Podcast Performance Intelligence
2026 · Client work
Predicting what works instead of describing what happened.
The problem
Podcasters make their most important decisions on gut feeling. Which format keeps people listening, which segments deserve to become clips, where an ad read should sit – the feedback they get answers a different question than the one they're asking.
View counts are a real signal – at the episode level. Across a catalog they tell you which topics and guests create demand: what your audience wants to click. What they can't see is the moment level. Views measure demand for the promise – title, thumbnail, topic, or clips that went viral – not whether the conversation delivered on it. An episode with 200k views may have lost its audience after eight minutes; one with 40k may have held it to the end. The view count cannot tell those two apart, and everything that decides whether a show grows durably happens in that gap.
The signal that actually describes content is retention: second by second, where do people stay, where do they leave. Every podcaster has this data sitting in their analytics – and almost none of them can act on it, because a retention curve tells you that people left at minute 31, not why, and not what to do about it next episode.
The bet
The claim is that the interesting product is not showing podcasters their data – dashboards do that, and decisions rarely change because of them. The interesting product is prediction: given a transcript, say in advance where attention will break, and which moments will work as clips. If a model can do that on episodes it has never seen, it has learned something real about the audience. If it can't, everything else it says is post-hoc storytelling.
That's why the eval comes before the product. The first artifact isn't an app – it's a model that ranks segments of an episode by clip potential, tested on held-out episodes against honest baselines (random, position in episode, simple heuristics), with the numbers published either way. Existing clip tools hand you a "virality score" with no published evidence that it beats guessing. The verifiable claim is the product.
Clips are the deliberate starting point because they are the one place where the loop closes on public data alone: which segments got clipped and how those clips performed is visible for any channel, at scale – no access to private analytics required. Everything deeper – retention modeling, drop-off prediction – needs exactly that access, and it is earned, not assumed.
The sponsoring angle
The clearest place this becomes money is sponsorships, because money already flows there – on remarkably thin evidence.
Look at typical podcast media kits, including those of large channels: reach, subscriber counts, age and geo breakdowns. What they almost never contain is proof that the ads work. How many viewers are actually still watching when the mid-roll starts? How hard is the skip dip at the ad read compared to the show's baseline? Which position in an episode loses the least audience? These are answerable questions with the data podcasters already own – and answering them changes the economics: a mid-roll at 82% retention and one at 61% are two different products currently sold at the same price. A podcaster who can prove the difference can price it, slot by slot, and back a higher CPM with evidence instead of reach.
That's the longer arc: a data layer under the podcast where the same model serves two sides – telling the creator what works, and telling sponsors what it's worth.