Probing the brain of Internet radio service Pandora is like watching Tron backward. It starts in the digital realm with busy bits of data serving a giant, whirring master control program but zooms back out through the wires to an Oakland, California, office full of headphoned humans staring at glowing screens–a scene not too different from the one in Flynn’s arcade, circa ’82.
Instead of feeding quarters into games, though, these music analysts are feeding data into Pandora’s “Music Genome Project,” a music discovery and recommendation algorithm that crunches users’ interests and matches them with song and artist attributes. But the tech and code alone can’t completely explain how Pandora knew you, a staunch Lady Gaga fan, would also love Robyn or Hercules and Love Affair.
“It’s true that the algorithms mathematically match songs, but the math, all it’s doing is translating what a human being is actually measuring,” says Tim Westergren, who founded Pandora in 2000 and now serves as its Chief Strategy Officer. “You need a human ear to discern.”
Pandora’s secret sauce is people. Music lovers. Professional players who pass job application tests requiring them to pick, for example, one of four jazz tunes and “describe the harmonic language,” answer whether it’s “tonal or modal,” and “outline the progression.” These wonks pour over every track that Pandora downloads or rips (yes, rips from hundreds of CDs purchased each week) and catalog its 400-plus attributes before adding it to the 850,000-plus song library. It’s this hominal factor that has helped Pandora keep its engine serving so many new listeners in so many new places.
“That is the magic bullet for us,” Westergren says of the company’s human element. “I can’t overstate it. It’s been the most important part of Pandora. It defines us in so many ways.”
It’s also important, at least in the beginning, for these music analysts to sit, physically, in the same room. That way, they can regularly peel back their headphones and engage with their colleagues about the music they’re categorizing. Senior members are allowed to work remotely–often on the road. “These are musicians who play in bands and tour,” Westergren tells Fast Company.
Real people also help prevent a repeat of the Napoleon Dynamite conundrum that plagued coders working on Netflix’s Cinematch recommendation engine–the movie became a symbol of quirky works that cross genres and user tastes and stymie predictive bots and formulas. Artists such as Beck and King Crimson would have the same potential to cause problems at Pandora, were it not for the analysts, Westergren says. Plus, they’re always on the hunt for undiscovered artists who break out of typical genres and up Pandora’s serendipity levels. “If you really want to effectively surface that stuff, you need a human,” Westergren says.
Pandora now leads a noteworthy cross section of tech-savvy companies that are proudly powered by people. Groupon, for example, uses humans to curate its deal-of-the day service. Groupon and Pandora have been lumped together in reports suggesting they’re planning IPOs for early 2011 (Pandora reportedly met with bankers recently about a $100 million offering). Customer service tool eLoyalty uses caller attributes to define their personalities and match them with compatible customer service representatives instead of running them through robotic keypad options. Hello Music, a service for aspiring musicians and bands, promises that a real music industry pro and musician will listen to their songs and recommend specific ways to monetize their art–or recommend how to improve it. And Amazon’s Mechanical Turk service actually finds people to perform “human intelligence tasks” for companies that can’t afford to rely on code and computers.
“We want Pandora to feel like it’s talking to you,” Westergren says. “We also literally talk to people. We have a team of people who are called listener advocates. Their job is just to respond personally to every single email, phone call, or letter we get. The identity of Pandora is forged through those collective interactions.”
In 2010, Pandora’s people-powered machine purred like never before, forging strategic partnerships and securing spots on everything from iPhones to entertainment systems in the dashes of new cars–lots of new cars. “Half of radio listening happens in cars,” Westergren says. “It’s an important place for us to be.” Pandora’s long been on board Fords and Mercedes, and in the last few months Buick, Toyota, and Mini added options that include the service. Recent deals with Alpine and Pioneer have allowed in-dash components to link up to Pandora via web-connected smartphone apps, too. Pandora’s “anytime, anywhere” mantra seems like manifest destiny now. “In the last few years what we’ve really done is work on the efficiency in how we monetize the service,” Westergren says. “And we’ve gotten really good at it.”
Even after paying $30 million in royalties in 2009 (2010 royalty check was “a lot bigger,” though Westergren wouldn’t say how big, exactly), Pandora turned its first profit at the end of that year, earning $50 million in total revenues. Analysts predicted 2010 would end with $100 million in revenues for Pandora–Westergren declined to confirm or deny the number, saying only of revenue, “It’s all going in the right direction.” Nor would he comment on speculation about an impending IPO. He does claim to have lured more than 2% of radio listeners from its broadcast competitors and says Pandora now has more than 75 million users. And while conglomerates of broadcast stations were busy dumping taste-making disc jockeys for automated MP3 programs, Pandora was fine-tuning its automation, too, but keeping humans at the heart of its playlists.
“The general scaling of the Internet means that automation is tending to increase. As companies try to become more efficient, cut costs, do bigger things less expensively, automation is a common answer to that,” Westergren says. “And I think we’re going to continue to cut against the grain there.”
Follow Tyler Gray on Twitter.