White Paper 2 of the Beyond the Playlist Series
Abstract
Spotify provides several mechanisms beyond the playlist for listeners to engage with the catalog in an exploratory mode: artist radio, song radio, album radio, artist mixes, and the autoplay continuation function that activates when a queued session ends. This paper examines these mechanisms in analytical detail, describing their observable behavioral patterns, explaining the underlying algorithmic logics that produce those patterns, and evaluating them as genuine discovery tools against the standard of musical exploration established in Paper 1. The central finding is that Spotify’s exploration tools represent a genuine advance beyond pure playlist culture — they do introduce novelty and do lead listeners beyond their immediate starting point — but they exhibit characteristic failure modes that limit their effectiveness as discovery infrastructure. These include genre gravity wells that pull extended sessions toward commercially dominant subgenres, recency weighting that systematically disadvantages catalog depth, popularity bias that makes genuinely marginal or subcultural music difficult to surface, novelty decay over the course of extended sessions, and differential behavior depending on whether a session is seeded from a major hit, a deep album cut, or an artist’s broader catalog. Understanding these failure modes is essential both for listeners who want to use these tools more strategically and for any serious evaluation of what streaming platforms would need to do differently to support genuine musical exploration.
1. Introduction: Beyond the Playlist’s Edge
The listener who reaches the end of a curated playlist, or who deliberately steps outside it in search of something less familiar, encounters a different layer of Spotify’s architecture: the set of tools loosely grouped under the radio and mix functions. These tools share a common premise — that the listener has expressed a preference by identifying a seed (an artist, a song, an album, or a genre), and that the platform should use that seed to generate an extended, theoretically open-ended sequence of music that radiates outward from the starting point.
This premise is more promising than the playlist model described in Paper 1. Whereas a playlist is a closed document — finished before the listening session begins — a radio or mix function is generative, producing a sequence in real time that can theoretically adapt to listener behavior and introduce genuine novelty. In principle, seeding a radio function from an unfamiliar artist or a deep album track should produce a pathway into unfamiliar territory. In practice, the pathway has characteristic shapes and characteristic dead ends that repay careful analysis.
This paper proceeds from direct behavioral observation of Spotify’s radio and mix functions across a range of seeding conditions, combined with analysis of what is publicly known about the algorithmic mechanisms underlying these functions. It documents the observable patterns, proposes explanatory frameworks for those patterns, and evaluates the results against the discovery standard articulated in Paper 1.
2. Spotify’s Exploration Architecture: A Functional Map
Before analyzing behavior, it is useful to map the distinct tools Spotify provides for non-playlist exploration, since they are not identical in their construction or their outputs.
Song Radio is generated from a single track seed. The platform uses that track as a reference point and constructs a queue of tracks that it determines to be related, drawing on a combination of audio feature analysis and collaborative filtering data. Song radio is the most tightly bounded of the exploration tools — its seed is highly specific, and its initial output tends to stay closest to the sonic and cultural neighborhood of the originating track.
Artist Radio is generated from an artist seed rather than a specific track. Because an artist’s catalog may span multiple periods, styles, and sonic territories, artist radio has more latitude than song radio to range across different parts of a musical tradition. In practice, however, this latitude is frequently constrained by the algorithm’s tendency to anchor to the artist’s most-streamed or commercially recognizable material rather than their full range.
Album Radio begins from an album as seed object. This is perhaps the most interesting of Spotify’s exploration tools for present purposes, because it implicitly treats the album as a meaningful unit — a coherent artistic statement with a particular sonic character — rather than dissolving it into its constituent tracks. Album radio tends to produce outputs that are tonally and stylistically more consistent than artist radio, because the album’s sequencing and production provide a denser and more specific set of audio signals for the algorithm to work from.
Artist Mix is a distinct function from artist radio, though the distinction is not always clearly communicated in the interface. Artist mixes are described by Spotify as personalized to the individual listener’s taste profile, combining tracks from the seed artist with tracks from artists that Spotify’s model associates with that listener’s listening history. This means that two listeners seeding the same artist mix will receive different outputs depending on their individual taste profiles — a personalization feature that has implications for the echo chamber problem examined later in this paper.
Autoplay is the continuation function that activates when a defined listening session — a playlist, album, or other queued content — ends. Rather than stopping, Spotify generates a continuation queue using the recently played content as a seed. Autoplay is the least deliberately chosen of the exploration modes — it operates on listeners who have not actively elected to explore — and its behavioral characteristics are worth examining separately from the radio functions that listeners consciously invoke.
3. Algorithmic Mechanisms: Collaborative Filtering and Audio Feature Matching
Spotify’s recommendation systems draw on two primary technical approaches, and understanding their interaction is essential to understanding the behavioral patterns the radio and mix functions exhibit.
Collaborative filtering is the older and more widely deployed approach in recommendation systems generally. In its basic form, it works by identifying listeners whose overall taste profile resembles the current listener’s profile and recommending music that those similar listeners have enjoyed. The underlying assumption is that taste similarity is a reliable proxy for recommendation relevance: if listener A and listener B have listened to many of the same artists and tracks, then music that B enjoys but A has not yet heard is a plausible recommendation for A.
Collaborative filtering is powerful at identifying music that is socially adjacent to a listener’s existing taste — that is, music that belongs to the same overlapping communities of listeners. It is structurally limited, however, in ways that matter for discovery. It cannot identify relevant music that lacks a substantial listening community — whether because it is new, obscure, or associated with a small subculture — because the data density required for reliable similarity matching is simply not available. It also tends to route recommendations toward cultural consensus rather than toward genuine novelty, because the listener populations whose overlapping tastes drive the algorithm are themselves shaped by prior exposure to the same commercially dominant music.
Audio feature analysis is a complementary approach that operates at the level of the music itself rather than its listener community. Spotify’s audio analysis system, developed in part through its acquisition of the music intelligence company Echo Nest, extracts a set of measurable acoustic and structural features from each track in its catalog: tempo, key, mode, energy, danceability, valence (roughly, emotional positivity), acousticness, instrumentalness, liveness, and speechiness, among others. These features can be used to identify tracks that are sonically similar to a seed track, independent of whether their listener communities overlap.
Audio feature matching addresses some of collaborative filtering’s limitations — it can identify sonically similar music regardless of whether it has a large listener community — but introduces different problems. Sonic similarity is a real but limited proxy for musical relevance. Two tracks can share nearly identical audio feature profiles while being musically unrelated in any meaningful sense — one might be a country ballad and the other a soft rock song, both slow and acoustic and emotionally moderate, with nothing to recommend one to a fan of the other except superficial sonic resemblance. Conversely, tracks that are deeply musically related — part of the same tradition, produced by musicians who know each other’s work — may differ substantially in their audio feature profiles because they represent different moments or moods within a tradition.
In practice, Spotify’s algorithm combines both approaches, weighting them differently depending on the available data. For very popular artists and tracks with rich listener community data, collaborative filtering dominates. For less popular material where community data is thin, audio feature matching plays a larger role. This means that the algorithm’s behavior differs systematically depending on where in the popularity distribution the seed material sits — a fact with significant implications for the exploratory behavior of radio functions.
4. The Genre Gravity Well
The most consistent and consequential behavioral pattern observable across Spotify’s radio and mix functions is what this paper terms the genre gravity well: the tendency of an extended radio session to drift, over time, from the specific qualities of its seed material toward the commercially dominant center of whichever genre the algorithm has associated with the seed.
The phenomenon is best illustrated by example. A listener who seeds an artist radio function from a critically regarded but commercially modest jazz guitarist from the 1960s will initially receive recommendations that are reasonably proximate to the seed: other players from the same period and tradition, perhaps some adjacent hard bop or post-bop material. Over the course of the session, however, the queue will characteristically begin to drift. The recommendations will tend toward more commercially familiar jazz — the artists with the largest streaming audiences within the genre, the most frequently included tracks on editorial jazz playlists, the names most widely recognized outside the dedicated jazz listener community. By the latter portion of an extended session, the queue may bear little resemblance to its origin point in terms of specific musical character, having been drawn by the genre’s gravitational center toward a version of jazz that is most legible and most consumed in the streaming era.
Several mechanisms produce this drift. Collaborative filtering data, which grows denser toward the popular end of any genre’s distribution, exerts increasing influence as the algorithm exhausts the listener community data available for more marginal artists. The popularity weighting that Spotify’s algorithm applies — which explicitly favors tracks with higher stream counts as a signal of quality or relevance — pushes recommendations toward commercially successful material within the genre. And the personalization layer that shapes artist mixes specifically may, if the individual listener’s history contains more popular material in a genre than marginal material, systematically route the session back toward the familiar even when the seed suggested a desire for the unfamiliar.
The gravity well phenomenon is particularly acute for genres with a wide internal range of commercial visibility. Jazz, classical, folk, and electronic music all contain both highly accessible, commercially dominant subgenres and more demanding, subcultural traditions with small but serious listener communities. In all of these cases, radio sessions seeded from the marginal end of the genre tend to migrate toward the accessible end over time, regardless of whether the listener has expressed any preference for that migration.
The effect is less pronounced in genres with more uniform commercial profiles — where the most critically marginal material is not drastically less commercially successful than the most popular. In these genre spaces, the gravity well is shallower, and radio sessions tend to remain closer to their seed material for longer.
5. Seed Dependency: Hits Versus Deep Cuts
One of the most practically significant behavioral differences in Spotify’s radio functions is the divergence in output quality and exploratory range depending on whether the session is seeded from a major commercial hit or from a deep album cut.
When a radio session is seeded from a track with high stream counts — a canonical hit, a song that has appeared on major editorial playlists, or a track with broad cross-demographic appeal — the algorithm has access to rich collaborative filtering data from an enormous and diverse listener community. The recommendation pool is large, the similarity signals are strong, and the algorithm can draw on data from many different overlapping listener communities. The resulting radio queue tends to be stylistically coherent, musically credible, and reasonably varied in its specific selections, though subject to the popularity bias and genre gravity effects described above.
When a radio session is seeded from a deep cut — an album track that has received relatively few streams, a B-side, a live recording, or a track from an artist’s less commercially successful period — the algorithm is working with sparse data. The listener community for this specific track may be small, the collaborative filtering signals thin, and the audio feature matching therefore correspondingly more influential. The result is frequently a queue that feels less musically coherent: songs that share superficial sonic characteristics with the seed but lack the deeper musical relationship that a knowledgeable human curator might perceive.
This differential performance has a paradoxical implication for exploratory listening. The listeners most likely to seed radio sessions from deep cuts are precisely those with the most developed musical knowledge and the greatest interest in genuine discovery — listeners who are already familiar with an artist’s hits and want to explore beyond them. These are the listeners for whom Spotify’s radio functions perform worst, because their exploratory seeds are precisely the seeds that expose the algorithm’s data sparsity limitations most sharply.
Conversely, the listeners most likely to seed radio sessions from major hits — those less familiar with an artist’s full catalog, exploring from a position of partial familiarity — receive the platform’s most polished algorithmic performance, but in service of a form of exploration that stays closest to the commercially dominant mainstream. The algorithm performs best for the listeners who need it least and worst for the listeners who could benefit most from it.
6. Album Radio as a Special Case
Album radio deserves particular attention because it represents Spotify’s most structurally interesting exploration tool from the standpoint of musical depth. Unlike song radio and artist radio, album radio treats the album as the primary unit of musical meaning rather than the individual track — implicitly acknowledging that albums are coherent artistic objects with properties that exceed the sum of their tracks.
In practice, album radio exhibits several distinctive behavioral characteristics. Because it is seeded from an entire album rather than a single track, the algorithm has access to a richer and more internally diverse set of audio signals. An album that moves through multiple tempos, keys, and emotional registers provides a more complex seed than any single track, and the resulting radio queue tends to reflect this complexity with somewhat greater range than song radio. Album radio seeded from an album with a distinctive and unusual character — an avant-garde record, a concept album with an unusual tonal arc, a recording that sits at the intersection of multiple genre traditions — will typically produce a queue that engages more seriously with those unusual qualities than a song radio seeded from any individual track on the same album.
However, album radio also exhibits the genre gravity well and popularity bias effects described above, and these effects interact in a specific way with the album format. Albums that are well known and frequently streamed provide rich collaborative filtering data, and their album radio queues benefit accordingly. Albums that are critically significant but commercially modest — the case for a substantial portion of the recorded music catalog that is most rewarding for serious exploratory listeners — produce thinner data and correspondingly less coherent radio queues.
There is also a temporal dimension to album radio performance that merits attention. For recent albums within their initial promotional window — the period in which the label is actively promoting the record and listeners are encountering it for the first time — collaborative filtering data accumulates rapidly, and algorithm performance improves quickly. For catalog records released before the streaming era, data accumulation has been slower and more uneven, reflecting the different rates at which different catalog records have been rediscovered and streamed. This means that album radio performance is, among other things, a partial function of an album’s streaming-era cultural profile rather than its intrinsic musical significance — a fact that introduces a systematic distortion into the exploration experience.
7. Recency Weighting and the Catalog Depth Problem
A structural feature of Spotify’s recommendation system with significant implications for exploratory listening is the recency weighting applied to tracks and albums. Spotify’s algorithm gives preferential treatment to recent releases in its recommendation outputs, reflecting the platform’s interest in promoting new music (which aligns with label promotional priorities) and the assumption that recent music is more likely to be relevant to current listener interest.
For playlist contexts, recency weighting is relatively benign — it keeps playlist-based listening feeling current and ensures that the promotional economy of new releases functions as intended. For exploratory radio listening, it represents a significant distortion. A listener who seeds an artist radio or album radio function with the intention of exploring a tradition, a historical period, or a body of work that predates the streaming era will find that their exploration queue is systematically inflated with recent releases at the expense of catalog depth.
This recency bias compounds the genre gravity well effect for listeners interested in historical musical traditions. Not only does an extended radio session tend to drift toward the popular center of a genre, but it also tends to drift toward the recent end of that center — toward the contemporary artists who have inherited and commercially updated the tradition, at the expense of the historical figures who originated it. A listener exploring a radio session seeded from a foundational blues record may find themselves, an hour into the session, listening to recent blues-influenced rock and relatively recent artists with streaming-era commercial profiles, rather than the depth of the mid-twentieth century blues tradition that their seed suggested an interest in.
The practical implication is that Spotify’s radio and mix functions work reasonably well for exploring the contemporary landscape of a genre — finding current artists working in a tradition — but work less well for historical depth exploration, which requires sustained engagement with older catalog material that may have limited streaming-era data profiles.
8. Novelty Decay Over Extended Sessions
One of the most practically observable phenomena in extended Spotify radio listening is what this paper terms novelty decay: the tendency for the rate of genuinely unfamiliar recommendations to decrease as a session extends. Early in a radio session, the algorithm may present several tracks that the listener has never encountered. As the session continues, the recommendations increasingly draw from the listener’s existing listening history, introducing tracks they have heard before, tracks from artists already prominent in their library, and tracks from the most commercially familiar nodes in the genre network.
Several mechanisms produce this decay. The personalization layer that shapes artist mixes and influences radio queues has a finite pool of highly relevant unfamiliar material to draw from — once it has exhausted the most confident recommendations, it retreats toward safer ground. The collaborative filtering data’s density gradient, which thins as the algorithm moves away from the most popular material, produces increasing uncertainty in later recommendations, and the algorithm resolves this uncertainty conservatively by favoring familiar territory. And the listener’s own behavioral signals within the session — skipping unfamiliar tracks, allowing familiar ones to play through — are read as preference data that the algorithm incorporates in real time, steering the session back toward the comfortable.
This last mechanism deserves particular emphasis because it reveals a fundamental tension in the design of Spotify’s radio functions. The skip behavior that the algorithm reads as a signal of negative preference may, in an exploratory listening context, simply represent the listener’s unfamiliarity with the recommended music rather than their rejection of it. Genuinely novel music frequently requires multiple listens before it becomes fully legible; a first encounter with an unfamiliar tradition, artist, or style may produce confusion or mild discomfort that resolves into genuine appreciation with repeated exposure. By reading a skip as a negative signal and adjusting accordingly, the algorithm treats the exploratory listener’s appropriate response to genuine novelty as evidence that the novelty was unwanted. This is a structural design problem that no individual listener can fully work around.
The practical result is that extended radio sessions tend toward a kind of equilibrium that is less novel and more comfortable than the session’s starting point, regardless of the listener’s intent. The radio function that begins as an exploratory tool gradually reverts to a mode that resembles comfortable curation — not through any single decision but through the accumulated weight of real-time adjustments driven by listener behavior and algorithmic conservatism.
9. The Personalization Paradox
Spotify’s personalization features — the artist mixes and Discover Weekly-style recommendations tailored to individual listener profiles — are among the platform’s most widely praised and commercially successful offerings. The promise of personalization is intuitive: a recommendation system that knows your listening history in detail should be better positioned to find music you’ll enjoy than a generic system that treats all listeners alike.
For curation purposes, this promise is substantially fulfilled. Personalized recommendations tend to be more accurately targeted to individual taste than generic editorial playlists, and most listeners report that Discover Weekly and similar features have introduced them to artists they enjoy. The commercial case for personalization is strong.
For discovery purposes, however, personalization introduces a paradox. The more precisely a recommendation system is calibrated to a listener’s existing taste, the more difficult it becomes for that system to introduce genuine novelty — because genuine novelty is, by definition, outside the listener’s taste profile as currently modeled. A perfectly personalized recommendation system would converge on a model of the listener that is isomorphic with their existing preferences, and its recommendations would trace the same contours indefinitely. This is not a failure of implementation but a logical consequence of the personalization goal carried to its limit.
The practical expression of this paradox is that Spotify’s most personalized features — the artist mixes that incorporate listener history as a shaping parameter — tend to produce tighter echo chambers than the less personalized radio functions. Two listeners who both seed an artist mix from the same artist will receive different outputs shaped by their respective histories, and in each case the output will be steered toward what the algorithm predicts they will accept. For the listener whose history is dominated by mainstream pop, the artist mix will be steered toward the pop-adjacent end of whatever artist they seed. For the listener whose history is dominated by jazz, it will be steered toward the jazz-adjacent end. Both listeners encounter a version of the artist filtered through their own prior taste rather than the artist’s full range or the broader tradition within which that artist works.
This means that the listener who most needs to encounter music that lies outside their existing taste profile — the listener who wants to genuinely expand their musical knowledge rather than simply refine their existing preferences — is precisely the listener for whom personalized recommendation features work least well as discovery tools.
10. Strategic Use: Working With and Against the Algorithm
Despite the limitations documented above, Spotify’s radio and mix functions are not without value as exploratory tools. Understanding their behavioral characteristics allows listeners to develop strategies that partially mitigate the genre gravity well, recency bias, and novelty decay effects.
Seeding from unexpected starting points is one of the most effective such strategies. Because the algorithm constructs its recommendation pools from the intersection of audio features and listener community data associated with the seed material, unusual or marginal seeds produce recommendation pools that are less dominated by commercially familiar material. A radio session seeded from an album that sits at the intersection of two genre traditions — rather than at the center of one — will draw from two overlapping listener communities rather than one, producing a more varied and unexpected queue. Similarly, seeding from a track that is associated with a small but passionate listener community — a cult classic, a critically praised but commercially modest record — produces a recommendation pool that reflects the specific and developed taste of that community rather than the broad preferences of a mass audience.
Active engagement with the Dislike function (removing tracks from the queue) and the Like function (saving tracks to the library) provides behavioral feedback that the algorithm incorporates in real time. Used deliberately — consistently dislisting commercially familiar material and saving genuinely unfamiliar tracks — this feedback loop can be trained over time to produce radio queues that are more consistently oriented toward the marginal end of a genre’s distribution. This is a labor-intensive and imperfect approach, but it represents a meaningful way for dedicated exploratory listeners to partially overcome the algorithm’s default biases.
Cross-seeding — deliberately seeding radio sessions from material in an adjacent tradition to the one the listener is trying to explore — exploits the audio feature matching component of Spotify’s algorithm. A listener trying to explore a tradition that is underrepresented in streaming data may find that seeding from a well-documented adjacent tradition produces more useful results than seeding from the underrepresented tradition directly, because the audio feature overlap provides a more reliable pathway than the thin collaborative filtering data available for the underrepresented material.
None of these strategies fully resolves the structural limitations described in this paper. They represent workarounds rather than solutions — evidence of the gap between what Spotify’s radio functions are designed to do and what a genuine musical exploration infrastructure would require.
11. Comparative Context: What Radio Functions Are Not
It is worth clarifying what Spotify’s radio functions are not, in order to frame appropriately what they can and cannot be expected to accomplish.
They are not the equivalent of a knowledgeable human guide. A musician, critic, or deeply experienced listener who constructs a listening pathway for someone exploring an unfamiliar tradition draws on contextual knowledge, historical understanding, and musical judgment that no audio feature analysis or collaborative filtering system currently approximates. They know not only that two artists are sonically similar but why they are similar, what tradition they both inhabit, which of them is more historically significant, and what it would mean for a particular listener to encounter them in a particular order. Spotify’s radio functions produce sonic adjacency rather than musical understanding, and these are not the same thing.
They are not the equivalent of a record store browse. As Paper 6 will examine in detail, the physical record store provided a discovery environment with spatial, visual, social, and serendipitous dimensions that are absent from algorithmic recommendation. The possibility of pulling a record from a bin because its cover art is arresting, hearing it playing in the store, and discussing it with a knowledgeable clerk represents a qualitatively different form of discovery encounter than any queue-based recommendation system can offer.
They are not the equivalent of engaged music journalism. A review or essay that situates an artist within a tradition, explains their historical significance, analyzes their distinctive qualities, and articulates why they matter provides a form of contextual knowledge that is a prerequisite for genuine musical understanding. Algorithmic recommendation can deliver the music but not the understanding.
What Spotify’s radio functions are is a system for generating musically adjacent content at scale, personalized to individual taste profiles, with a genuine capacity for incremental discovery within a listener’s existing taste neighborhood. This is not nothing — it represents a real advance over the closed playlist model — but it is substantially less than the genuine exploration infrastructure that the richness of the recorded catalog would warrant.
12. Conclusion: The Algorithm as Narrow Corridor
Spotify’s radio and mix functions represent the platform’s most serious attempt to provide tools for musical exploration beyond the playlist. They exhibit genuine sophistication in their construction and produce real value for listeners who use them deliberately. But the behavioral patterns documented in this paper — the genre gravity well, the seed dependency gap, the recency bias against catalog depth, the novelty decay over extended sessions, the personalization paradox — collectively describe a system that functions as a narrow corridor rather than an open field.
The corridor is not arbitrary. It is shaped by specific economic incentives (the promotional value of new releases, the commercial interests of major labels with large streaming catalogs), specific technical constraints (the data sparsity problem in marginal genre spaces, the limits of audio feature matching as a proxy for musical relevance), and specific design choices (the real-time behavioral feedback loop that reads skips as negative signals, the personalization features that filter recommendations through existing taste). Understanding these shaping forces is not merely an academic exercise — it clarifies both what can realistically be asked of current algorithmic exploration tools and what would need to change, structurally and economically, for streaming platforms to develop genuinely exploratory architectures.
Paper 3 broadens this analysis by examining how Apple Music, Tidal, Amazon Music, and YouTube Music approach the exploration problem, asking whether their different institutional logics and design philosophies have produced meaningfully different solutions or merely differently shaped versions of the same structural limitations.
This white paper is the second in the Beyond the Playlist series. Paper 3, “Platform Comparison: How Apple Music, Tidal, Amazon Music, and YouTube Music Handle Exploration,” examines the exploration architectures of Spotify’s major competitors, analyzing how each platform’s institutional logic, user base assumptions, and technical infrastructure produce distinctive approaches to the music discovery problem.
