these dimensions; more popular tracks and artists tend to benefit from the removal, whereas less popular counterparts are generally more negatively impacted. For further details on the heterogeneity analyses, please refer to Online Appendix D.

5.3. Robustness Checks

We now describe a series of robustness checks on the analysis presented above. First, in Section 5.3.1, we test the assumptions required for the validity of the DiD models we used. Next, in Section 5.3.2, we examine the sensitivity of our findings to outliers and to different time frames for the pre- and posttreatment periods; we also report on analysis that incorporates time-varying variables. Lastly, in Section 5.3.3, we consider a variety of alternative model specifications.

5.3.1. Model Assumptions and Validity Tests.

The validity of the DiD model we employed in our analysis depends on a few key assumptions. We briefly describe these assumptions and our empirical tests to validate them in the text below, with further details available in Online Appendixes E, F, and G.

Parallel Trends.

A key assumption of the DiD model pertains to parallel trends; if the treatment group had not received the treatment, then the trend in this group’s outcomes would have been the same as the trend in the control group’s outcomes (Angrist and Pischke 2009). Therefore, we first compare the pretreatment trends in music demand on Spotify and YouTube for the treatment and control groups by estimating a relative-time model (Autor 2003) at the daily level over our entire time window. The results show that some pretreatment coefficients are significant, indicating a violation of the parallel trends assumption.

However, in our case, the pretreatment period is quite long (113 days), and there is a lot of variance in demand over time because of seasonality, holidays, platform-specific time-varying shocks to traffic, etc., so naïve relative-time trend tests can be noisy. Therefore, we estimate a linear trend model (Bhuller et al. 2013, Goodman-Bacon 2018) to assess the magnitude of the difference in pre-trends across treatment and control groups. We find that the linear trend difference estimates are either insignificant and/or very small (compared with the magnitude of the estimated treatment effects, e.g., 0.3%–1.5% of the size of the treatment effect).

In addition, we provide an exhaustive sensitivity analysis of the extent to which treatment effects could be influenced by potential violations in parallel trends. The standard way to do this is to use the Honest Difference-in-Differences (Honest DiD) approach proposed by Rambachan and Roth (2023). This method allows the researcher to bound the estimated treatment effects for a given magnitude of violation of the parallel trends assumption, that is, quantify the magnitude of the violation of parallel trends that can be tolerated before it nullifies/reverses the estimated treatment effect while considering the worst-case scenario over the pretreatment periods. This approach has been used extensively in applied DiD work in recent years to bound treatment effects (Ang 2021, Dustmann et al. 2022, Christensen et al. 2023). Thus, we also adopt this approach, and our analyses suggest that even when allowing for significant deviations from the linear extrapolation across consecutive periods (e.g., 1,774 times greater than our current linear trend in the pretreatment period), we cannot reject our main effect or heterogeneous treatment effects. All the details of the parallel trends tests are shown in Online Appendix E.

Finally, we note that although there are alternative approaches to estimating treatment effects when the parallel trends assumption is violated, such as Synthetic Difference-in-Differences (SDiD) (Arkhangelsky et al. 2021), these methods come with drawbacks. First, SDiD requires a balanced panel data set, which would result in dropping many observations from our data and render the sample nonrepresentative. Second, constructing weights for a long pretreatment period (113 days in our case) in SDiD is computationally inefficient. As such, we would need to aggregate the data to weekly/monthly levels to effectively match on pre-trends, and this would lead to a significant loss of power. Third, given the variance in demand for music across time and tracks, matching too closely on pre-trends can lead to matching on noise/variance in the pretreatment period. Indeed, such matching can introduce its own source of bias (Daw and Hatfield 2018), leading to results that are quite sensitive to the length of the pretreatment period used for matching, the level of aggregation, and which time periods and tracks are dropped to achieve a balanced panel. As such, a more commonly adopted approach in applied work is to bound the estimated treatment effects using the Honest DiD method, as described above. Moreover, as we later discuss in Section 5.3.3, the results are consistent even when we use other specifications.

Level Differences Between Treated and Control Groups.

In addition to the parallel trends, it is usually a good idea to confirm that the levels of the treatment and control groups in the pretreatment period are comparable (McKenzie 2020). This provides additional assurance for the validity of the DiD analysis. To that end, in Online Appendix G.1, we plot the distributions of the daily demand of tracks from all three music labels—UMG, SME, and WMG—and confirm that these distributions are largely similar in levels. Moreover, we recenter the outcome distribution of the treated group to align the baseline means and rerun the analysis after ensuring that there is no disparity between the treatment and control groups. Our findings remain consistent after rescaling, as detailed in Online Appendix G.4.

SUTVA.

Another potential concern is that the treatment has a spillover effect on the control group, resulting in a violation of the SUTVA (Stable Unit Treatment Values Assumption) required for the DiD model. In particular, one issue could be that when UMG’s music disappeared from TikTok, users on the platform may have switched to using music from SME and WMG, leading to a substantial upswell in their TikTok usage (which in turn could have impacted the control group’s demand on Spotify and YouTube). To examine whether this is the case, we collect data on the number of new TikTok videos uploaded daily that use music from different labels before and during the treatment period. We do not see any significant jump in the use of SME and WMG music after the dispute. Furthermore, we analyzed Spotify streams and YouTube views for SME and WMG music by fitting a linear trend model to examine whether the silencing of UMG music on TikTok led to a significant surge in demand for SME/WMG tracks in the post-dispute period. Our analysis did not reveal any consistent demand shifts across either platform. Lastly, we examined global downloads and usage of the TikTok app before and after the licensing dispute, finding continued growth in monthly active users. Thus, it is unlikely that there were any significant spillover effects from the treatment on the control group during the treatment period. See Online Appendix F for details of the analysis and results. However, as discussed earlier, our findings indicate that there are likely spillover effects between units within the treatment group (i.e., among UMG tracks). As such, our findings and counterfactual estimates are valid only in settings where all UMG tracks are treated (go silent) and all control group tracks remain available on the platform.

In summary, we find that the data patterns are generally supportive of our identification strategy. However, some of the identifying assumptions are not directly testable with our data, even though they are unlikely to have been violated. First, a comprehensive review of public news sources reveals no evidence that other major labels, such as Sony, Warner, and various independent labels, altered their policies regarding TikTok during the dispute period. Second, we do not observe any coinciding changes in demand or artist-driven factors that would suggest endogeneity in the timing of Universal’s music withdrawal from TikTok. Based on a comprehensive search of public news sources (factiva), third-party databases (data.ai), and UMG’s quarterly financial reports (https://investors.universalmusic.com/report), we found no indications that UMG altered its advertising or marketing strategies in response to the dispute. Third, we find no anecdotal or empirical evidence suggesting that Universal’s withdrawal of its content from TikTok triggered any immediate changes in Spotify’s/YouTube’s recommendation algorithms. For example, we found no reported evidence of manipulation efforts by UMG—such as the use of bots—to influence Spotify or YouTube streaming behavior during the disputed period. Given the high-profile nature of the action taken by UMG and the ramifications of the dispute, we would have expected any major strategic shifts or efforts—either by UMG or its competitors—to have attracted media attention. The absence of such media coverage provides some additional reassurance. Nevertheless, we acknowledge that we cannot empirically rule out these considerations.

5.3.2. Other Robustness Checks.

Sensitivity to Outliers.

To ensure the robustness of our main log-specification, we also assess the sensitivity to potential outliers. Specifically, we exclude observations with Spotify streams or YouTube views above the 99th percentile when estimating our main and heterogeneous treatment effects. Full details of this robustness check are provided in Online Appendix H.1. The results remain consistent with our main findings in Section 5.1 as well as with the heterogeneous treatment effects discussed in Section 5.2.

Different Time Windows.

We also examine the sensitivity of our results to changes in the lengths of the pre- and posttreatment periods. First, we shorten the pretreatment window from 113 days (as used in the main analysis) to 90, 75, and 67 days. Notably, the 67-day window results in a symmetrical design, with both the pre- and posttreatment periods spanning an equal number of days. Next, we vary the posttreatment window from 67 days to shorter durations of 60 and 45 days. Details of this robustness check are provided in Online Appendix H.2. Across all these different time windows, the results remain broadly consistent with both our main findings in Section 5.1 and the heterogeneous treatment effects in Section 5.2.¹⁰

Control for Time-Varying Factors.

One potential concern is that track fixed effects may not fully account for time-varying factors that influence music demand on Spotify and YouTube, which could potentially challenge the validity of our substitution argument. To address this, we conduct a robustness check that accounts explicitly for such factors. Specifically, we collect daily media coverage data for the Big Three labels and their top artists using Factiva, a comprehensive digital archive of global news content from more than 33,000 sources, including newspapers, industry publications, and websites. Even controlling for daily media coverage, our results remain consistent with the main findings presented in Section 5.1, as do the heterogeneous treatment effects discussed in Section 5.2. Full details of this robustness check are provided in Online Appendix H.3.

Definition of Treatment and Control Groups.

We examine the robustness of the results to the definition of the treatment and control groups. Specifically, one concern is that some tracks that were produced by other labels were licensed to UMG and hence, may not have been silenced on February 1, 2024. To examine this issue, we first manually checked the TikTok music pages for a random sample of 100 UMG tracks after our data collection and found that all were unavailable, with the associated videos rendered silent.¹¹ Next, we examined the number of TikTok video creations for UMG tracks in our data set. We found that only 303 UMG tracks continued to show positive video creation numbers after February 6, 2024, most of which were removed by February 27, 2024. This was likely due to UMG’s subsequent takedown of songs by songwriters under contract to Universal Music Publishing Group (Variety 2024). Although this second removal affected only a small number of tracks in our data, it could, in principle, affect the estimated treatment effects. Therefore, we conduct another analysis where we exclude the 303 tracks that had not been removed by February 6, 2024. The results are consistent with our main analysis in Section 5.1 and the heterogeneous treatment effects in Section 5.2. Details of this analysis can be found in Online Appendix H.4.

5.3.3. Specification Checks.

So far, we used a logged dependent variable in our analysis. We did so because of the nature of the data, which are highly skewed, and because we are interested primarily in understanding the outcome in percentage terms (Fouka 2020). However, a recent working paper by McConnell (2024) suggested that using a log specification in DiD models, as opposed to a levels specification, can result in the sign of the treatment effect flipping under certain conditions (especially when the mean levels of the treated and control groups show large differences in the pretreatment period). To examine whether this is a concern in our setting, we consider additional tests and alternative specifications.

First, we examine the extent to which the outcome distributions for the treated and control groups are different. Although we observe some differences, they are relatively very small in magnitude. Nevertheless, we formally test whether the moments in our data satisfy the condition for sign flipping when going from levels to logs (McConnell 2024). We find that this condition is not satisfied in our data; that is, based on the data patterns, we would not expect the sign to flip. See Online Appendix G.1 for details.

Second, we estimate a DiD model in levels as follows:

Table 9 presents the results from estimating the levels-specification DiD model in Equation (2). Column (1) suggests that the demand for UMG’s music tracks on Spotify streams increased by 684,307.2, compared with the counterfactual scenario where tracks would have remained available on TikTok. Column (2) shows the effect of the results on YouTube demand, which is null. We also confirm that the parallel trend assumption holds for both of these level models; see Online Appendix E.3 for details.

Table 9. Main Effect of Removing UMG Music from TikTok on Music Demand (Levels-Specification)

The main difference between this analysis and that presented in Section 5.1 is the large positive treatment effect on Spotify. Based on the daily average pretreatment demand for UMG streams on Spotify (135,222 per Table A.1 in Online Appendix B), this translates into a 506% increase in daily Spotify streams. This unreasonably large effect likely stems from the poor fit and extreme skewness of our data; notice that the $R^2$ for Spotify demand in column (1) in Table 9 is considerably lower than that for the log specification results in Table 3. To examine the extent to which this effect is driven by outliers, in Table 10 we also report results for the case where observations with Spotify streams above the 99th percentile were excluded. These results remain directionally consistent with Table 9, but $R^2$ improves significantly, and the coefficient decreases by orders of magnitude. This implies a high degree of sensitivity in the magnitude of the estimated treatment effects when using a levels specification in the DiD model.¹² Overall, these results suggest that the licensing dispute had a null or a possibly small positive effect on UMG’s demand in streaming platforms.

Table 10. Main Effect of Removing UMG Music from TikTok on Music Demand (Levels Specification, Dropping Observations Larger than the 99th Percentile)

Importantly, all the heterogeneity analysis results using the levels DiD are consistent with our earlier findings reported in Table 7; see Table A.22 in Online Appendix G.3 for details. That is, we find that because of the dispute, tracks that had been available on TikTok exhibit a positive (substitution) effect, whereas tracks that had not been available on TikTok exhibit a negative (complementarity) effect. In a recent working paper subsequent to ours, Bairathi et al. (2024) used a levels specification on a weekly DiD model, focused only on tracks available on TikTok, and documented a negative treatment effect for those tracks. However, despite multiple stress tests, we consistently see a positive treatment effect for tracks on TikTok, including the levels DiD analysis described above, the log DiD analysis described in Section 5.2, and an additional rescaled analysis described in Online Appendix G.4 (where we scale the outcome variable to align the baseline outcome means before the log transformation).¹³ Furthermore, as we will discuss in Section 6, our findings and treatment effects are largely consistent with the eventual resolution of the dispute (unlike the implications of the negative effects on Spotify demand found in Bairathi et al. 2024) because TikTok increased the compensation for UMG artists. Finally, because the results for the analysis by subgroups (of tracks available versus not available on TikTok) are always consistent across all types of robustness checks, we will focus on these heterogeneous effects in our economic impact discussion section.

6. Economic Impact

We now present a simple back-of-the-envelope assessment of TikTok’s economic impact on UMG’s streaming revenues based on our estimates and data. For this analysis, we use the heterogeneous estimates from Section 5.2. We calculate the impact on UMG’s annual streaming revenue from Spotify in the scenario where its tracks are removed from TikTok. Based on the findings in Section 5.2, we know that (1) for music available on TikTok, there is a substitution effect on Spotify, which implies an incremental gain in demand if UMG’s music is removed from TikTok, and (2) for music not available on TikTok, there is a complementarity effect, which implies an incremental loss in demand if UMG’s music is removed from TikTok. To the extent that the baseline demand for these two groups is different, it is possible that the overall net impact on revenue is nonzero (i.e., not null). Therefore, we calculate the net impact on annual revenue from Spotify as follows:

We can further expand the two terms on the right-hand side as

where ${\beta }_{\text{OnTikTok}}^S$ and ${\beta }_{\text{NotOnTikTok}}^S$ are the incremental impacts on the daily demand for the two groups in the counterfactual scenario. Based on the parameters in columns (1) and (3) in Table 7, this translates to ${\beta }_{\text{OnTikTok}}^S=+2.32\%$ and ${\beta }_{\text{NotOnTikTok}}^S=-1.41\%$. Furthermore, ${\mathit{\text{BaselineDemand}}}_i^S$ denotes track i’s average daily demand on Spotify in the pretreatment period, $0.003 represents the per-stream average payment that Spotify remits to the music label (RouteNote 2022),¹⁴ and 365 refers to the number of days in a year. Note that the calculation is over the set of tracks in the two groups, and we know that $N_{\text{OnTikTok}}=35,837$ and $N_{\text{NotOnTikTok}}=77,961$. This yields ${\text{Gain}}^S=340.87$ million USD and ${\text{Loss}}^S=24.15$ million USD, for a net revenue gain of $\approx 316.72$ million USD per year.¹⁵ This suggests that by removing its music from TikTok, UMG could gain more than $300 million per year in revenues from Spotify.

To provide managers and policymakers a better understanding of the distribution of outcomes, we enrich our back-of-the-envelope calculations by considering the confidence intervals for the treatment effects. Based on the parameters in columns (1) and (3) in Table 3, the 95% confidence interval for ${\beta }_{\text{OnTikTok}}^S$ is [0.01483, 0.03071], whereas the interval for ${\beta }_{\text{NotOnTikTok}}^S$ is [−0.01922, −0.00915]. These intervals imply that the removal of UMG tracks from TikTok led to an estimated increase in Spotify demand for tracks that were available on TikTok prior to the dispute by 1.49%–3.12% (i.e., $e^{0.01483}-1$ to $e^{0.03071}-1$) and a corresponding decrease in Spotify demand for tracks not previously on TikTok by 0.91%–1.90% (i.e., $1-e^{-0.00915}$ to $1-e^{-0.01922}$). Thus, the estimated annual Spotify revenue gain (${\text{Gain}}^S$) for UMG ranges from $228.328 million to $478.111 million, whereas the estimated loss (${\text{Loss}}^S$) ranges from $15.587 million to $32.543 million. This yields a net revenue gain of between $195.8 million and $462.5 million per year.

In summary, the removal of UMG tracks from TikTok’s music library leads to an expected net revenue gain of approximately $317 million from streaming on Spotify, with estimates ranging from $196 million to $463 million. We can contrast this number with the status quo at the time of the dispute. In 2023, UMG’s annual revenue was approximately $11.11 billion, of which TikTok contributed only 1%, or approximately $111 million (Universal Music Group 2024b), which implies a revenue shortfall of approximately $206 million, with a range of $85 million to $352 million.¹⁶ Taken together, these calculations suggest that UMG can make the case that it should gain more in revenues than it currently does from its partnership with TikTok, even if our calculations err somewhat on the side of supporting UMG’s claims.¹⁷ Notably, on May 1, 2024, UMG and TikTok announced a new licensing agreement (Universal Music Group 2024c) that promises to “improve remuneration for UMG’s songwriters and artists,” a move that aligns with our findings. Of course, the final resolution likely reflects a broader set of considerations, including concerns over a possible decline in TikTok user engagement following the removal of popular music (D’Souza 2024), as well as the platform’s potential long-term gains from restoring its partnership with UMG.

More broadly, our findings and analysis also invite further discussion on the optimality of the licensing/compensation model between social media platforms like TikTok and music labels like UMG. So far, UMG and other music labels do not receive any direct compensation for the number of views/streams on TikTok of a given track; rather, the compensation is based on the number of TikTok videos that used the track. As the aforementioned Kate Bush example highlights, these two metrics can be orders of magnitude different from each other. Recall that the top 1,000 TikTok videos featuring Bush’s track “Running up the Hill” garnered nearly 5 billion views (Hypebot 2023). If UMG were to treat these views in a similar fashion to Spotify streams, then this would translate to a very significant monetization opportunity for UMG. Although views of TikTok videos that use music tracks (as their audio backdrop) should likely be compensated at lower rates than Spotify streams, because the videos also include new original content made by TikTok users and often do not play the entire track, they could still represent a substantial revenue stream for UMG.

To get a sense of the scale of this potential revenue, in Table 11 we list the top 10 UMG tracks on TikTok based on the number of videos that incorporate these tracks (as of April 7, 2024, and compiled from Chartmetric). For each of these tracks, we show the total number of views that the top 100 videos using that track garnered. For example, the top 100 videos featuring the soundtrack “good 4 u” garnered more than 24 billion views on TikTok. Collectively, the videos featuring the top 10 UMG tracks garnered more than 129 billion unpaid TikTok views. If UMG were to charge TikTok a streaming fee similar to Spotify ($0.003 per view), this would translate to a revenue gain of more than $387 million, which is quite significant. Note that this revenue calculation considers only the top 100 video recreations for the top 10 UMG tracks; if we were to consider the full UMG collection on TikTok and all their video recreations and views, this number would be much higher. As such, this estimate should be considered as a lower bound on the potential revenue gains from moving to this alternative revenue model (or to some combination of compensation for track usage in a TikTok video creation and the subsequent views of that creation). In summary, we find that pulling UMG’s music from TikTok can lead to significant positive revenues from other sources and that UMG may be under-monetizing its music on TikTok by not charging for views directly. These findings also suggest that music labels can further sharpen their licensing agreements with social media platforms like TikTok without undercutting their streaming revenues.

Table 11. Views of the Top 10 UMG Tracks Used on TikTok Based Upon Their 100 Most Popular Video Creations

Finally, we note that our economic impact calculations make a series of simplifying assumptions. As such, they are intended to give readers a sense of the scale of the economic impact (rather than serve as exact numbers) and should be taken with the appropriate caveats. For instance, we do not account for music streams on other platforms like YouTube (where the estimates are marginally significant based upon fewer observations) and SoundCloud (where UMG music is also available but we do not have access to data from this platform). We further do not consider the potential impact on digital music sales (e.g., on Apple Music) or direct album sales. Additionally, our estimates are based on the short-term change in demand (within a few months of the removal of UMG’s music from TikTok). It is unclear whether the long-term effects on streaming demand would be similar in magnitude. Lastly, platforms like TikTok may provide artists and studios with other benefits not quantified in our analysis, for example, a channel to shape the popular zeitgeist, a venue for interacting with fans and other artists, and an outlet for influencing popular trends in music and culture. Nevertheless, the analysis is intended to serve as a conservative first step in quantifying the impact of social media platforms like TikTok on music streaming demand and revenue and also to provide some insights into the potential profitability of alternative revenue models.

7. Conclusion

Our study focuses on a recent music licensing dispute between UMG and TikTok, which highlights important questions about the consumption, promotion, and monetization of music in the era of social media. At the heart of the dispute is UMG’s argument that TikTok’s compensation had been “unfair” because it failed to adequately compensate the label, its artists, and songwriters for the usage of and exposure to tracks on the platform. In particular, extensive exposure and repeated consumption of music tracks on TikTok could potentially diminish listeners’ interest in other paid streaming services such as Spotify. Conversely, TikTok maintained that its platform “fairly” compensates the label and its artists because it enhances their visibility and fosters discovery, which in turn can boost demand across various music streaming platforms.

We leverage this dispute as a natural quasi-experiment, using the removal of UMG’s music tracks from TikTok as the treatment group and comparing them to tracks from Sony Music Entertainment (SME) and Warner Music Group (WMG), which remained available. Our Difference-in-Differences analysis shows that the removal of tracks that had been available on TikTok led to a 2%–3% increase in the consumption of these tracks on Spotify and YouTube, indicating a substitution effect and supporting UMG’s concerns about unfair compensation. Conversely, tracks not previously available on TikTok experienced a 1%–3% decrease in streams on Spotify and YouTube, suggesting a complementarity effect. Our results further indicate that the complementarity effect is driven by the promotion and discovery role that TikTok can play for artists with a partial presence on the social media platform. Specifically, once their tracks that had previously been available on TikTok were removed, it negatively impacted the streaming of these artists’ other tracks that were not previously available on TikTok. Taken together, the findings support the arguments of both UMG and TikTok, albeit with respect to different groups.

We further note that several of our results suggest possible mechanisms underlying the findings. For example, because the substitution effect is linked to the more popular tracks that had been available on TikTok prior to the dispute, it could reflect user “wear-out” (because of repeated exposure), and because the complementarity effect is associated specifically with tracks by artists that had partial availability on TikTok, it could reflect a discovery and promotion role of the platform. Notwithstanding, future research, particularly with access to listener-level data, can further explore the underlying forces responsible for these cross-demand effects.

Our back-of-the-envelope calculations indicate that UMG can make the case for additional compensation, because of the usage of its music on TikTok, of approximately $206 million, with an estimated range between $85 million and $352 million (beyond the $111 million it was already earning). This assessment does not account for potential added compensation on other platforms, like YouTube, Apple Music, and SoundCloud. Notably, on May 1, 2024, UMG and TikTok reached a new licensing agreement that promised to “improve remuneration for UMG’s songwriters and artists” (Universal Music Group 2024c), aligning with our findings.

In closing, we note that our work has significant managerial implications for a number of key stakeholders. For music labels and copyright-protected content owners, our analysis suggests that social media and digital media consumption platforms often function as “substitutes,” especially for the content available on the social media platform (which also tends to be more popular). In contrast, for content not available on the social media platform (which tends to be less popular), they can serve as “complements.” This asymmetry suggests that labels and content owners may benefit from differentiated compensation models with social media platforms like TikTok. For example, compensation could be tied to the profile of the track or artist, with higher payments for established hits and more flexible terms for emerging artists. Alternatively, payment structures could be tied to engagement metrics of the posted creations leveraging the track, such as views, or to revenue directly generated by track usage on social media platforms, thus aligning compensation more closely with the platform’s actual value creation and the economic pie cocreated by both parties. For social media platforms, our analysis suggests that investment in tools for better discovery and promotion of music tracks can mitigate substitution concerns and strengthen their appeal. Another direction could be the development of analytics dashboards that quantify the downstream impact of music exposure, such as lifts in search or streaming by track type. By offering transparency into these spillover effects, social media platforms can make a stronger case that they also drive meaningful value for the music industry. At the same time, platforms should also carefully consider the revenue implications of music usage. With the rise of short-form video formats, platforms like TikTok are becoming increasingly important advertising channels, where music plays an essential role in content virality and engagement. Hence, any compensation model shared with music copyrights holders (e.g., UMG or other labels) may need to account not only for engagement but also for advertising revenue generated from music-driven video content. For artists, our study underscores the strategic importance of track selection and release timing across platforms. Artists can actively choose which tracks to feature on social media platforms to minimize substitution and enhance discovery and promotion opportunities. Additionally, they can consider staggering releases across social media and streaming platforms to create a promotional ramp-up effect, amplifying momentum and engagement across platforms.

More broadly, our findings shed light on the strategic interdependence between platforms. The observed spillover effects suggest that consumption on social media platforms can help or hurt consumption on music streaming platforms. As digital ecosystems grow more interconnected, this also has implications for public policy. Regulators should thus be attuned to these cross-platform dynamics when assessing issues related to competition, compensation, and market power behavior, particularly given the dominance of a few key players in both industries.