1Where this started
I've been titling podcast episodes for a long time, and I've been in an embarrassing number of meetings where reasonable adults debated whether a colon or an em dash would perform better.
Most of those meetings ended with someone's gut feeling winning. That always bothered me. Gut feelings are often right, but nobody could point to any evidence. Podcast advice is full of respectable-sounding recommendations about hooks, trailers and brand consistency, and almost none of it deals with a more basic question: how does anyone find the episode in the first place?
My working theory was that podcast platforms increasingly behave like retrieval systems. If that's true, metadata is part of how the work gets discovered at all, rather than packaging you add at the end of production. Titles, in other words, are search surfaces pretending to be names. I wanted to test that.
2The experiment
The test show was Moving to Canada, an AI-generated informational podcast I built from the Canadian government's documents for newcomers. It had ten episodes and no promotion, and its titles were written to sound like what a real person might type into a search bar, not like tasteful podcast episodes. “How to Get a Canadian Work Permit” is not elegant titling. It is functional titling, which turns out to be the more useful category.
I set it against two professionally produced, promotion-supported narrative documentaries from our network at Snack Labs, then looked at download trajectories and Apple Podcasts engagement data over twelve months. That comparison turned “I noticed something weird” into a sturdier argument, or at least a more defensible one.
Table 1. How the study was set up.
| Element | Implementation |
|---|---|
| Treatment show | AI-generated, search-optimized Moving to Canada |
| Comparators | Two professionally produced, promotion-supported documentary podcasts |
| Primary outcomes | Acast downloads, Apple engagement, and episode-level performance patterns |
| Reproducibility | Data tables, figure-generation script, statistics verification script, and LaTeX manuscript |
3Three shows
This isn't “AI versus humans” in the broad, annoying sense. The comparison is much narrower than that. One show was optimized for retrieval and direct usefulness. The other two were optimized for professional storytelling and conventional promotion. The question was which kind of packaging produced more durable discovery over time.
- Case A
Moving to Canada
AI-generated, search-optimized, no promotion, practical topic, ten episodes.
- Case B
Professional Documentary Show
Human-produced, promotion-supported, narrative format, conventional launch spike.
- Case C
Second Documentary Comparator
Same network context, professionally packaged, stronger launch support, faster decay.
The comparison is intentionally unfair in an interesting way: one show had less polish and less promotion, but much tighter alignment between title, listener intent, and information need.
4What happened
With no promotion, the search-optimized show reached 6,033 downloads across twelve months, and its growth was delayed but sustained. The two documentaries behaved more like traditional launches: stronger spikes at the start, followed by decay. None of that is shocking if you think like a search engine. It is a strange thing to watch happen to shows you helped produce.
I also pulled Apple Podcasts engagement summaries, which mattered because “people clicked it once” is a much weaker claim than “people found it, stayed and kept finding it.” None of this means AI audio is secretly better. The takeaway is that when the query, the title and the episode's actual usefulness line up tightly, that can make up for a remarkable amount of aesthetic compromise.
Table 2. The headline numbers after twelve months.
| Metric | Value |
|---|---|
| Treatment show downloads | 6,033 |
| Observation window | 12 months |
| Episodes in treatment show | 10 |
| Marketing spend | $0 |
5Checking my work
The part I care about most is that this isn't an argument from vibes. The project includes the monthly downloads table, engagement summaries, episode-level metrics, a script that regenerates every figure in the paper and a separate one that recomputes the summary statistics. I submitted a double-blind version for review and maintain a credited preprint. That's a satisfying amount of procedure for something that started with me side-eyeing podcast titles over coffee.
It also changes the claim. It stops being “I tried some SEO-ish titles and things went well” and becomes “here are the numbers, the figures and the manuscript, and here's the precise, narrow claim I'm willing to defend.” I find that more persuasive. It's also more fun.
6So, do titles matter?
I don't think every podcast should sound like a help article. Not every show benefits from titling itself like a search typed by a stressed-out newcomer. But I do think we routinely understate how much distribution logic has crept into editorial decisions.
The broader lesson is uncomfortable, which probably makes it useful: metadata is part of the product. The meaner version is that metadata is the part of the product platforms are willing to reward first.

