Listen Notes deleted 4,389 podcasts in June.
Not for being bad. For being fake, AI-generated, or both. And that was one month of a purge it’s been running since January.
Spotify did something bigger in the same month, pulling tens of thousands of fraudulent shows that turned out to be fronts advertising illegal online pharmacies. CNN reported it. Senator Maggie Hassan had an investigation open.
Meanwhile, if your inbox looks like mine, you’re getting two or three pitches a week from some AI startup promising to halve your production time. Auto-edit, auto-transcribe, auto-clip for TikTok, YouTube Shorts and Meta‘s platforms, auto-write your show notes.
Most of those tools work. I use several of them.
The problem is that they’re all solving the bottleneck podcasting had in 2019, and the bottleneck has moved somewhere else entirely.
Fewer shows and more shows, at the same time
Two figures from Listen Notes covering January to June this year.
82,642 new podcasts launched worldwide. At that rate 2026 finishes with fewer launches than any year since 2017.
519,708 shows published at least one episode. That’s more than double the whole of 2025, and the highest active count since the 2021 peak.
So: fewer people starting podcasts, considerably more people actually making them. That’s the best pair of numbers this industry has produced in five years.
About 3.8 million podcasts exist on paper. Roughly half a million are alive. The directories are now spending real engineering budget on deleting the difference.
Nobody is building a detector, because detectors don’t work
On Deezer, more than 30% of new music uploads are now fully AI-generated. In the platform’s own listening tests, 97% of people couldn’t tell which ones. Spotify removed 75 million spam tracks in a single year and barely dented the supply.
You can’t filter something that’s infinite and that nobody can identify.
So in May, Spotify quietly changed the question. It stopped asking whether the content was real and started asking whether the audience was.
Read the badge criteria like a spec sheet
Spotify’s podcast verification badge, the green check that now shows up on show pages and in search, has two stated criteria.
Sustained listener activity, meaning consistent engagement held over time.
Verified audience authenticity, meaning actual humans rather than bots inflating your numbers.
There’s also a ban on cloning another host’s voice or likeness without permission. But look at what isn’t in there. Nothing about audio quality. Nothing about how the episode was made. Spotify hasn’t banned AI-generated podcasts and isn’t going to, because it can’t reliably find them.
What it can measure is whether anyone comes back.
So the old advice is now the risky advice
For fifteen years the standard guidance was: publish weekly, stay consistent, the algorithm rewards you.
That’s a liability now. Machine-schedule publishing with shallow listen-through is the exact behavioural signature of the spam these platforms are hunting, and the systems making that call are automated, tuned to be aggressive, and effectively unappealable.
Being listed in a directory is worth nothing. There are 3.8 million feeds and around 7 million titles on Spotify.
Being verified with a real audience attached is worth a lot, mostly because it’s the one thing on the list that can’t be generated.
Three things follow from that.
Retention beats reach. Completion rate and your episode-two return rate aren’t vanity numbers anymore, they’re inputs into your standing on the platform. Four thousand downloads that people finish is a stronger position than fifteen thousand who quit at minute eight.
Cadence should follow capacity. Fortnightly with your audience intact beats weekly while you bleed them.
Identity has to be legible. Named hosts, real bios, a claimed feed, a site you control, verification claimed wherever it’s offered. This used to be housekeeping. It’s now what separates you from the pile getting deleted.
Where to point the tooling
Everything after the record button, automate freely. Transcription, filler removal, levelling, chapter markers, show notes, translation, clips. The output is either right or it isn’t, and a decent tool saves you a week. There’s no romance in noise reduction.
Everything before the record button, do it yourself. Who you book, what you ask, which twelve minutes you cut, what you’ll say that a competitor won’t. That’s the machinery producing the retention signal the platforms are now grading you on.
The failure mode here isn’t that AI-assisted shows sound bad.
They sound fine. Clean, smooth, competent, and completely interchangeable with the other 300,000 clean, smooth, competent shows in your category. In a market where the platform is explicitly measuring whether anyone bothers to return, fine is fatal.
This quarter
Claim and verify every profile you have. Open your analytics and find your completion rate and your episode-two return rate. If you’ve never looked at either, that’s your answer.
Then take the hours the tooling gives you back and spend them on better guests and harder cuts, not on shipping more episodes.
The purge is good news. Three-odd million dead and synthetic feeds are being cleared out of the way, and the field you’re actually competing in is about half a million shows, not 3.8 million.
That’s small enough to stand out in.