Managing by Data: Algorithmic Categories and Organizing

Tracking user listening behaviour

To offer personalized music recommendations, Last.fm needs to gather, store, order and categorize music data. Differently from previous systems of categorizing based on expert or industry classifications of artists and music styles, Last.fm operates by gathering data produced by the real-time music listening behaviour of users which it clusters and categorizes bottom-up, using big data and machine learning techniques. To personalize its music recommendations, Last.fm uses an item-based collaborative filtering recommender system which it extends and qualifies with a tagging system (Aggarwal, 2016). Item-based collaborative filtering is one of the oldest and most widely used recommendation systems. Collaborative filtering, as the name indicates, filters data from the behaviour of large groups to recommend relevant items to individual users. The system embeds the assumption that relying on real-time data of listening habits from a large music community instead of individual experts is more effective in finding new and relevant ways of categorizing music and filtering recommendations.

By listening to music on any online music player application, computer or portable device with an Audioscrobbler plugin or extension, users automatically submit their listening choices to Last.fm. Scrobbling creates and transmits to Last.fm a playback event every time a user listens to a track. Playback events, in turn, generate playcount data which are the counting of how many times a listening event occurs. Playcount data are the main data entity of the system, connecting an individual user to playback events and building the individual and collective history of user listening behaviour. Playcount data are deployed to generate personalized charts and recommendations.

As with many other data tracking technologies, there is no playcount data out there which Last.fm records or tracks. Rather, the platform creates a playcount every time it receives an event that meets specific design requirements and fits a certain predetermined data format. Put differently, tracking listening data is an organizing process that requires a number of rules and parameters which define all the characteristics the event needs to have in order to produce playcount data. For instance, a playback event needs to contain specific metadata such as artist name, track name and timestamp. Only if the signal contains such parameters is it recognized as playback event and encoded into listening data (playcount). ¹ The careful crafting of how listening is encoded into data reflects a long process of negotiation between the organization and its users and developers which entails more than just the making of inscription rules. Once recognized as listening data, playcounts, in fact, need to pass through numerous filters which are applied to standardize the heterogeneous data gathered from the 600 connected applications and devices. ²

Last.fm uses artist names (which are data submitted with playback events) as data buckets or pigeonholes. Artist names represent a decisive intermediate passage between the organization’s making of data out of selected and encoded events (playcounts) and the subsequent creation of wider and algorithmically derived categories (similar artists). In contexts such as the one Last.fm represents, artist names are no more than data objects created ad hoc by the system to contain the information it needs to automate the computation of recommendations. They become the key objects that function as placeholders within which the system stores the playcounts and other data (metadata) it extracts from the track-file. Although an artist name may exist already in the music field, the organization moulds it into an object which sustains its own data-driven approach to music listening.

Effectively, artist names already exist both offline as names of (real) artists and online as metadata or descriptions of music tracks. Yet, because there is no agreed-upon standard or institutional music identifier to name digitized music, the same music track can have different descriptions (different artist names). Out of an arguably uncontested entity such as the name of an artist, digitized music has created a multitude of entities (the many different versions of the same artist name such as nicknames, wrong spellings, etc.). Machines cannot yet identify misspellings as different versions of the same name and thus the Last.fm’s system automatically creates a new object out of any misspelling of a given artist’s name (see Figure 1). The system, for instance, produces as many objects as the different spellings of Louis Armstrong’s name, where ‘Louis Armstrong’, ‘Armstrong Louis’ and ‘L. Armstrong’ each creates an object in the system.

OPEN IN VIEWER

Figure 1.

Example of differently spelled tracks from Last.fm blog, see http://blog.last.fm/2007/09/10/fingerprinting-update.

This seemingly trivial problem creates data inconsistencies whose scale has always been one of the most challenging computational problems for the system. Last.fm has, over the years, implemented different strategies to select the correct version of an entry: user vote (so-called community-based solution of flagging incorrect entries), manual data cleansing and cleansing robots. Here is what a Last.fm engineer said on the issue:

Until recently, the system had no way of identifying variation in spelling of artist and tracks, which led to many duplicate pages for the same artist or track. With millions of scrobbles coming in everyday, it doesn’t take a genius to figure out that you will soon have a big metadata mess on your hands. ³

In 2009 Last.fm implemented Autocorrection, an automated system that identifies incorrect artist names and maps them to their correct version. Albeit automated, the task remains hard to conclusively address and is further complicated by the fact that new (correct and incorrect) artist names are constantly ingested to the system. This adds a real-time exponential complexity to an already quite intensive computational task ⁴ which requires constant updates from internal developers and continuous feedback from users.

From objects to similarities

A further step in the production of music personal recommendations is the establishment of similarity between artist names. Similarity in this context is not constructed on the basis of the intrinsic attributes of the artist-items grouped. In the system, artist names are basic objects created by counting and clustering user plays. In most essential respects, they are data objects. All that is known of an artist in these online contexts is how many times its name is associated with a user playback. It is on the basis of this link between user playback and artist name that the system establishes similarity between artists. Similarity measures to what degree two or more artists tend to be listened to together by two or more users over time. The system assumes that artists that tend to be listened to by the same users have something in common; it therefore uses the listening patterns of these users to rank artists and compute suggestions.

Once similarity between artist objects is established, similar artists are put into clusters or, as the jargon goes, neighbourhoods. Such clusters or neighbourhoods are categories of similar artists that work as groups of predictors (see e.g. Aggarwal, 2016). To rank how similar two artists are within a group of already similar artists, the system will select a measure of similarity (see Alaimo & Kallinikos, 2020). Recommendations are essentially predictions cast in the form of probabilities. To put it simply, a prediction to suggest Ella Fitzgerald will be computed on the basis of the ranking of her nearest neighbours (i.e. Louis Armstrong, Billie Holiday, etc.).

There are a number of limitations to this formal approach to categorization. Similarity in music discovery systems cannot distinguish genres of music, attributes or specific features of artists or bands. Given that similarity is just a score, based solely on listening data, it may very well determine two artists belonging to different genres as similar to one another and group them into the same neighbourhood. The only solution to this problem would seem to be to gather more data or to implement a hybrid approach to recommendation by adding a different recommender system (Alaimo & Kallinikos, 2020). Not surprisingly, Last.fm has blended its collaborative filtering by implementing a user-generated tag system. Users can tag by attributing any label to any track. Labels are keywords and can be genres, names, years, but also personal labels, based on experience or use such as for instance moods, occasion (i.e. dinner or party), memories (i.e. summer 2018) and so on. Tags seek to provide to the system the missing embedment into a meaningful structure of relations which the counting of listening data lacks. Tags were envisioned to offer an additional dimension to Last.fm’s music discovery system, as the statement by a Last.fm engineer illustrates:

[T]his diversity and eclectic view of the musical landscape is what Last.fm is all about. We don’t live in a cookie cutter world of hackneyed generic labels for music. Most music sites include the standard dozen genres (pop, rock, urban, etc.) and that doesn’t adequately describe the diversity of music out there. Our tags system encourages the weird and the wonderful, the micro-communities and new scenes that are springing up as fast as new, independent bands are formed. ⁵

The results, however, proved to be a mixed blessing. On the one hand, tagging delivered a staggering variety of labels that mixes mostly traditional genres and general attributes with functional value such as ‘Saturday night’ or subjective feelings such as ‘moody’. These miscellaneous labelling practices (tags) certainly brought in novelty and, effectively, Last.fm was one of the first platforms to experiment with suggestions based on its tags with its Discovery web app. ⁶ On the other hand, the cumulated four million tags have left gaps in identifying novel music characteristics. The tags turned out to be either too personal or altogether predictable. ‘Although you can make up any tag you like, we noticed that in practice most people use tags that describe genre, or closely related things such as the era, or nationality of an artist’, noted an engineer from the Music Information Retrieval team at Last.fm. ⁷ Additionally, the massive scale reached by tags and their intrinsic aggregate value required the development and implementation of new automated methods of analysis, a task that eventually turned out to be a big challenge for Last.fm.

The last step of this music recommendation cycle is the computation of prediction. Albeit merely an output of a complex organizational process which involves all the passages we have illustrated above, prediction is the ultimate purpose of recommender technologies. The system is designed to produce a form of knowledge that is not descriptive but predictive and, in the final analysis, performative. The service the organization provides, the design of the system, the encoding of events, the making of basic data objects and the algorithmic assembly of data to categories are constantly optimized to achieve efficient prediction outputs on a massive scale.

The making of data

Playcounts are the key data inputs through which the listening behaviour of users is registered, rendered legible and, ultimately, computable. The making of data that can serve as essential inputs for further operations requires some upfront organizational decisions, which are then embedded in the system. To implement such decisions further requires addressing a series of technical problems which we described in some detail in the preceding pages. Playcount is the count of a user’s plays that pass the filters, rules and parameters of Audioscrobbler. Data produced this way are therefore marks or inscriptions of user listening behaviour as this is engineered and subsequently tracked by the system. Playcount is the elementary and fundamental data unit of the system. Via the playcount, the listening of music undergoes a subtle but decisive transformation: it becomes a formal operation made of discrete data, engineered in specific formats and put together to build a computable object.

Data are often, and rather mistakenly, equated to facts out there. Yet, the connection of data to facts is very complex, often insidious (Bailey, Leonardi, & Barley, 2012; Kallinikos, 1999). In organizations such as Last.fm, the format of data is produced by the functional prerequisites of the systems in place (i.e. the Audioscrobbler with its music discovery functionalities). As we showed earlier, data are the outcome of a number of contingent processes and organizational negotiations (Jones et al., 2019) rather than the straightforward mapping of an area of reality out there, as the listening behaviour may suggest at first sight. If they represent facts, data often do so in a rather narrow sense. In some cases, such as those of simulation, data may entail strong correspondences or structural affinities to a reality out there (Bailey et al., 2012; Knorr-Cetina, 1999). More often than not, the specific formats of data and the formal language used to record them considerably determine the particular event that is recorded, stored and sorted out (Alaimo & Kallinikos, 2019; Bowker & Star, 1999; Flyverbom & Murray, 2018; Gitelman, 2013). An event is inscribed into data only because it fits the format of the digital token and possesses functional attributes such as metadata which are important to its interoperability and computability.

As such, playcounts have no value or meaning on their own. They are made valuable thanks to their volume and the function they perform in the broader system of knowing set up by the organization. Playcounts make listening behaviour countable and possible to enter into a series of permutations with other data. Making data thus entails an act of selection, interpretation and encoding that is at the same time an important organizational decision that conditions subsequent organizational operations, decisions or tasks. Last.fm’s system works on the basis of the organizational interpretation of what constitutes a music event (playback events) in online platform settings and, accordingly, negotiates data rules and parameters with users, developers and other stakeholders (e.g. music labels). At the same time, these negotiations are heavily limited by a number of other constraints such as the kind of technology employed, its functional prerequisites and other technical and industrial standards prevailing in the field.

Artist names as basic objects

Playcounts are aggregated into artist names. Artist names operate as basic objects, in the ways defined earlier in this paper, reducing the variability of the world (here playcounts) yet delivering entities concrete enough to aid perception and action (Rosch, 1975; Rosch et al., 1976). As basic objects, artist names allow the organization to perform and coordinate its main activity – that is, the encoding of music listening behaviour into data – with little or no concern for being representative of entrenched social practices linked to cultural processes of categorization. From its very beginning, the platform’s mission has been to innovate online music consumption by relying on automated methods of data management. The making of artist names out of playcounts does not draw from cultural music conventions (i.e. real artist names). It rather provides the system with the scalability and standardization of the data it needs to run accurate statistical parsing and offer recommendations of music out of user listening behaviour data.

Cast in this light, the engineering of basic objects in online settings presents some important differences to the socio-cultural making of basic objects. The balance, in particular, between abstraction and concreteness that is key in coordinating action and communication within and across organizations and fields (Bowker & Star, 1999; Rosch & Mervis, 1975; Rosch et al., 1976) is perturbed and shifted towards abstraction or, perhaps, more correctly, formality. The processes we describe displace the trade-off between concreteness and abstraction underlying basic objects with a vaguely recognizable cultural entity (artist name) that serves the principles of scalability, standardization and computational efficiency, and the cascade of automated operations that enact these principles (Couldry & Mejias, 2019). The flipside of having a music discovery system based on a data object such as artist name with vague socio-cultural roots and no real social context is the production of a tower of Babel, where it is no longer possible to distinguish a real artist name from a fabrication (Boast, Bravo, & Srinivasan, 2007; Ekbia, 2009; Iliadis, 2019). Up until 2009, Last.fm ran with the idea that massive volumes of artist names and the technology already in place would end up with the correct artist name. After different attempts at dealing with misspelled names and inaccurate entries with community flagging and other means, Last.fm implemented the autocorrection system, the automated mapping of non-standard artist names into correct ones. Algorithmic categorization does not admit conflicting rationalities and the only way of solving an issue caused by automation is by implementing further automation (Arthur, 2009; Ciborra, 2009; Kallinikos et al., 2013).

Artist names as basic objects provide the foundation upon which Last.fm derives higher-order categories, based mostly on similarity but also on other time-contingent measures such as popularity and trending. But artist names also operate as boundary objects of a very specific type. In many ways, artist names are the meeting point between Last.fm, users, developers and the platform ecosystem in which Audioscrobbler operates. The knowledge artist names embed allows and constrains the coordination of organizational activity across multiple contexts. Organic boundary objects are both abstract and concrete and have both local and shared meaning which need to be constantly negotiated and balanced out to maintain coherence across intersecting communities (Bowker & Star, 1999; Star, 2010). Artist names however are engineered boundary objects. While they enable massive collaboration across organizational and field boundaries, they at the same time lack the plasticity that would allow them to adapt and speak to different communities of practice. In the case of Last.fm this led to the proliferation of rules and parameters governing their production and use and ended up excluding data sources and membership categories (i.e. certain genres of music that do not fit the parameters of data ingestion, external developer practices) and triggered protests from users and developers. Ultimately, the rigid nature and formalism of data boundary objects affected music recommendation.

Categorization and user behaviour

Artist names qua basic objects provide the basic material out of which higher-order categories such as similar artists, popular artists and other time-bound clusters of artist names such as daily, weekly or monthly trends are derived. Similarity, in particular, is an axial principle and the backbone of all recommender systems that populate the online world of platform organizations. Categories of this sort activate the potential meaning of playcounts and represent a crucial technological step through which entities are related to one another via recommendations.

These recommender systems operate by sorting data objects into similar groups on the basis of the computation of patterns of linked occurrences. Louis Armstrong and Ella Fitzgerald are deemed similar because they are listened to (play counted) by the same users over time. In this regard, the activity music recommendation systems perform is functionally equivalent to categorization processes, insofar as they both treat non-identical objects or events as largely equivalent and sort them out, accordingly, into similar groups. Yet, the means and premises by which they achieve these goals are vastly different. In settings such as Last.fm, higher-order categories are the outcome of clustering and aggregating the objects and events such settings bring into being on the basis of formal operations rather than via resemblances or other intrinsic or practice-related affinities.

Once artist names are created and filtered by the autocorrection system, they need to be assigned into similar groups and mapped into artist similarity networks. In this process, similarity is defined as the step-by-step procedure of calculating the distance between two items (artists) according to how much the rating patterns (listening is unary rating) of two or more users agree. The category of similarity is thus established by computing, not by experience, cognition or knowledge. The results obtained are groups of artists that are made visible to users as similar and, importantly, acted upon as similar. Because of the similar artists category, the system is able to produce a number of derived ephemeral categories such as ‘Artists for you’, ‘Tags to explore’, ‘Tracks you need to hear’, ‘Blast from the past’, etc. which direct user listening behaviour and further update modes of online music consumption across the online music field. It is in this sense that recommender systems operate in a functionally equivalent mode to traditional categorization. Recommender systems deliver larger cognitive clusters, such as similar artists, yet they do so in ways that differ remarkably from how culturally embedded categorization schemes mediate the perception of reality.

Categorization and organization

Institutionally based classifications inevitably develop, Mary Douglas (1986) claimed, beyond community-based categorization schemes. The rising complexity of production (technical, material, economic) brings the establishment of new ways of labelling and classifying to accommodate the matrix of operations underlying this complexity. The path from community to expert or industry-based classifications is, of course, seldom linear and exhibits considerable empirical diversity. Douglas’s claim is nonetheless a good approximation of how the complexity of production and the advances in knowledge that underpin that complexity are linked to new schemes for ordering and categorizing work in organizations and across fields.

Our own findings indicate that these ideas are, mutatis mutandis, applicable in the case of Last.fm and, perhaps, more widely to how music experience (and consumption) is organized in online settings. Platform organizations, in particular, orchestrate the participation of highly dispersed and atomized populations of users which they seek to maintain and enlarge. At the two endpoints of the process we have analysed stand the rendition of user listening habits into data and the making of recommendations that steadily connect back to users. Both these fundamental operations require categorization of listening data to basic objects such as artist names and the derivation of higher-order categories such as similar artists, processes that entail a long and hidden data-work at a remove from the platform interface. The steady production of recommendations renders the atomized user platform populations as a sort of surrogate community, whereby users are made similar to one another on the basis of a computed similarity (Alaimo & Kallinikos, 2019). Current experience beyond Last.fm indicates that recommendations increasingly become native in online worlds, and algorithmic categories, such as similar artists, naturalized (Bowker & Star, 1999; Douglas, 1986).

Cast in this light, personalized recommendations make visible similar artists and music genres derived ad hoc while holding obscure the work of the apparatus, that is, the long journey of formal operations by which similarity is produced in platform organizations. Once re-contextualized and re-socialized, these categories emerge as a common currency that effectively coordinates user interaction online and the organization’s encounter with users and other stakeholders (e.g. developers, other platforms, artists). Categories of similar artists are navigated by users and naturalized by their listening behaviour which, in turn, reinforces further algorithmic categorization and aids system learning and optimization. In this regard, categorization redefines experience and long-entrenched cultural habits (Couldry & Mejias, 2019). The apparatus of data-technologies-algorithms in which these developments are embedded operates by sorting things out on behalf of users. Its computational outputs mingle, develop adjacently and, occasionally, take the place of existing socio-cognitive structures and cultural patterns by facilitating the processing of available categories and shaping the emergence of new ones.

Our investigation points out how these processes often require larger time scales to manifest, over which several tensions and conflicts are settled or addressed. Our analysis illustrates how these tensions and negotiations are considerably framed by technological functionalities and developments, the path-dependencies of the digital music field and, more generally, the performativity and self-reference of automation, whereby technologies and technological operations steadily beget new ones (Arthur, 2009; Kallinikos et al., 2013). Once this apparatus is at work, its operations become blackboxed and inaccessible to end users and, to a certain degree, developers and organizational members. The basic objects of categorization such as artist names embed machine-readable knowledge that facilitates automation but makes the objects and the process of categorization non-readable to humans. In this regard, algorithmic categorization does not, and cannot, rely on existing socio-cultural contexts while its operations remain at a remove from and opaque to human inspection, understanding and control. After all, the entire apparatus of data-technologies-algorithms is a technical response to the low commensurability, permutability and computability of cultural objects and categories (Alaimo & Kallinikos, 2017; Espeland & Sauder, 2007) which the apparatus reconstructs as data-based operations.

Paraphrasing Knorr-Cetina (1999, p. 40), we can claim that under the online conditions we have outlined in this paper, the social order is reconfigured as a data order. The idea is not entirely new and is encountered across the social sciences (e.g. Bailey et al., 2012; Borgmann, 1999; Foucault, 1970). In the field of organization studies, it can be traced back to Zuboff’s (1988) pioneering research on the computerization of work. Computerization, Zuboff foresaw, reconstitutes the order of organizations in the form of data tokens that bring about the far-reaching textualization of work. This electronic text, as she called it, is comprehensive and systemic, exhibiting considerable depth and existing independently of space and time. This text Zuboff claimed has no author in the conventional sense. It is just the patchwork of vaguely connected individual acts structured and mediated through automation and the power matrix from which it derives (Zuboff, 1988, pp. 179–81).

Never before have these ideas been as relevant as they are today. Recasting them in the current context of the web and the online environments of platform organizations certainly entails some important modifications (see e.g. Faraj et al., 2018; Flyverbom, 2019; Kornberger et al., 2017; Monteiro & Parmiggiani, 2019; Von Krogh, 2018). Our own research updates and extends these poignant observations by showing how the apparatus of data-technologies-algorithms encroaches upon and restructures the cultural exchanges of social actors. Organizations such as Last.fm become possible thanks to the codification of cultural habits and the remaking of music preferences by means of data production processes and the algorithmic categorization schemes they entail. On this account, technology emerges not simply as a mechanical force but also as a pervasive medium (Beverungen et al., 2019) that reorganizes the human sensorium (categories) and the ways social agents perceive and organize their world and their interactions with others (Rosch & Mervis, 1975; Rosch et al., 1976). The study of categories and categorization schemes offers in this regard an important avenue for confronting the challenge raised by the pervasive involvement of data and data management techniques in organizations and fields.

The tension between field-specific processes and decontextualized techniques of control is of course a pervasive theme in organization studies and, particularly, the branch of organization studies known as institutional theory. Field-specific processes often develop endogenously encoding experiential knowledge in a field, and other professional and organizational exigencies (Scott, 2008). By contrast, decontextualized techniques of control embody general principles of calculation and measurement that are largely field-independent (Fligstein, 1990; Kallinikos & Hasselbladh, 2009; Miller & O’Leary, 1987). While developing in particular social settings, such decontextualized techniques often diffuse across a variety of fields and organizational types (Abdelnour et al., 2017; Labatut et al., 2012), including public agencies (Bejerot & Hasselbladh, 2013) and also fields such as finance, travel and insurance (MacKenzie, 2006; Shiller, 2003). Our own study links to and, at the same time, renews the relevance of this type of research by bringing it to bear on organizational fields that are pervaded by data and the technologies and operations by which data are made to matter.

Technology and organization

We have so far shown how the apparatus of data-technologies-algorithms is linked to the establishment of objects and categories in organizations such as Last.fm. The data production process cycle we have unravelled (playcounts, artist objects, artist similarity networks, recommendations) is in most essential respects a process of categorization that is built on artist names as the basic objects and on the higher-order categories of similar (and popular) artists. Playcounts link Last.fm with an atomized population of users while basic objects qua boundary objects shape the interaction patterns between users, developers and the platform. Such findings address the first two of the questions we raised at the front end of this paper.

However, our third question concerning the wider organizational implications of these developments still requires further commentary. Organizations are broader social entities that embody much more than just technologically embedded tasks and processes. Since its very beginnings as a scientific field, organization theory has construed organizations in terms of structural attributes such as jobs and positions, authority lines, resource compounds and knowledge systems and capabilities or in terms of processes that describe the entanglement of resource flows and actions, including routines, sustaining the production of goods and services. Over the years, such structural and processual attributes of organizations have variously been linked to the core technologies they deploy (Hickson, Pugh, & Pheysey, 1969; Mintzberg, 1979; Noble, 2017). This understanding of technology’s involvement in organizations has implied a clear ontological separation between the technological and the social or organizational system as, for instance, in Mintzberg’s (1979) distinction between the operative core and the administrative superstructure or in Hickson et al.’s (1969) distinction of operations technology and organization structure (for a further and more recent discussion see Leonardi, 2012; Leonardi & Barley, 2010).

The underlying idea has always been to find out how structural, processual and other attributes of organizations are shaped by the technological system and its features, and vice versa. While historically studied in the context of industrial technologies, the conception of technology and organization in these terms has by and large continued to underlie the study of information technologies and their organizational implications. One can retrace that separation even in Zuboff’s (1988) work mentioned above and other studies on information technology and its link to social structure or processes (e.g. Kallinikos, 1999, 2007; Labatut et al., 2012; Munir & Jones, 2004; Orlikowski, 1992; Zammuto, Griffith, Majchrzak, Dougherty, & Faraj, 2007). Much current research on algorithms, AI and machine learning carries that tenor into our own age.

Organizations such as Last.fm and similar platform-like organizations are however different. In a sense, they collapse the difference between technology and organization. The standardization of the human interface and the automation of the backstage operations the apparatus brings about increasingly envelop the layout of organizational tasks and the processes with which they are associated and blur the distinction between humans and machines. The diffusion of smart and light technologies through which most users are currently linked to platforms and the web more widely makes the distinction fuzzier. Taken together, these developments reframe membership rules and the separation of organizations from their environments that many organizations, no matter how laboriously, have sought to maintain. Users are not employees nor are they simply customers. Critically, technologies are used in ways that reclaim the instrumentation of key organizational operations that have traditionally been performed by experts and work groups. The processes of categorization we have analysed in this paper and the process cycle with which it is linked is a case in point.

Much of what we describe in this paper demonstrates that the process of categorization is considerably hardwired to technologies and technological operations. At the same time, categorization is centrally linked to the production of data as a new and pervasive resource and medium characteristic of this hyper-technological age. Users are profiled on the basis of the music preferences they express through their clicking behaviour. The translation of cultural taste by other modes of expression (buying behaviour, social participation, income, class belongingness) to click data already marks a significant change of social modes of expression and identification and establishes an entirely new channel (data and their outputs) that links organizations to their surroundings. Once transformed to data, preferences undergo an alchemy of operations and can be added, subtracted, aggregated and computed to generate other descriptions (classifications) of the social such as similar artists or similar users. This is how similarity and also popularity and trending scores produced daily and amassed on these platforms work.

Cast in this light, the difference between one set of operations that could be described as social from that of the operative technological core (Mintzberg, 1979) is reduced and the two sides mingle and interlace with one another in ways that make their separation a challenging task. Paraphrasing McLuhan (1967; the medium is the message), one could go so far as to claim that the technology is the organization. This is certainly a hyperbole but a useful one that casts the current processes of organizing in a new light. It definitely calls for further research on these matters that can help unpack the empirical variability through which the fusion rather than simply the entanglement or imbrication of technology and organizations takes place (Leonardi, 2012; Leonardi, Nardi, & Kallinikos, 2012; Orlikowski, 2007; Von Krogh, 2018; Zammuto et al., 2007).