3 How Does Spotify Package Music?
“Dear Spotify, sometimes I don’t know what to listen to.” We know what you mean. Sure, having millions of songs in your pocket has changed the way you listen to music, but can you figure out what to play on a lazy Sunday afternoon?1
In 2013, those were the words with which Spotify introduced a “new and entirely personal way of discovering music,” which aimed to predict user preferences based on previous listening. Followed by the launch of “expert playlists for every mood and moment” a few months later,2 the new features pointed to a shift in the company’s way of attending to both music and its users. Period E—as this phase is referred to in chapter 1—meant a turn toward algorithmic and human-curated recommendations, through which the company promised to offer a personalized experience with “music for everyone.” The emphasis on recommendations illustrates how Spotify, similar to other digital content providers, not only delivers music but also actively frames and shapes data. The service thereby promotes certain values and identities over others, with music files being contextualized in a range of different ways: through playlists and other classificatory systems, through visual and textual elements of the interface, and through recommendations delivered to particular groups of users. Such operations are central for turning digital music into goods, but they also constitute a politics of content through which the delivery of music implicates prescriptive notions of the streaming user.
The integration of music files into larger value systems is dependent on two elements: the curatorial processes designed to be carried out behind the scenes (as described in chapter 2) and the interface through which users engage with the perceptible features and materialized effects of these processes. While the previous chapter was centered on the back end and various technical systems beneath the surface of Spotify’s desktop client, this chapter is concerned with Spotify’s front end. It engages primarily with the level of the interface as part of a larger apparatus that shapes digital music delivery and users’ modes of relating to musical content in the forms of various “expert curations.”
While scholars such as Nick Montfort and Matthew Kirschenbaum have cautioned against one-sided “screen essentialism[s]” that privilege visual aspects of digital media at the expense of material analysis,3 any attempt at uncovering the workings and effects of data contextualization must take into account both media-specific properties and representational content. As Johanna Drucker argues, graphical interfaces are in themselves “zone[s] of affordances” that organize data in particular ways and thereby foreground some things rather than others.4 In this sense, the interplays between users and interfaces—and thus the sprawling system of human and computational actors underlying them—come to constitute particular realities and ways of being.5
By drawing on playlists and album recommendations that were observed and collected between 2015 and 2017, this chapter sets out to discuss the inherently political aspects of Spotify’s recommendation features and how they (re)instate certain notions of music, of streaming, and of the ideal user. More specifically, the chapter builds on data from a number of case studies in which both manual data collection and bots were deployed to retrieve information about recommendation features as seen from a user angle. The bots were programmed to play particular songs and then capture the response in terms of generated recommendations. Together with manual observations and documentation of 224 playlists from Spotify’s mood- and activity-based categories, this has enabled us to perform close readings of visual representations and interface elements while also analyzing processes of personalization in music deliveries.6 Contextualized through critical analysis of promotional materials, journalistic accounts, and online user discussions, the collected data has provided us with an understanding of the performative effects of Spotify’s way of packaging music.
What’s in a Playlist?
It is often claimed that the digitization of music distribution involved a disaggregation of the otherwise dominant form of music commodity: the album.7 Relatedly, the popularization of file sharing by Napster around the turn of the millennium is typically mentioned as the point at which individual tracks were set loose from the album format, lending themselves instead to be mixed and matched according to the preferences of their listeners. Early media player software such as Winamp provided functionality for reaggregating tracks into customized playlists—an approach to music that built on previous assembling practices and technologies, such as broadcast radio, compilation albums, and mixtapes.8 Playlists thus have a longer history but emerged in their digital form as a way of recommodifying the individual tracks of the disaggregated album. Now a staple of every streaming media service, they are, to cite Jeremy Wade Morris, metacommodities that “rewrap individual commodities into a bundle under the assumption that the new whole is greater than the sum of its old parts and that another new whole is only a recombination away.”9
The streaming metaphor itself implies a continuous flow of music, reminiscent of a never-ending playlist—and playlists have indeed been one of the main devices by which users have created and managed their personal music collections on Spotify since its launch. As we have charted in detail in chapter 1, Spotify was initially centered around a search-based interface where users would have to enter keywords to find their desired music, unless they had imported their locally stored MP3 files. Tracks could then be added to a personal library or dragged and dropped into sharable and collaborative playlists. In other words, a notion of sociality was built into the service early on, but it has—somewhat paradoxically—come to be backgrounded through the service’s continuous transformation toward an increasingly personalized experience. For the purposes of this chapter, the single most important change to Spotify was the so-called curatorial turn, in 2013, from a search-based interface focused on simply accessing music to its current emphasis on delivering crafted music recommendations.
Introducing a range of prepackaged lists for different tastes and occasions, Spotify then consolidated the playlist’s status as a privileged object of its streaming universe and also reframed the playlist from being a primarily social and interactive element to being an object of editorial and algorithmic expertise.10 Both the search box and the recommendation deliveries can be understood as ways of managing overabundance in an archive so vast that it makes other browsing practices—such as shuffling through the entire collection—impossible. Importantly, however, they also represent two quite different notions of the streaming user.11 As we have argued elsewhere, the former reliance on the search engine made music consumption “dependent on a conscious reflection on the part of the listener, who is forced to ask herself: ‘what music do I want to listen to right now?’” This upholds the fictions of individual taste and consumer choice. In contrast to this positioning of streaming listeners as knowledgeable consumers, the reorganization toward a recommendation-based interface—still a form of music navigation largely “characterized by the centrality of text (and image) rather than sound”—imagined the listener as being in urgent need of musical advice and guidance from experts.12
More specifically, the 2013 turn toward tailored music recommendations, with Spotify being a producer of “entirely personal” musical experiences, was materialized in the (now slightly rearranged) Discover, Browse, and Follow features. Together, these functionalities allowed users to follow their favorite artists and tastemakers and to check out selected music based on previous listening patterns and editorial decisions.13 Through such collaboratively filtered album recommendations and prepackaged playlists, Spotify now took on a new role of delivering music “for every moment,” thus increasing the service’s reliance on both curatorial expertise and algorithmic systems. The latter, however, was typically downplayed in the initial promotion of the new features, as Spotify attempted to invoke a relationship of trust with users: “We’ll be your new best friend and offer handpicked recommendations that we think you’ll love. We’ll give you a full explanation of why we’ve picked each song too—like any best friend should.”14 Moreover, as discussed by Rob Drew, the trope of the mixtape added a nostalgic and affective value to this supposedly personalized mode of music streaming.15 Rather than dedicated mixtapes, however, the computational approximations of music preferences and ready-made playlists can be seen to reference commercial mixes, such as compilation albums, as markers of “an increased level of commercialization of the music industry,” in which business-related aspects are more important than the creative process.16
At present, the Spotify desktop client allows a user to access musical content via four different frames: through the search box in the banner; the menu bar to the left, which contains the Browse and Radio options, as well as links to the user’s saved and favorited content; the feed of followers’ streams to the right; and the main content frame in the center, which presents search results or automatically selected playlists. The player is found in the footer. For Spotify Free accounts, the otherwise sleek interface typically has a banner area that contains advertisements. Figure 3.1 offers a vivid depiction of the sponsored promotion of artist Selena Gomez, which at first glance is not easily distinguished from the other, Spotify-promoted musical content.17
Home page of the Spotify desktop client, as seen from a US listener account in May 2017.
The Browse tab is open by default when logging in to the service, and the home page displays a selection of twelve Featured Playlists, together with a short and typically cheerful greeting: “Have a great day!”; “Focus with Your Favorite Coffee!”; or “New week, new opportunities!” The message and the selected playlists are refreshed several times each day and also vary between countries.18 As seen in figure 3.1, playlists are visually represented in the form of square-shaped icons—a form of remediated album covers—and they are further framed by catchy titles and short descriptions that add affective value and context to their usage and users.
From the home page, users can move on to explore top charts, new releases, local concerts (based on their internet protocol, or IP, address), and—of special interest in this chapter—the so-called Discover and Genres & Moods features. Discover includes album recommendations based on the user’s recent streaming, as well as a weekly updated playlist with supposedly new musical discoveries. Genres & Moods presents a vast number of what seem to be (at least partially) human-curated playlists, divided into a number of categories.19 These represent both conventional music genres (typically broad ones, such as pop, electronic/dance, and hip hop) and generic moods and activities (such as Sleep, Travel, and Chill).
While Spotify provides access to a countless number of public playlists shared by individual users as well as independent curators (e.g., Indiemono, Soundplate, All Things Go), such user-created playlists can only be found through active searches. The playlists that are visually foregrounded in the interface—as Featured Playlists or Genres & Moods playlists—are the official ones compiled either by Spotify’s team of curators or by playlist brands connected to Spotify partners, such as Filtr (Sony), Digster (Universal), and Topsify (Warner Music).20 Not only are the in-house playlists thematically tailored to match advertisers’ potential target groups, they can also be sponsored by advertising clients.21 Moreover, as musical discovery through playlists is a prominent selling point for Spotify, playlists work as promotional devices for record labels and musicians. Because curation “has become a neutralised marketing term for taste-making and gatekeeping,”22 the selection and inclusion of specific artists on Spotify-curated playlists—some of them with millions of followers—have enormous effects for building a fan base and for increasing the number of streams and generating more revenue.23 Consequently, record labels work heavily with playlist promotion, aggregators such as CD Baby advise their members on how to best reach out to curators, and online discussions abound among independent artists on how to get their music on popular playlists.24 Meanwhile, Spotify keeps asserting the independence of its in-house content curators.25
In addition to being an important avenue for music marketing, the promotion of prepackaged playlists—as well as specific albums—should be acknowledged as a value-laden practice that is constitutive of certain subjectivities. The emphasis on delivering recommendations in the Spotify interface forges an asymmetrical relation between casual users and musical experts, with the former being understood as requiring assistance to find their way through the overwhelming collections. The implicit and explicit interactivity of the service—such as the possibility for users to share and jointly work on personal playlists, or the collaborative filtering of listening habits on which personalized recommendations are based—suggest that any user could potentially become a tastemaker, knowingly or not. The favoring of Spotify’s official playlists by way of the interface design, however, still insists on the significance of expertise in selecting the right, “handpicked” music for every user.
Soundtracking the Lives of Happy Subjects
“You can soundtrack your entire life with Spotify. Whatever you’re doing or feeling, we’ve got the music to make it better.”26 That was the bold declaration on the webpage of Spotify Running, a recently retired mobile functionality that used tempo detection to match “songs you love to the tempo you’re running.”27 The Running feature provided an illustrative example of Spotify’s attempts at selling a “branded musical experience” oriented toward personal improvement, something that can be seen throughout the service and especially in the promotion of ready-made playlists.28 In a previous publication, we explored the ideals promoted through the rotation of Featured Playlists and their accompanying greeting messages on the home page.29 This content category is what users first meet when they sign in to the service, and it provides thematically selected playlists and messages for mornings, afternoons, evenings, and weekends. Because it is refreshed several times a day, its content is essential to how the Spotify client produces a sense of real time–ness through filtering streams by time of day.30 As we demonstrated in our previous study, the particular notion of temporality presented here is also bound up with chrono-normative prescriptions of “the good life” that instruct users to get out of bed, go to work (in an office), work out in the afternoon, and then socialize with friends, family, and lovers in the evening. Meanwhile, music is presented as a way of increasing productivity and performance in these time-bound activities.31
However, the larger collection of Genres & Moods playlists, from which Featured Playlists seem to be retrieved, is not dayparted in the same way. It is currently organized into roughly forty different genre-based and thematic music categories, the exact content of which varies slightly over time and between geographic places. For example, in the spring of 2017, a snapshot comparison showed that some genres were provided to users in only one or a few countries (for example, Dansband in Sweden and Cantopop, Brazilian Music, and Mandopop in Singapore and Hong Kong). A Christian music category was present in many parts of the world but absent in European countries. Latin music, while present in all of the locations we explored, was typically found toward the bottom of the page—except for users registered in Spain and Mexico, to whom it was suggested among the top choices.32 Judging from language and musical content, a substantial number of the specific playlists found in each category were likewise based on a user’s country. Unsurprisingly, this illustrates the significance of local music cultures in shaping the appearance of the service, something that arguably could also relate to copyright and licensing issues.
In addition to presenting content according to a conventional, genre-based structure, Genres & Moods also includes ten playlist categories that are specifically related to activities and mind-sets: Mood, Party, Chill, Workout, Focus, Dinner, Sleep, Travel, Romance, and Kids & Family. Some of these categories, as well as their actual playlists, can be traced back to Tunigo—the playlist app discussed in chapter 1—which was acquired by Spotify right before its shift toward a recommendation-based service in 2013. Mundane as it may seem, this taxonomy is noteworthy in its attempt to schematize every aspect of daily life, not least because the music categories correspond to target activities offered to advertising clients.33
Looking at the short descriptions that accompany each playlist in these activity-related categories, most lists are promoted as functional tools for accomplishing a task or reaching a certain state of mind. In our manually collected data, for instance, the Workout category included a “Motivation Mix” playlist (with 1.5 million followers at the time), which was full of “Upbeat songs that keep you motivated while doing your cardio.” The Focus category held the even more popular “Deep Focus,” which encourages users to “Keep calm and focus. This playlist has some great, atmospheric rock to help you relax and concentrate.” The Mood category also included more specifically targeted lists, such as “The Writer’s Playlist”: “Suffering from writer’s block? This collection of writerly and literature-inspired songs might help!”
The use of music as a functional device has a long history, especially in terms of productivity requirements in workplaces and the exercise of and resistance to power more broadly.34 However, while the idea that music can be used to control one’s body and mind is not new, the mode of “ubiquitous listening” facilitated by streaming services seems to correlate with a broader turn toward a utilitarian approach to music, whereby music consumption is increasingly understood as situational and functional for certain activities (rather than, for instance, a matter of identity work or an aesthetic experience).35 This shift is evident not only in Spotify’s classification scheme (figure 3.2) but also in other features delivered by the service. The previously mentioned Spotify Running is one example, as well as the short-lived Party feature, which included a mood slider that allowed users to indicate their desired music energy level.36 More recently, Spotify has called out to users to “Show your mind some love” with new partner Headspace, an app for voice-guided meditations that will “help you feel happier, healthier, and more confident.”37 Whereas these examples suggest that music streaming and listening should be used for utilitarian purposes, they also privilege specific ways of thinking, feeling, and acting. In particular, they insist on self-governance through mood control, which can also be seen in how the Mood and Chill categories are found among the top music categories in all our explored countries, and how supposedly productivity-enhancing playlists are collected under the mood-related label “Focus” rather than, say, “Work.”
Spotify’s top eight Genres & Moods categories, as seen from a Swedish account in May 2017.
Mood management is thus a central theme of Spotify’s way of delivering music, as we have also noted elsewhere.38 The same has been said in relation to other music streaming services: Jeremy Wade Morris and Devon Powers, for example, conclude that emphasis on the affective dimensions of music consumption is one of the key components of a shift toward a branded musical experience.39 In a similar vein, Paul Allen Anderson discusses how streaming services increasingly work to create musical moodscapes in the tradition of older mood-delivery systems such as Muzak. Music recommendations, then, can be understood as products for mood enhancement and the management of psychological capital. According to Anderson, “Neo-Muzak services encourage listeners as never before to explicitly play with moods, to try out their tonalities and colorings at work, home, and anywhere with network reception.”40 However, complementing Morris and Powers’s broad discussion of mood management, and in contrast to Anderson’s emphasis on the circulation of and play with different moods, we want to emphasize the singularity of “Mood,” as the category is labeled on the Spotify interface. The branded musical experience delivered by Spotify does not seem to involve much playing around with different moods. Instead, it evokes fantasies of one specific state of mind and the moral values that come with it: happiness.
By happiness, we refer to contemporary formations of emotional wellbeing, as well as their improvement through techniques of the self that stem from popular positive psychology. The discourse of positive psychology promotes the idea that happiness is a result of one’s cognitive outlook, as well as “a task, a regimen, a daily undertaking in which the individual produces positive emotional states just as a fitness guru might shape a desired muscle group.”41 Looking specifically at the presentation of top playlists delivered in the Mood category of our data, these are almost exclusively devoted to the maximization of positively charged emotions, with descriptions such as the following:
- “Set it off with these epic anthems. Only good vibes here!” (“Good Vibes”)
- “The perfectly brewed cup, the perfect songs to match. Your happy place is right here.” (“Coffee + Chill”)
- “Hits to boost your mood and fill you with happiness.” (“Happy Hits!”)
- “Let your worries and cares slip away …” (“Relax & Unwind”)
Affective responses to and connotations of music are, of course, highly subjective. The same songs can be included in differently themed playlists, and the same playlists are also found in different music categories. However, our interest lies not with the subjective experience of musical content but with the universalizing and homogenizing aspects of its packaging: “Need a burst of inspiration?” “Need to cheer up?” “Feeling tired?” Met with such rhetorical questions in playlist descriptions, users are repeatedly called upon to cultivate their optimism by using music to manipulate or cure their thoughts and behaviors. The promotion of music as a technique for achieving the (most likely unattainable) ideal of happiness is reflected not only in calls to feel “happy-go-lucky to a lush and glossy mix of bouncy and peppy tunes. Stay positive—life is good!” but also in how the service offers help to overcome difficulties in times of emotional hardship:
- “Breathe deep and release that pressure right now with The Stress Buster.”
- “When someone you love becomes a memory … find solace in these songs.” (“Coping with Loss”)
- “The best cure for a broken heart or lost love. Angry or sad songs. You never know when you need comfort for a heart break—breakup songs is here for you!” (“Breakup Songs”)
The mood-based recommendations simply promise to provide the conditions necessary for fostering happiness—even in times of utter darkness.42
In this sense, the positive thinking characteristic of prepackaged playlists is intimately tied to the privileging of “entrepreneurial subjectivity,” as users are encouraged to direct their desire for change inwards and “capably manage difficulties and hide injuries.”43 Spotify, however, also offers a small number of playlists that provide a corrective to the happiness imperative, such as the “Life Sucks” list (“Feeling like everything just plain sucks? We’ve all been there. These songs will probably only make you feel worse, but at least they’ll let you know you’re not alone”)—or “Down in the Dumps” (“It’s a horrible day and nothing can change your mind”). These rare examples of negative sentiment open up spaces for reflection on the compulsory positivity of the interface, precisely because of their otherness in this context.
The general self-help ethos and cheerfulness of playlist descriptions is echoed in the graphical display of playlists, as seen in the (mostly) brightly colored covers of the top Mood playlists. The square-shaped cover icons conform to a visual aesthetic that combines the conventions of stock photography with traits of what Lev Manovich terms “Instagramism.” The concept refers to “the aesthetic of the new global digital youth class that emerge[d] in early 2010s,” which overlaps in many ways with the currently dominant commercial aesthetic.44 In addition to the square format, other visual characteristics of Instagramism present in the playlist covers include the privileging of flat rather than deep space and bodies cut by frame rather than symmetrically arranged full figures. This visual form resonates with the lack of denotative excess and contextualization in commercial stock photography,45 while at the same time constituting what Manovich describes as an “urban/hipster sensibility” by staging unique moments, feelings, and states of being. Props such as bubble gum, sunglasses, or a jetty invoke relaxed and joyful atmospheres through which individuals are turned into cultural stereotypes of the young and happy middle class.
Figure 3.3
Spotify’s top eight Mood playlists, as seen from a US account in June 2017.
Such visual elements also come to constitute gender and gender relations in specific ways. The visual materializations of curated playlists tend to foreground people presenting as women, especially in the Mood and Chill categories: laughing women, smiling women, bubble gum–chewing women, relaxing women, dreamy-looking women in nature. While the use of women as playlist-marketing instruments might be understood as building on historical gender conventions in advertising, it could also be considered in light of one of our previous case studies. There, we found that an overwhelming majority of Spotify’s recommended artists were male, suggesting that Spotify’s curatorial authority is deployed in ways that maintain male privilege in the music industries.46 Pitted against the visual aesthetic of playlists, this indicates that while the service reproduces an often-criticized notion of music production as a domain of masculinity, music consumption—especially for the sake of mood management—is portrayed as a female undertaking. The alluring promise of happiness and positive thinking gleaned from Spotify’s mood boards can thus be seen as reproducing a gendered form of neoliberal subjectivity, where young women in particular are invited to identify themselves as entrepreneurial subjects and to embrace the values associated with “consumption, self-transformation and notions of choice.”47 Possibly reflecting the entrepreneurial mind-set demanded of contemporary cultural workers, as discussed by Rosalind Gill, it seems to favor myths of egalitarianism and individual achievement while disavowing structural power relations.48
While implicit modes of governance through mood management and calls for self-enhancement are pervasive throughout the service, it should be noted—as we discussed in “Intervention: The Swedish Unicorn”—that Spotify has also been known to occasionally raise traditional and overt political issues. For instance, our collected data included the somewhat controversial “Refugee Playlist,” which was published in early 2017 in response to Donald Trump’s travel ban.49 In 2016, Spotify published the “Black Lives Matter” playlist, which was removed shortly after its publication, possibly due to fierce criticism.50 Furthermore, during the US elections in 2016, Spotify launched “Clarify,” a podcast targeted at young voters.51
The causes supported in these examples clearly target a specific audience and implicitly position the ideal user as a millennial with progressive values. And while the individual cases might be seen as providing counterpoints to the focus on interiorization in the construction of a happy, entrepreneurial subject, they are at the same time obvious examples of how a media company capitalizes on social injustice and politically charged events. In a similar vein, Spotify presents its updated “Feminist Friday” playlist every week, and the genre category WHM (Women’s History Month)—“celebrating women in music and culture”—was highlighted in the spring of 2017 for Swedish users. These content categories serve to emphasize the role of women in music production while simultaneously singling out female artists as the other to the male norm. In contrast to the visual association of women with mood management, such explicit references to gender might be said to illustrate the new “post-postfeminist” sensibilities of mainstream media, in which a commercialized version of feminism is constructed as “a desirable, stylish, and decidedly fashionable” identity.52
The Unruliness of Algorithmic Profiling
Our analysis of Spotify’s ready-made playlists points to the ways in which Spotify’s packaging of music comprises elements of gendered consumerism, individualism, and psychologism. The promise of happiness, in short, often invokes the ideal of a self-governing subject (at least potentially) in control of their inner life and social circumstances, so long as they stream the right playlists, with the right attitude. However, user input may also affect what content is being delivered, adding further layers of meaning to the construction of the user and to ways of packaging and conceptualizing music.
Alongside the supposedly human-curated playlists of Genres & Moods, Spotify’s Browse menu provides access to the Discover feature, one of the service’s main selling points and part of what marked its transition in 2013. The feature has changed over the years, from the initial display of a collage-like presentation of different musical objects (songs, playlists, albums, apps) to today’s streamlined organization of content in strips, including the weekly updated playlists “Discover Weekly” and “Release Radar,” as well as album recommendations framed as “Top Recommendations for You,” “New Releases for You,” “Because You Listened to …,” “Similar to …,” and “Suggested for You Based on …” These tailored recommendations are supposedly based on their resemblance to the user’s recently streamed artists, calculated through comparison with other users who appear to have similar taste.53 Thus, they “organize individuals into collective forms,”54 and the sheer number of Spotify users (140 million) provides a fertile ground for such collaborative filtering operations.
Based on accumulated data and feedback loops between recommendation deliveries and users’ actions, algorithmic systems work prescriptively; they attempt to predict user preferences and therefore also tend to shape user practices. The murky workings of proprietary algorithms have spawned scholarly interest over the last decade, and a large body of work now discusses the ways in which algorithmic systems are constitutive of taste, knowledge, and cultural practices.55 Software has performative capacities, as algorithms may attribute meaning to patterns of user behavior. Scholars have demonstrated how algorithmic content delivery has implications for the production of gender, race, and other categorizations.56 Building on Max Weber’s and Erving Goffman’s notions of the “ideal type,” John Cheney-Lippold argues that algorithmic identity can be thought of in terms of “measurable types”—that is, sets of observed data patterns that produce norms against which new user data can be compared and by which users can be categorized. However, as Cheney-Lippold stresses, such measurable types do not necessarily correspond with our nondatafied self-identifications or sociopolitical identities. Instead, they form constantly refreshed classifications that are usually inaccessible to ordinary users—classifications that act as modes of soft biopolitics by defining and governing large populations.57
While Spotify’s recommendations are based precisely on this type of continuously updated data, we cannot know exactly how user behaviors are classified and sorted into normative measurable types. What we do know is that users are invited—or obliged—to have their listening habits turned into “taste profiles,” which record clusters of preferred artists and genres, as well as affinity scores that measure how heavily, actively, and regularly those artists and genres are played and how much is streamed from an artist’s full catalog of music.58 Taste profiles are “mapped against wider ‘cultural knowledge’ about how those artists are described online, and the characteristics of their music” (such as “mainstreamness,” “freshness,” “diversity,” and “hotness”), as well as against the profiles and listening patterns of other users.59 Thus, each user becomes part of a larger collective intelligence from which recommendations are derived. As Matthew Ogle, product manager of the “Discover Weekly” playlist, puts it: “It’s still humans who are doing the song selection and arranging, but instead of outside experts, it’s users like you and me.”60 Currently, these taste profiles are not revealed to users, although there have been requests to make them public, resettable, or exportable between services.61
What is not known, then, is the precise content of these profiles and the wider cultural knowledge against which they are measured. Neither do we know whether or how traditional demographics factor into these calculations. Dynamic behavioral data, determined by input, is typically understood as more relevant for marketing purposes or content delivery than strict demographic categories. However, in a series of blog posts, Paul Lamere (The Echo Nest) claimed that certain artists are in fact skewed toward listeners of one gender or a certain age group.62 More specifically, Lamere argued that “for mainstream listening about 30% of the artists in a typical male’s listening rotation won’t be found in a typical female listening rotation and vice versa.”63 Similarly, when comparing listeners by age, Lamere stressed that the artist overlap between thirteen-year-olds and sixty-four-year-olds was only about 35 percent. Lamere’s blog posts also highlighted that streaming service users aged twenty-four to thirty-five played more music and listened to more artists than younger and older users. A similar argument has been made by Ajay Kalia, Spotify’s product owner for taste profiles, who has stated in interviews that, as people get older, they tend to stop keeping up with popular music. By their mid-thirties, users’ tastes have matured into a “taste freeze.”64 Despite these patterns, both Lamere and Kalia stress that demography is not as useful as actual listening patterns for predicting preferences, but it is still not clear whether or to what extent it plays a part in Spotify’s recommendation deliveries.
What is evident, however, is that before creating a new account, Spotify requires every user to self-identify their gender and age (via their birthdate). The service also detects a user’s location based on their IP address, thereby assigning a national identity to each user. At a later stage, users can add details such as postcode, cell phone brand, and carrier to their profile. The drop-down menus of the mandatory gender and (to some extent) birthdate fields testify to the enforced selection of what Lisa Nakamura calls “menu-driven identities,” that is, a limited range of simplistic and mutually exclusive identity categories through which users have to make themselves known to a service.65
Menu-driven identifications are thus performative acts required for the constitution of intelligible users. The fact that they are mandatory and yet offer only a limited range of options, as Rena Bivens has shown, can be understood as symbolic violence—such as when a binary gender menu (as was included in the Spotify registration form for many years) forces nonbinary users to either misrepresent their gender or abstain from using the service.66 After many years of criticism and discussions in the user community, Spotify rolled out a third gender option (“non-binary”) during the fall of 2016—but only in select locations, including Sweden, the United Kingdom, the United States, Australia, and New Zealand.67 Arguably, the selection of countries relates to market-driven assumptions about cultural and political appropriateness in different national contexts. It illustrates the significance of geography as a main structuring device and geographic difference as a performative effect of the service.
The fact that this third option is only available in a few countries—and that gender registration, within or outside the binary, is still mandatory at the time of writing—indicates that gender is perceived as vital to the functioning of Spotify, at least for marketing purposes.68 In order to investigate whether demographic self-categorization matters for music delivery, in 2016 and early 2017, we set up two case studies using the previously described bots to explore the meaning of gender and age. The intention was not to reveal the secrets of Spotify’s proprietary algorithms—an undertaking obviously beyond the scope of this project—but to gain insights into how the service appears to different users and thus comes to constitute situated notions of music, taste, and listening.
Figure 3.4
Was the introduction of a nonbinary gender option in the fall of 2016 a token of progressiveness or market-driven appropriateness?
The results of the gender case have been reported elsewhere, and we will therefore mainly focus on some observations from the age case.69 Here, we created Spotify accounts for thirty-two bots that were identical in all respects—apart from their assigned birthdates. They were divided by birth year into four groups: 1924 (aged ninety-three), 1954 (aged sixty-three), 1984 (aged thirty-three), and 2004 (aged thirteen). Four sets of ten songs were selected, based on one of the Billboard charts for the genres Hot Country, Hot Latin, Kids, and Spotify Rewind (music from the 1960s and 1970s), respectively, and two bots of each age were assigned identical music. Each bot was instructed to play its assigned songs and document the albums presented as “Top Recommendations for You.” Data collection went on for six weeks, during which recommendations were retrieved on a daily basis. Focusing specifically on the recommended artists (and not albums), the data was analyzed on both the aggregated and individual bot levels. The rationale behind our methodology is discussed at greater length in the forthcoming intervention; suffice it to say, we took a special interest in two questions:
- To what extent could age-specific artist recommendations be identified?
- How many unique recommendations were given to bots in each music group?
All in all, we collected over twenty-one thousand recommendations, including 1,809 artists. In other words, many recommendations were repeated; for each individual bot, about 70 percent of the recommendations were duplicates that were offered several times during the course of data collection. The pace at which content changed seemed to match Spotify’s claim that Discover content is refreshed every three to five days.70 However, it should be noted that two of the music types took weeks longer than the others to generate recommendations: Kids and Spotify Rewind. Kids music is a special case in this context, as it is not supposed to affect recommendations. For Spotify Rewind, however, the reason why songs needed many streams to generate recommendations was possibly related to them being more diverse in terms of genre. Thus, the taste profiles of these bots might have been seen as less uniform and predictable. This might also explain the fact that Spotify Rewind bots were provided with the largest number of artists in their recommendations, compared to the other music types.
Generally, the proportion of what we termed “age-specific” artist recommendations was very low—only 0.3 to 2.0 percent of the total recommendations.71 There was, however, one exception: the pair of ninety-three-year-old bots in the Spotify Rewind group, for which such age-specific recommendations made up just over six percent of their total artist recommendations. Besides attesting to the multifarious musical landscape of this Billboard category, it indicated that there was something special about our old-timer bots, as they were clearly set apart from the others. Looking at the ways in which unique artist recommendations—that is, artists recommended to only one bot, regardless of age—were distributed among the Spotify Rewind age pairs, we found further indications of this status.
Figure 3.5
Spotify Rewind: unique artist recommendations as a percentage of each age group’s total artist recommendations.
The pairs with the oldest bots (ninety-three and sixty-three years) seemed to be provided with a much more diverse set of artists than the younger age pairs (thirty-three years and thirteen years).72 A similar pattern was seen in the country group, in which the pair of sixty-three-year-olds also received a larger and more varied set of artist recommendations. While the case study was not designed to produce generalizable results, at first glance, the data seemed to indicate that Spotify and its music deliveries—at least at this particular moment in time and for these particular users—attributed meaning to some taste profiles that reproduced the demographically based categories assigned at registration. On the one hand, the results could be taken as a suggestion that our older bots, supposedly suffering from taste freeze, needed special—that is, more—help to find music and thus received a greater number of recommended artists. On the other hand, we wondered whether the music defined as similar to the streamed Rewind songs was age skewed in Paul Lamere’s sense, so that there were simply more albums to offer older bots listening to this particular music style.73
Figure 3.6
Spotify Rewind: unique artist recommendations for individual bots as a percentage of each bot’s total artist recommendations.
However, when delving into the data on an individual level, what had initially looked like musical diversity in the pair of sixty-three-year-old bots turned out to be the result of only one of them having received a vast number of unique artists—more than twice the average in this music group.
This was in line with the observations from our previous gender experiment, in which a large majority of all bots within each music group were given almost identical artist recommendations (regardless of gender), while a few of them received a much larger and more varied set of recommendations—similar to our odd sixty-three-year-old.74 We decided to call the former mainstreamers and the latter outliers, although it is not immediately clear whether the mainstreamers were given more popular recommendations. From a user perspective, though, being positioned as an outlier means being attributed a more alternative taste, in contrast to the crowd of slightly more conformist and generic mainstream listeners.
Another outlier was also found among the sixty-three-year-old country listeners, and similar constructions of users seemed to be at work in the Kids music group. We included Kids music in the case study precisely because children’s music is supposed to be filtered out in algorithmic calculations of listening patterns. According to Matthew Ogle, Spotify “made the decision early on in testing that we wouldn’t include children’s music. A lot of Spotify users are parents who play Disney tunes for the kids. They do a pretty good job of telling their parents when a new soundtrack has come out, you don’t need us to help with that.”75 During the course of the case study, one of the eight bots in the Kids group (a sixty-three-year-old, again!) actually received a total of 127 different artist recommendations, while the seven other bots received none. This bot, too, was seemingly positioned as an outlier in a group where being mainstream meant not getting any recommendations at all. Notably, the recommended music included album tracks such as “Pokémon (Trap Remix),” music from the TV series Glee and Family Guy, and music from video games, as well as music by major artists Zara Larsson, Beyoncé, Bon Jovi, Katy Perry, and Christina Aguilera. While presumably aimed at a younger audience, it was still pretty far from the Kidz Bop songs that constituted the original playlist in the Kids group.
Obviously, the number of bots in the age case was too small to make any inferences about the general relationship between outlier status and specific age groups, but the results can provide a backdrop for considering how supposedly personalized recommendations are unruly and beyond the control of individual users—despite ostensibly being based on previous listening. In fact, the very notion of personalization might itself be implausible. As Cheney-Lippold notes, measurable types cannot automatically be mapped onto lived social and political identities, and it might be better to talk about “profilization” as the constantly shifting “intersections of categorical meaning that allow our data, but not necessarily us, to be ‘gendered,’ ‘raced’ and ‘classed.’”76 Users, then, are not governed in terms of individuals but as members of populations, or even clusters of correlations.77 Nevertheless, the results of algorithmic governing are likely to affect how individual users conceive of themselves, the service, and the content delivered. Looking at the way in which recommendations were distributed among users in both the gender and the age case, it would be tempting to describe the algorithmic consequences as somewhat random and arbitrary. While always matching the musical style that our bots had previously listened to, the large variations in content and in the number of recommendations surprised us—in quite the opposite way from how the Spotify Radio function tended to repeat itself and not respond to user interaction, as described in chapter 2. As seen from the angle of our bots, then, Spotify was obviously a different service for each particular user: scant and mainstream for some, and exploratory and dynamic for a few others.
The Politics of Personalization
Despite—or precisely because of—the ongoing profiling, the trope of personalization looms large in Spotify’s marketing. As users are drawn into a supposedly tailored universe that seeks to provide the right music for “everyone” and “every mood,” they are also encouraged to enter into an affective and intimate relationship with the service. This sense of intimacy is invoked partly by phatic cues in terms of cheerful imperatives and real-time exclamations on the home page of the desktop client (“Music for your afternoon,” “Get the week started with upbeat music,” “Afternoon Energy?”). Moreover, it is promoted by the provision of curated playlists meant to facilitate daily activities and mood management, thereby bringing up topics otherwise understood as belonging to the private sphere.78 The invocation of intimacy and affect can also be seen in how algorithmically calculated Discover recommendations are framed as being built on the shared collective knowledge of human users—where any user could be seen as a potential tastemaker—rather than on computational processes.79
These constructions of intimacy, we argue, form part of a politics of content that seek to map and shape the lives of streaming users. The selling point of Spotify is not necessarily music but music streaming framed as a deeply personal and intimate—even happiness-inducing—practice. More specifically, this chapter has shed light on how the organization and presentation of recommendations privilege certain ways of attending to music and, consequently, to oneself as a subject of contemporary digital media culture. There are at least two types of disciplining logics at play here. First, curated playlists seem to inscribe music streaming in a gendered discourse of positive psychology. While alternative points of identification were present in our collected data, music streaming at large was rendered intelligible through references to neoliberal and capitalist values of individualism, self-fashioning, and self-responsibility. The mode of packaging and (re)presenting music mostly served to reinforce the notion of the user as a happy, entrepreneurial subject—young, urban, middle-class. At the same time, happiness and an entrepreneurial ethos were promoted as the taken-for-granted ideals toward which users should strive.
Second, the profiling of users, on which the delivery of customized album recommendations is based, illustrates the other disciplining logic at play. The molding of users into types or taste profiles can itself be seen as an expression of the “soft biopolitics” that, according to Cheney-Lippold, regulate our lives without us being fully aware of it.80 While recommendations are marketed as based on a user’s previous streaming practices—and thus suggest a notion of user agency—our study indicated that users had limited control over some of the content designated as personal and recommended “for you.” In fact, our interrogation of parts of the Genres & Moods and Discover features showed that the apparent personalization of music delivery, brought about by a multitude of actors and algorithmic processes and materialized on the surface of Spotify’s interface, was not very personal at all. Instead, music delivery appeared overall to be structured around a combination of broad and universal claims—such as the call to deploy music in order to achieve happiness—and particularities that seem haphazard rather than personal, such as the inconsistently profiled recommendations we found in the age case.
To our knowledge, the features found on Spotify’s Discover tab are the only ones that are dependent on a user’s past streaming. However, the constant feedback loops of user data are invaluable to the company’s business strategy, even if this is not immediately visible to users. Playlists, for instance, occupy a central role in Spotify’s strategy for attracting advertisers. Targeting users either according to their moods and activities or to demographic categories ensures that advertisements can be delivered to particular groups of users at particular points in time, as will be discussed in our next chapter. Here, the trope of personalization encourages users to work on their future recommendations by streaming more music and inputting more data into the system. The functional and intimate framings of playlists insist that listeners share data not only about their streaming behavior but, implicitly, also about their state of mind at any given moment, which in turn generates revenue for the service. Hence, while the act of providing playlists for every moment is in itself constitutive of intimacy, this intimate relation is monetized at the very moment when users click play.