The Song Went Viral

A great song finds its audience. This is the foundational belief of every musician, every listener, every streaming platform's marketing department. Spotify has 253 million tracks. The algorithm learns what you like. Release something good enough and the system will surface it. Talent rises.

The belief feels true because the mechanism feels personal. Discover Weekly delivers 30 songs every Monday that sound like what you already listen to. Release Radar surfaces new music from artists you follow. The home screen rearranges itself around your habits. The system appears to discover music for you, one listener at a time.

The belief is structurally incorrect.

In 2006, Matthew Salganik, Peter Dodds, and Duncan Watts published an experiment in Science that tested a simple question: does the quality of a song predict its success?¹

They built an artificial music market. 14,341 participants listened to 48 songs by unknown bands and decided which ones to download. In the control group, participants chose without seeing what anyone else had done. Quality roughly predicted ranking. The songs that sounded better were downloaded more often.

Then they added one variable. Visible download counts. Participants could see how many times each song had been downloaded by previous listeners. Nothing else changed. Same songs. Same interface. Same quality. Same listeners drawn from the same population.

The inequality of market outcomes nearly doubled. The Gini coefficient for market share rose from approximately 0.34 in the independent condition to approximately 0.50 with social influence visible.¹

The most important finding was not the inequality. It was the unpredictability. The researchers ran eight parallel versions of the social influence condition. Eight separate worlds, same 48 songs, same starting point. Across the eight worlds, the same song could be ranked #1 in one world and #40 in another. The rankings were not reproducible.

"The best songs rarely did poorly, and the worst rarely did well, but any other result was possible." Salganik, Dodds & Watts, Science, 2006

Quality set a floor and a ceiling. The best songs stayed near the top. The worst stayed near the bottom. Everything in between, the vast majority of the market, was determined not by the music but by the accident of who listened early and whether that early signal compounded.

One design variable. Same songs. Nearly double the inequality. The experiment was published in Science eighteen years ago. Its structural finding describes the streaming economy more precisely than any industry report since: the primary predictor of which songs get heard is not the quality of the song. It is the initial allocation of visibility.

The silent catalog

If Salganik's experiment describes the mechanism, the Luminate 2025 Year-End Report describes the outcome at industrial scale.²

253 million tracks are available on streaming platforms. In 2025, 541,000 of them drove 49.4% of all global audio streaming. That is 0.2% of the catalog capturing half of all listening.

120.5 million tracks, nearly half the catalog, received fewer than 10 plays in the entire year. Not per month. Per year. Ten listens across twelve months across every country on earth.²

106,000 new tracks are uploaded every day. The catalog expands by roughly 39 million songs per year. The listening does not follow. It concentrates.²

In 2004, Chris Anderson published "The Long Tail" in Wired, arguing that digital distribution would shift demand from hits toward niches. The aggregate of non-hits would become a bigger share of total consumption. Infinite shelf space would liberate the market from the tyranny of the blockbuster.³

Will Page tested the theory in 2008. As Spotify's future chief economist, he obtained transaction data from a major digital music provider. Eighty percent of all songs had zero transactions. The top 3% of songs accounted for 80% of all sales.⁴

Anita Elberse tested it at Harvard using Rhapsody data. The top 10% of titles accounted for 78% of all plays. The top 1% accounted for 32%.⁵ David Hesmondhalgh and colleagues, in the most comprehensive study of music creator earnings commissioned by the UK government, found that the top 1% of artists accounted for approximately 80% of all streams.⁶

The long tail exists. It is available. It is functionally silent. More choice did not produce more diverse consumption. It produced the same concentration with a longer silence underneath it.

The viral moment

On the scatter plot of streaming, "Old Town Road" is the dot that sits alone.

In December 2018, Montero Hill bought a beat for $30 from a Dutch teenager on BeatStars. He recorded "Old Town Road" in a bedroom and uploaded it to SoundCloud. He promoted the song through TikTok memes he made himself. There was no label. No manager. No playlist pitch. No marketing budget.⁷

By March 2019, the song was on the Billboard Hot 100. In April, Billboard removed it from the Country chart, a gatekeeping decision that generated more attention than any playlist could have. By June, under the name Lil Nas X, "Old Town Road" was the #1 song in America. It stayed there for 19 weeks, the longest run in Billboard history.⁷

The standard telling is a story about disruption. A teenager with no connections made a song in his bedroom and beat the industry at its own game. Streaming democratized music. The gatekeepers lost.

The structural telling is different.

Columbia Records, a division of Sony Music, signed Lil Nas X within weeks of the song charting. The Billy Ray Cyrus remix was a Columbia production. The 19-week run at #1 was supported by label marketing, playlist relationships, and the promotional infrastructure of a major. Every subsequent release was made inside the system.⁷

The viral origin story became the marketing narrative for a Sony-distributed career.

This is a description of the architecture, not an indictment of the artist. One song out of 253 million found an audience without the machine. The career that followed required the machine. The outlier sustained the belief that the system is open. The subsequent releases confirmed that it is not.

The system did not get disrupted. It got a new scouting department.

The architecture

The concentration of listening is produced by a set of interlocking mechanisms that compound each other. Together they form a system where the distribution of attention is self-reinforcing.

The pool

Spotify uses a pro-rata royalty model. There is no fixed per-stream payment. Each month, Spotify collects all revenue from subscriptions and advertising, takes approximately 30%, and distributes the remaining 70% to rights holders based on their share of total global streams.⁸

A listener who plays nothing but underground jazz all month has their $10.99 subscription distributed based on total global streams, not what they personally played. The bulk of every subscriber's fee flows toward the most-streamed content regardless of whether that subscriber ever listened to it.⁸

The alternative, a user-centric model where each subscriber's fee goes only to the artists they actually played, would redirect money toward niche and mid-tier artists. Deezer has experimented with it. Spotify has not adopted it.

The effective per-stream rate has declined roughly 40 to 50% since 2014. Not because Spotify cut the rate as a policy decision. Because the royalty pool is divided among more total streams each year. In 2015, Spotify hosted approximately 8 million tracks. In 2025, 253 million. Revenue grew. Total streams grew faster. The denominator consumed the numerator.⁹

The bottleneck

In 2021, Luis Aguiar and Joel Waldfogel measured the economic value of a single playlist placement. Being added to Today's Top Hits, Spotify's flagship playlist, raised a song's streams by approximately 20 million and was worth between $116,000 and $163,000 in a single editor's decision.¹⁰

Spotify employs an editorial team that selects songs for playlists like RapCaviar, New Music Friday, and Today's Top Hits. Artists can submit unreleased tracks through Spotify for Artists at least seven days before release. Less than 2% of submissions make it onto significant editorial playlists.¹¹

Tiziano Bonini and Alessandro Gandini studied this process through ethnographic research with platform curators. They found a dual logic they termed "algo-torial power." The first week after release is editorial. Editors select tracks for playlists based on pitch submissions, data signals, and label relationships. The second week is algorithmic. Engagement signals from the editorial placement, skip rates, save rates, completion rates, determine whether a song stays, rises, or falls across algorithmic playlists like Discover Weekly and Release Radar.¹²

Editorial gatekeeping sets the initial conditions. The algorithm amplifies them. The mechanism Salganik demonstrated in a laboratory, where initial visibility compounds into large outcome differences, operates at industrial scale inside Spotify's playlist infrastructure.

Major labels account for approximately 70% of tracks on flagship playlists. They maintain dedicated streaming teams whose job is to build relationships with editors. The pitch process is technically open to anyone. The outcomes are structurally predictable.¹¹

The three companies

Universal Music Group, Sony Music Entertainment, and Warner Music Group collectively control approximately 70% of global recorded music revenue. On Spotify, the three majors plus Merlin, an independent label consortium, accounted for 87% of all streams in 2017. By 2024, that share had fallen to 71%.¹³

Before Spotify's IPO in April 2018, the three major labels held approximately 18% of Spotify's equity. They received these shares as a condition of licensing their catalogs. The total price paid was roughly 8,800 euros. At the IPO, Sony's stake alone was worth approximately $1.5 billion. Warner sold most of its stake for $400 million within weeks. By late 2024, Universal's remaining shares were worth approximately $3 billion.¹⁴

The three companies simultaneously negotiate licensing rates with Spotify as suppliers, own equity in Spotify as shareholders, and influence how the royalty pool is allocated as the dominant rights holders. Most-favored-nation clauses in their contracts guarantee that if any label negotiates better terms, the others receive the same. The structure produces coordinated economics without requiring explicit coordination.¹⁴

Tamas Tofalvy and Julia Koltai tested whether this offline power translates into algorithmic outcomes. Their network analysis, published in New Media & Society, found that the primary determinant of an artist's position in Spotify's recommendation network was not genre similarity or musical quality. It was their label's international connections. Artists on international labels received genre-appropriate recommendations across borders. Artists on local labels were clustered by country of origin regardless of what their music sounded like.¹⁵ The algorithm reproduces the power structure it was trained on.

The design

In 2025, Rackowitza and Haamplandb published a study in Information Economics and Policy that used Spotify's 2018 mobile redesign as a natural experiment. The redesign reduced the visual prominence of the platform's Top 50 playlists on the home screen.¹⁶

Consumption of those playlists dropped by more than half.

Same songs. Same catalog. Same users. Same algorithm. A change in where a playlist appeared on a screen determined what millions of people listened to. The platform is not a neutral delivery mechanism. It is the architecture that determines consumption.

Spotify's own researchers confirmed a related finding. In 2020, David Holtz and a team including Sinan Aral ran a randomized field experiment on the platform. Personalized recommendations increased engagement by 28.9%. They also decreased individual listening diversity by 11.5%.¹⁷ The algorithm makes each person's consumption more narrow. It optimizes for the next click, not the next discovery. The better it gets at predicting what you will listen to, the smaller the range of what you hear.

Two catalogs

Spotify does not have one catalog. It has two.

The visible catalog is 253 million tracks. Every song ever uploaded. The promise of infinite choice. The marketing copy that streaming democratized music by making everything available to everyone.

The functional catalog is what people actually hear. 541,000 tracks drive half the listening. The distance between the two catalogs is the entire argument.

Twelve million people uploaded music to Spotify in 2024. Here is what they earned.¹⁸

Eighty artists earned more than $10 million. Fifteen hundred earned enough to constitute a professional income in a Western country. 97.5% earned less than $1,000 for the year.¹⁸

These are gross numbers. What Spotify pays goes to rights holders, usually labels or distributors. A major-label artist on a traditional deal receives 13 to 25% of what Spotify pays the label. At the midpoint rate of $0.004 per stream, a million streams generates $4,000 in gross royalties. The artist's share on a standard major deal: roughly $500 to $1,000. Before taxes.¹⁹

To earn the US federal minimum wage from Spotify alone through a major label, an artist would need approximately 50 to 83 million streams per year. That is roughly 4 to 7 million streams per month, sustained for twelve months. The number of artists who achieve that level is not published separately, but it falls somewhere between the $10,000 tier and the $100,000 tier. A few tens of thousands, globally, across every genre and every country on earth.¹⁸

"A wealth of information creates a poverty of attention." Herbert Simon, 1971

Simon was describing organizations. He was describing Spotify before it existed. 253 million tracks represent a functionally infinite supply of music. The binding constraint is not content. It is the cognitive capacity to process it. This makes the distribution of attention, playlists, algorithms, the arrangement of a home screen, more economically important than the production of music.²⁰

The artists at the top know this. The labels at the top know this. The platform knows this. The only people operating under the belief that great music gets discovered are the ones uploading 106,000 songs per day into a catalog where half of everything will receive fewer than 10 plays this year.

The pattern

Every system I've written about operates on a belief nobody checks.

In Formula 1, the belief was about what the product is. It looked like a racing series. The economics said it was a franchise model disguised as a sport.

In retail, the belief was about whether the economics work. Founders held a margin number that had nothing to do with what ended up in their bank account.

In the Premier League, the belief was about what determines the outcome. It looked like a meritocracy. The data said it was a payroll receipt with a ten-month delay.

In pricing, the belief was about when the economics get decided. Founders believed they set the price. The price was determined when they committed the COGS. Everything after was discovery.

In music, the belief is about how the audience finds the song.

It feels organic. The algorithm feels personal. The recommendation feels like discovery. The platform presents itself as a neutral surface that connects listeners to the music they will love. 253 million tracks. Personalized for you.

Salganik proved in a laboratory that initial visibility, not quality, determines success in cultural markets. The streaming economy reproduces that mechanism 5.1 trillion times per year. A royalty pool that rewards total volume. A playlist system where one editorial decision generates 20 million streams. Three companies that control 70% of what flows through the system and own equity in the platform itself. A recommendation algorithm that narrows what each person hears while optimizing for engagement. A UI redesign that halved consumption of an entire playlist category by changing where it appeared on a screen.

The song did not go viral. The song was placed where the architecture needed it to be. And the architecture called it discovery.