Published July 29, 2024 · By CampaignsLive · Case Study
Spotify’s Wrapped campaign has been, since its 2016 launch, one of the most reliable annual engagement events in consumer marketing. The mechanic — a personalized data summary of a user’s year of listening, designed for social sharing — is unusually well-suited to the platform’s strengths and to its audience’s behavior. Wrapped works because it gives users something to share that is about them, that flatters them mildly, and that participates in a shared cultural moment with other users doing the same thing.
The 2023 edition introduced a new element: AI-generated personality categories. Each user was sorted into one of twelve archetypal “Sound Towns” and assigned an AI-generated descriptive persona — categories that the trade and social-press coverage variously labeled “Vampire,” “Alchemist,” “Hunter,” “Luminary,” and so on. The campaign was widely covered. The engagement numbers were, by Spotify’s accounts, strong. The conversation about whether the additions improved Wrapped, made it worse, or simply changed it has not fully resolved more than two years later.
This is a working postmortem from outside the brand.
What was new
The AI persona layer was not the only new element in Wrapped 2023, but it was the most visible. The persona was placed prominently in the user’s shareable summary, alongside the standard top-artists, top-songs, and minutes-listened panels. The persona description was a short paragraph of AI-generated copy, tailored to the user’s listening profile, that read as a personality summary rather than a statistics summary.
Two structural choices in the implementation are worth noting.
The first was that the persona was generated, not retrieved. Each user’s persona description was produced for them, at the moment of their Wrapped reveal, rather than selected from a pre-written set of archetypes. The volume of personalization Spotify was running was, structurally, only achievable through generation. A human-written set would have been one persona per archetype across hundreds of millions of users.
The second was that the persona was a sorting operation as much as a description. The twelve Sound Towns categorized users into groups. The AI description was the user-facing surface of the categorization. The category structure was finite; the descriptions were generated against it.
This combination — finite categorization with generated description — is the architecture that personalization-at-scale tends to converge on, because the alternatives have practical problems. A finite, human-written set is not personalized enough to feel earned by the user; an entirely generated approach is too unstable to maintain brand voice.
What the campaign got right
Three things.
The categorization was finite enough to be shareable. Users who learned they were “Vampires” or “Hunters” could find each other on social platforms. The shared categories produced shared cohorts, which is where Wrapped’s social-sharing engine actually runs. A fully bespoke persona — different for every user — would not have produced this cohort effect, because there would have been no shared identity to rally around.
The descriptions were finishing on top of structure, not the structure. The AI was decorating the categorization, not driving it. The system stayed coherent because the underlying sort was deterministic and explainable. Spotify could, if pressed, defend why a user was a “Vampire” in terms of listening patterns. The AI description was the user-facing translation of a structured outcome.
The risk surface was narrow. Wrapped is a year-end consumer engagement campaign, not a brand-equity or commercial conversion campaign. The brand could absorb a category of personalization-related risk that a different campaign type could not. If a user’s persona description was off, the cost was a minor disappointment, not a commercial harm.
What the campaign got wrong, or at least less right
Two things.
The descriptions were uneven. The generated persona text varied in quality. Some users got descriptions that felt earned and specific. Others got descriptions that felt generic — the AI-flavored uncanniness of language that is grammatically correct, broadly applicable, and not quite anchored in the user’s actual listening. The variance was visible enough that social-platform commentary about it surfaced as part of the Wrapped 2023 conversation.
The personas drifted toward flattery. The descriptions, on aggregate, told users they were creative, distinctive, ahead of trends. The bias was a familiar one in AI-generated personality content — the model has been trained to please — and it produced descriptions that, taken in aggregate, were less interesting than a more honest set would have been. Some Wrapped users explicitly noted that their persona description felt like a horoscope: vague enough to be true, flattering enough to be appreciated, not informative enough to be insightful.
The categorization carried gendered and cultural assumptions. The archetypes — drawn from a set that included some explicitly gendered framings — surfaced in user reception in ways the brand probably did not intend. Several users on social platforms remarked that their assigned archetype seemed to encode assumptions about their identity that did not match their listening. The cost was small but real.
The lesson the campaign left behind
The longer lesson from Wrapped 2023 — visible in the personalization-at-scale work several brands have done since — is about the relationship between structure and generation.
The campaigns that have worked have used generation as a finishing layer on top of structured categorization. The structure makes the personalization defensible; the generation makes it feel earned. Users get the cohort effect from the structure and the specificity effect from the generation.
The campaigns that have not worked have inverted this. They have used generation to drive the categorization, leaving the structure implicit or absent. The result is personalization that does not produce cohorts, does not scale to social sharing, and does not give users a stable identity to participate in.
Wrapped 2023 sat on the right side of this line. The campaigns that have failed in the same space — and there have been several since — have generally been on the wrong side. The lesson is that AI in personalization works best when it is the surface, not the spine.
What Wrapped did next
Wrapped 2024 and Wrapped 2025 continued the AI personalization layer but expanded the surrounding structure. The Wrapped 2024 edition introduced an AI-generated podcast version of each user’s summary, an experiment that produced more conversation than commercial traction. The Wrapped 2025 edition pulled back from the AI-podcast extension and leaned more heavily on the persona model that had worked in 2023.
The pattern across three years suggests that Spotify’s internal read on the AI-personalization play has converged on a narrower set of choices than the initial Wrapped 2023 announcement implied. The persona model has stuck. The AI extensions that broke the structure have not. The campaign has settled into a stable version of the architecture the 2023 implementation introduced, with the parts that did not work pruned away.
For the related discussion of personalization at scale in consumer creative work, see What 500,000 Brand Campaigns Taught Us.