Samurai and the difficulty of predicting a hit
At Sanremo 2026, Samurai Jay’s ‘Ossessione’ finished in seventeenth place out of thirty entrants. A few months later, it went double platinum: number one on Spotify Italia, making its debut in the platform’s Global Top 200, spending a long time at the top of the official Italian singles chart and going viral on TikTok. When a track is everywhere, its success seems obvious. But is this really the case, or is it, rather, merely a retrospective observation, which confuses the outcome of social mechanisms – which can, to a large extent, be directed and amplified through algorithms – with a fate that, nevertheless, no one could have predicted?
Predicting a hit is the oldest problem facing the music industry, and one of the most intractable: for twenty years, an entire field of research – known as ‘Hit Song Science’ – has been attempting to solve it, combining sociology and statistics to analyse the measurable aspects of a song: musical characteristics, lyrics, platform data and initial audience reactions. Predictions of this kind are both possible and fragile, since predicting a hit means estimating the probability that a new track will trigger collective dynamics of attention, imitation, sharing, recommendation and reposting, which depend only partly on its formal merits. Not even the most advanced algorithms can overcome this limitation: they can interpret early signals, however faint, but cannot predict the cumulative effects produced by the song’s circulation itself.
The research has identified certain recurring turning points along the path that leads a song to success. Not a formula, but the stages at which that path is determined. The first concerns its sonic characteristics – what makes the track an immediate contender for market selection. Two sociologists, Askin and Mauskapf, have sought to measure the optimal distance from the epicentre of convention – that core of shared stylistic elements characteristic of a musical era – which hit songs statistically tend to maintain. By studying the Billboard Hot 100 between 1958 and 2016 – nearly twenty-seven thousand tracks – the pair reduced each song to eleven sonic traits (acoustic quality, danceability, energy, instrumentation, key, live performance potential, presence of spoken word, tempo, mode, metre, emotional valence): a sort of formal and perceptual fingerprint of the track. Using a similarity index, they then calculated how closely that fingerprint matched the characteristics of the songs that had appeared in the charts the previous year, deriving a measure of typicality. The result refutes the idea that the market rewards those who conform the most: tracks that are too close to convention tend, in fact, to climb the charts less; conversely, those that reach the top are the ones with a distinctiveness slightly below average. The conclusion is that one must be recognisable, yet stand out from the crowd.
The second key point concerns the social experiment. In a study that has since become a classic, Salganik, Dodds and Watts set up an artificial music market – involving 14,341 participants and 48 tracks by unknown bands – and compared an independent condition, in which participants made their choices without any indication of others’ preferences, with eight parallel ‘worlds’ subject to social influence, in which each participant could see how many downloads each track had already received – the same tracks in each world, but different individuals. This information alone was enough to make the market more unequal and less predictable. The best tracks rarely flopped and the worst rarely took off, but in the middle ground in terms of quality, outcomes could diverge: the same track might triumph in one world and remain marginal in another. In other words: quality narrows the field of contenders, but it is the mechanism of social conditioning that selects the actual winner.
The final stage concerns distribution and streaming platforms. Here, success is not merely observed, but organised, amplified and converted into value. Winkler and colleagues have shown that TikTok acts as a promotional channel for Spotify, but in a highly asymmetrical way. Not all tracks benefit equally from this. The benefit is concentrated in the ‘viral head’ of distribution, where the few tracks most frequently featured in videos account for the lion’s share of views, streams and revenue; in the ‘long tail’, by contrast – the endless array of tracks with a minimal audience – the effect is weak or non-existent. The threshold between the head and the tail, however, does not mark a difference in quality; it arises from the same reinforcement mechanisms observed in the social experiment by Salganik, Dodds and Watts. The recommendation system amplifies what has already gained traction, exposes it to new audiences and cements it in playlists, thereby helping to turn an initial advantage – whether due to chance or not – into a dominant position. The algorithm does not determine on its own whether there will be a hit, but it does influence which trajectories, amongst the many set in motion, rise to the top or remain at the bottom.
