There are well over 500,000 podcasts in the world, covering everything from ABBA to zoology. They’ve reached such a saturation point, in fact, that there’s almost certainly a podcast out there that’s perfectly attuned to your interests—and no great way to find it. Streaming radio veteran Pandora thinks it has a solution in—what else, in 2018—an algorithm. Its Podcast Genome Project, first announced nearly a year ago, launches Tuesday in beta.
That name should sound familiar; it was the Music Genome Project, after all, that propelled Pandora’s early success as a streaming music provider. Its podcast-focused counterpart shares the same underlying objectives, slicing and dicing individual episodes, using more than 1,500 tags to power recommendations. It looks a little different under the hood, though.
“In the case of podcasts, we can rely on machines more heavily than we can with music,” says Pandora CEO Roger Lynch. “Machines can actually determine content, determine intent, there’s many more things they can determine about a podcast than they can about a song.”
As with the Music Genome Project, humans provide both guidance and a backstop for algorithmic choices. But they play a less prominent role with podcasts, at least so far. Instead, Pandora leans on natural language processing to parse the content of a given podcast episode, assigning descriptors related to content type, production style, the host profile, and lots more. If you’ve hit thumbs-up on a podcast episode that talks about, say, parents and kids rebuilding engines for classic cars together, Pandora will find you another, says chief product officer Chris Phillips.
“You can imagine that almost anything that’s discussed in the podcast is a candidate to become a genomic trait,” Phillips says. As it happens, the Podcast Genome Project already pulls from nearly four times as many descriptors as its musical counterpart. “What we’re doing is starting with what we can extract out that we know is a topic someone might search on, but we also know can grow, and then you can put combinations of those together with context.”
Humans come into play as guardrails; natural language processing may not pick up on satire, for instance, or be able to tell fictional stories from fact.
“Machines can actually determine content, determine intent, there’s many more things they can determine about a podcast than they can about a song.”
Pandora CEO Roger Lynch
The granularity of Pandora’s Podcast Genome also hints at some of the challenges it might face. Phillips notes that a recent episode of the Questlove Supreme podcast tackled Atlanta politics at around the 50-minute mark, helping Pandora surface it for anyone who has expressed interest in that fairly specific topic. At present, though, that listener would have to sit through nearly an hour of unrelated conversation, or scrub until they found the relevant discussion.
Which itself raises another potential issue: Unlike songs, many podcasts are a substantial time commitment. You can only listen to so many in a day. Given that, how long might it take for the Podcast Genome Project to garner enough information from your listening habits to make truly useful suggestions?
Clearly it’s not impossible to create a recommendation engine for longer-form content. Netflix and other video streamers do it all the time. But while podcasts may be easier to categorize than music, they may prove trickier to suggest.
“With a music recommendation, if you get one song wrong or one song is off slightly, it’s OK. Each song is only two or three minutes. You can give it a thumbs-up or thumbs-down, and fine-tune along the way,” says Erik Diehn, CEO of Midroll Media, a podcast advertising network that works with shows like Freakonomics Radio and WTF with Marc Maron. Midroll also owns podcast listening app Stitcher. “But when you’re making a choice about a podcast, the stakes are a little higher. You have a smaller set of samples to choose from as you try to figure out what somebody likes or doesn’t like.”
Pandora foresees a future in which it can point people not just to specific podcast episodes, but relevant moments within those episodes. In the more immediate term, Lynch says, the company is encouraging podcast producers to focus on shorter formats, bite-size morsels that represent less of a commitment, and clear the way for less intrusive advertising.
Which, unsurprisingly, is the other half of Pandora’s podcast equation. Its music algorithms don’t just help listeners find songs; they also help advertisers find the right listeners. The company sees even more potential in targeted advertising for podcasts, which inherently cater to more clearly defined interests. And streaming a show through a platform could potentially offer better analytics, which in turn would help podcast producers—and Pandora—charge higher rates.
“We have over 3,000 targeted segments to sell advertisers,” says Lynch. “All the ads that we deliver on Pandora are targeted. If you compare that to podcasts, most podcasts are downloaded. There’s a measurement issue. Do you know if the podcast was even listened to? And how long was the podcast listened to? The answer to all of this is, you don’t know.”
At launch, most podcasts on Pandora will still have the host-read mattress ads to which you’re accustomed. But in time, Lynch hopes to replace many of those with an audio form of the programmatic advertising that permeates much of the rest of the internet. The benefit to the listener hinges on the ol’ “highly relevant ads” argument—how much better do you really want advertisers to know you, really—that Facebook and others have made for years. But for podcast producers, especially outside of the headline acts, it could help make their efforts more sustainable.
“Most podcast players just pull from an RSS feed. Monetization is left entirely to the podcaster. That’s a good thing and a bad thing, depending on how you look at it,” says Diehn. “Obviously it makes it harder for somebody with no infrastructure, no sales team, no real place to start to generate meaningful revenue. There’s no YouTube for podcasts.”
Then again, efforts to provide that sort of ad scaffolding already exist, as do models for discovery, whether from an integrated podcast company like Midroll, popular apps like NPR-owned Pocket Casts, or Spotify, or Google, which recently launched its first native podcast app for Android. Unlike its music discovery engine, Pandora’s Podcast Genome Project won’t be the first major player to take the field. And while the podcast project launches with content from heavy hitters like Gimlet, American Public Media, The New York Times, and NPR, among others, it will still need to aggressively build out its stable in order to provide listeners with the truly unexpected.
At the very least, though, Pandora is giving podcasts a new potential path for success—and if its recommendation engine works as advertised, giving itself newfound relevance in the process.