In one version of an online music experiment, a song called “Lockdown” by 52Metro finished first. In another, it came 40th out of 48. Both groups had the same recordings to choose from. Each group saw a different history of what people had downloaded.
It’s a question you can’t normally ask of a hit: if you could take it back to the day nobody had heard it and start over with a similar audience, would it win again? In 2006, Matthew Salganik, Peter Dodds, and Duncan Watts found a way to run that experiment. Their MusicLab study, published in Science, let 14,341 participants listen to unfamiliar songs and download the ones they wanted to keep.
Some listeners could see how often songs had already been downloaded. The researchers randomly assigned these listeners to eight separate versions of the site within each experiment. Every version started at zero and showed only its own download counts. In the second experiment, songs with the most downloads also moved to the top of the list. An early download could give a track a better spot, and a better spot could give it another chance to be heard.
The researchers also had listeners who saw song and artist names without download totals. They compared two groups drawn at random from those listeners, repeating the split 1,000 times. On average, a song’s share of downloads varied less between these groups than between the worlds with visible totals. The groups’ overall preferences lined up better when listeners couldn’t see what others had picked.
A song’s appeal still mattered. Tracks that did well among listeners who couldn’t see the totals generally had a better shot in the other worlds, too. But the experiment shows how a chart can do two jobs at once: record what people chose and steer what the next person hears. A chart may look like a consensus from inside one group. Try to build it again with a second group, and you can see how much of that consensus holds up.
That’s a question we could ask on Kallitone. A blind vote starts with two recordings. Listeners hear at least 15 seconds of each, starting at a randomly selected point within each song, with the artist names hidden. Their picks go toward a ranking people can browse separately. We reported 25,000 completed comparisons across 500 tracks in our beta. To find out how much of a chart would repeat, we’d need to build it twice.
One way to test that would be to randomly assign voters to two groups. Give them the same catalog and the same set of song pairings, then build a separate chart from each group’s votes. Keep the charts hidden while voting is open and compare them after the same number of votes. Do they pick the same winner? How many songs appear in both top tens? Does the overlap grow as more votes come in?
Imagine the groups disagree on number one but put eight of the same songs in their top tens. You’d have a strong shortlist even if the winner depended on which group you asked. For someone looking for their next favorite artist, that might be the more useful result: two rooms full of strangers, listening separately, keep finding the same songs.