lately

Who are podcast titles really for?

Brian Stever

2026 · Python, RSS pipelines, causal inference, multilevel modeling, LaTeX

The short version. After years of writing titles and descriptions for client shows at Snack Labs, I kept noticing a pattern I couldn't shake: the packaging that seemed to help discovery wasn't always the packaging that helped listeners. This project examines that tension formally. Using a network dataset of 2,585 episodes across ten shows, it looks at how metadata works as labor performed for platforms as much as for audiences. The main finding is that packaging for reach and packaging for quality often point in different directions, except when titles match what listeners were actually searching for. Metadata turns out to be a negotiation between discoverability, listener expectations and the quiet pressure to make every episode legible to a machine.

1Who it's working for

I wasn't asking whether metadata matters anymore. The previous two projects had settled that. The more interesting question was who it's working for. When I make a client show's title longer, its description denser or its keywords more explicit, am I improving things for the listener, or am I doing free optimization work for a recommendation engine?

After years of writing metadata for shows across the network, I'd started to suspect the answer was “both, but unevenly.” This project calls that work platform labor, which I think is exactly the right name for it. Metadata is editorial work, and it's also optimization work: it asks creators to translate cultural objects into something recommendation systems, search indexes and app interfaces can classify. That translation is rarely neutral, and nobody teaches you that part in a production course.

2Building the dataset

I wanted this to be more than a hand-wavy theory argument, so I built it as an end-to-end data problem. There are scripts that discover feeds, ingest RSS metadata, engineer features, merge in download outcomes, validate the dataset and run both preliminary and advanced analyses. I'd had enough of unverifiable claims in podcasting advice, including some I'd made myself.

That structure mattered because the core claim depends on scale and comparability. The analysis covers 2,585 episodes from ten shows in our production network, with audits, sensitivity checks and advanced modeling layered on top. It's less “here's a correlation I found at 1 a.m.” and more “here's the machinery I needed to stop myself believing the first pretty pattern.”

Table 1. The main stages of the pipeline.

StagePurpose
Feed discovery + ingestionBuild a network-wide metadata base from RSS sources
Feature engineeringTranslate titles and descriptions into measurable metadata properties
Outcome merge + validationJoin packaging features with plays, completion, and diagnostics
Advanced modelingRun DML, mixed-effects, and other analyses beyond simple correlation

3More plays, less listening

The central finding is one of those results that makes everyone in the room slightly unhappy. The metadata moves that increase reach are often not the ones that improve listening quality. Before any causal modeling, longer and richer descriptions go with more plays and weaker completion. Nobody wants to hear that, which is usually how you know a finding is worth keeping.

Reach strategy

Longer, richer descriptions and broader packaging can increase plays.

Quality outcome

The same packaging can attract shallower listening unless the title aligns with real search intent.

The point is not that metadata is bad. The point is that metadata often performs labor for platforms and creators simultaneously, and those incentives do not always point in the same direction.

Figure 1. The reach and quality tension from the study, sketched rather than measured. Packaging aimed at discovery can help one side of the funnel while weakening the other.

The exception is search-intent alignment, which I now think of as the least glamorous and most powerful metadata property. When titles more closely match what listeners are actually looking for, they appear to improve both plays and completion. That's inconvenient if you think of titles as pure branding, and welcome news if you care whether people arrive expecting what you actually made.

4What the modeling changed

One result I keep coming back to: the naive partial correlation says longer descriptions hurt completion, but Double Machine Learning reverses the sign once you account for confounding, mainly that longer episodes tend to have longer descriptions. The DML number is a single point estimate without confidence intervals, so I read it as a direction rather than a verdict. It's a good reminder not to publish the first tidy regression in observational media research and call it insight. I almost did exactly that.

The moderator analysis was just as useful. Metadata effects that matter for shows still finding new listeners are weaker for established shows whose audiences auto-play every release. The gaps were large but not statistically significant across ten shows. That makes intuitive sense, but seeing it formalized changed how I think about our newer shows compared with our more established ones. Not all metadata labor is equally necessary. Some of it makes up for an audience that hasn't formed the habit yet. Some of it is ornamental once the habit exists.

Table 2. Four findings, and what they mean.

FindingInterpretation
Reach and completion often divergePackaging for clicks does not automatically package for satisfaction
Search-intent alignment helps bothListener expectations and platform legibility can occasionally cooperate
Causal modeling reverses one naive storyConfounding is doing real work in metadata research
Audience type changes the effectsMature shows rely less on packaging hacks because habit does the lifting

5Calling it labor

I like this project because it gives metadata the right amount of dignity and suspicion. It's real work that shapes discovery, expectations and how people listen. It's also labor that platforms demand, often in forms that reward compliance more than expression.

The gentle conclusion is that podcast creators should think more carefully about packaging. The meaner one, which I tell myself after a long day of writing episode descriptions, is that we're doing unpaid optimization work for recommendation systems and should at least get the satisfaction of naming the arrangement accurately.