Why Podcast Discovery Is Broken in 2026 (And the Editorial Fix)

Apple Podcasts and Spotify recommend the same ten shows to everyone with a security or SaaS interest: Lex Fridman, Acquired, All-In, Tim Ferriss, Lenny's Podcast, My First Million, Huberman, Pivot, How I Built This, and Darknet Diaries. Open any "best tech podcasts in 2026" Google result and you get the same ten in a different order, padded with three you have never heard of and one your competitor sponsors.

This is not a content shortage. There are more than four million podcasts indexed across Apple and Spotify. The shows you should be hearing about exist. The discovery layer is just incapable of surfacing them.

This post explains why discovery broke, why "best of" listicles are worse than the algorithms, and what an editorial alternative looks like in practice.

The universal-recommendation problem

Run a small experiment. Ask three different people in B2B SaaS what Apple Podcasts recommended to them last week. The lists overlap by about 80%. Now ask them what they actually listen to. The lists overlap by maybe 20%.

That gap is the discovery layer failing. Apple Podcasts and Spotify recommend the same shows to everyone because their recommendation systems are optimised for retention and engagement at the platform level, not for genuinely useful discovery at the individual level. The two objectives diverge sharply once you get past the top 50 shows.

The result is a strange consensus reality where every "podcast person" knows the same ten shows, regardless of what they actually need to learn. The shows that are excellent in narrow verticals (the second-best SOC analyst podcast, the third-best B2B SaaS pricing podcast) effectively do not exist in the discovery layer.

Three structural reasons platforms recommend the same shows

1. Creator-economy incentives reward winner-take-all

Spotify pays out programme deals worth tens of millions to a small number of shows. Apple promotes shows in editorial spots that drive seven-figure download lifts. The economic model of both platforms makes them structurally biased toward concentrating attention on the shows where the platform has economic interest.

This is rational from the platform's perspective. A megadeal with Joe Rogan generates more retention than a thousand promotions of excellent niche shows. The platform monetises consensus better than it monetises differentiation, and the recommendation system follows the money.

2. Engagement optimisation kills depth

Recommendation systems on both platforms optimise for predicted listen time and predicted return. The shows that score highest on both metrics are the ones with the lowest barrier to entry: well-known hosts, broadly appealing topics, conversational pacing, no required prior context.

A 90-minute deep dive into PKI internals will lose on both metrics to a 35-minute conversation between two famous people about productivity. The algorithm is doing exactly what it was trained to do. The problem is what it was trained to do.

3. Popularity-anchored cold start

When the system has no signal about you, it defaults to the platform-wide top shows. This is a reasonable cold-start strategy. The failure mode is that it never stops being the cold-start strategy. Even after you have listened to fifty hours of a vertical, the recommendations still anchor on the platform-wide top.

The reason is that vertical-specific signal is weaker than category-level signal for the model. A model that recommends Acquired to a CIAM listener is technically wrong but never catastrophically wrong, because Acquired is genuinely good. The cost of a safe recommendation is low. The cost of a bold, specific recommendation that misses is high. So the system never makes bold recommendations.

Why "best of" listicles are worse

You might think the editorial alternative to algorithmic recommendation is the listicle. "Best 25 cybersecurity podcasts in 2026." "Top 15 B2B SaaS podcasts for product managers." Google these and you will find hundreds.

They are uniformly bad. There are three reasons.

They are SEO-driven, not listener-driven. The writer's job was to rank for a keyword, not to recommend. The list is therefore optimised for completeness (cast a wide net to capture more long-tail traffic) rather than for opinion. Twenty-five shows in a single list is a tell that the writer is hedging.

They are undifferentiated. The descriptions read like the show's own marketing copy because they were lifted from the show's own marketing copy. There is no opinion about what makes the show worth your time, what it covers well, where it falls short, or what kind of listener should bother.

They are stale. A 2023 listicle for cybersecurity podcasts still ranks for the 2026 query. It recommends shows that have stopped publishing, hosts who have left, and episodes that aged badly. Nobody updates the listicle, because nobody is responsible for it.

The combined effect is that the listicle layer is worse than algorithmic recommendation. The algorithm at least knows what you actually listened to. The listicle does not even know that.

What the editorial alternative looks like

The fix is editorial curation by someone who actually listens, with three properties that algorithms cannot replicate.

Curated paths by depth, not category

The right unit of curation is not "podcasts about cybersecurity." It is "three shows to listen to if you are new to identity, and the four shows to graduate to once you are not." The unit is a path through a topic, not a flat list.

Path-based curation gives the reader an arc. You start somewhere accessible (high production value, low prior-knowledge requirement) and you graduate into the harder, more rewarding shows once you have the vocabulary. Algorithms do not understand arcs. Listicles do not even try.

The Podcasts portal organises this way: see the curated listening paths for the working examples.

Opinionated takes that survive being wrong

A useful recommendation has a specific opinion, including the parts that the show does badly. The reader needs to know what they are signing up for. "Risky Business is the best general-purpose cybersecurity podcast, but the production has gotten chattier over the last two years and the Australian context can be jarring." That sentence is more useful than four paragraphs of neutral description.

Opinionated takes also age more gracefully. When a show changes, you can update the take without invalidating the whole recommendation. When a neutral description ages, the only fix is a rewrite.

Dated freshness

Every recommendation needs a visible date and a maintainer. "Last reviewed: April 2026." "This show stopped publishing in 2024, kept for reference only." "This show changed hosts in late 2025; the older episodes are still recommended, the new ones are weaker."

This is the basic editorial discipline of a working library. It does not exist in either the algorithmic layer or the listicle layer. It is what editorial curation can offer that the other two cannot.

The shows that prove the model

The Podcasts portal is the working example of this approach. Some specific recommendations that show what editorial curation can do that algorithms cannot:

• Risky Business for general-purpose security news with a specific point of view.
• Darknet Diaries for narrative-driven security stories, ideal for the non-technical listener you need to bring along.
• Identity at the Center for the actual best identity-and-access-management podcast that the algorithm will never surface.
• Acquired for company-history deep dives, where the path recommendation is "start with TSMC, then ARM, then go where you want."
• Lenny's Podcast for product management, with the caveat that the audio production quality and host preparation make it stand out from the much larger PM-podcast field.
• Latent Space for AI-engineering practitioners, where the recommendation is conditional on you already having the vocabulary.

None of these recommendations is novel. What is novel is the structure: opinion, dated freshness, path placement, and an explicit case for what kind of listener should bother.

What I changed about my own listening

The portal started as a private listening log. I had been quitting more podcasts than I subscribed to for about a year, and I wanted to write down why I quit each one so I would stop relapsing into them. I wrote up the keepers and the quits as the twelve podcasts I listen to weekly and the six I quit in early 2026. That post outperformed every other discovery format I have ever used.

The lesson was that a specific opinion from someone who actually listens is dramatically more useful than a long list from someone who Googled the topic. The portal scales that observation into a structured library.

The discovery problem in one sentence

Algorithmic recommendation optimises for the median, listicles optimise for keyword coverage, and neither optimises for the listener who is trying to get smarter at a specific thing. Editorial curation is the only format that does.

If you have been frustrated with podcast discovery, the fix is not to switch from Apple to Spotify or to read another listicle. It is to find a curator whose taste maps to yours and follow their library. The Podcasts portal is my attempt at being one of those curators for B2B SaaS, cybersecurity, identity, and AI.

Adjacent to this: the 34 blogs every developer and designer should start reading applies the same editorial logic to written sources. The underlying problem is the same, and the fix is the same.