Name: AI podcast feed benchmark, June 2026
Creator: Chain of Thought

Five of the eight AI podcasts in this benchmark declare zero transcripts in their RSS feeds. Across the 2,148 episodes those eight feeds currently serve, 190 carry a machine-readable transcript tag. That’s 8.8%. For systems relying on feed-declared transcripts, most episodes in these catalogs provide no machine-readable transcript text.

This report benchmarks Chain of Thought against seven AI engineering podcasts: Latent Space, Practical AI, The TWIML AI Podcast, Gradient Dissent, AI Engineering Podcast, No Priors, and The Cognitive Revolution. Every number comes from public signals: each show’s live RSS feed and Apple’s iTunes lookup API, fetched 2026-06-11. No download estimates, no chart scraping, nothing private. Methodology below, and the full dataset is downloadable above.

Here’s the argument the data makes: the feed is one important input to podcast discoverability in the AI-search era, alongside the audio and text published elsewhere. When someone asks ChatGPT or Perplexity what a guest said about evals, systems relying on feed-declared text have no transcript to read when an episode appears there only as an MP3. Transcript tags and channel-level metadata are a machine-readable discovery surface, and most leading AI podcasts are leaving that surface empty.

Key findings

All figures from public RSS feeds and the Apple iTunes lookup API, as of 2026-06-11.

5 of 8 AI podcasts benchmarked declare zero transcripts in their RSS feeds: Latent Space, The TWIML AI Podcast, Gradient Dissent, No Priors, and The Cognitive Revolution.
190 of the 2,148 episodes (8.8%) served across all eight feeds carry a podcast:transcript tag.
Chain of Thought leads on feed-declared transcripts at 98.4% (61 of 62 episodes), followed by AI Engineering Podcast at 93.7% and Practical AI at 15.2%.
No show scored 4/4 on feed hygiene. Chain of Thought and AI Engineering Podcast tie for first at 3/4; the other six score 2/4.
Publishing pace varies nearly 8x across the set, from Latent Space at 8.85 episodes per month to Gradient Dissent at 1.15.

The benchmark at a glance

Show	Catalog eps	Eps/month (6mo)	Avg min	Transcripts %	Hygiene /4	Days since last ep
Chain of Thought (subject)	62	2.62	45.4	98.4	3/4	7
Latent Space	207	8.85	71.1	0.0	2/4	6
Practical AI	361	3.28	46.4	15.2	2/4	6
The TWIML AI Podcast	787	1.80	46.3	0.0	2/4	1
Gradient Dissent	137	1.15	53.6	0.0	2/4	15
AI Engineering Podcast	79	1.31	53.8	93.7	3/4	106
No Priors	166	3.77	38.6	0.0	2/4	0
The Cognitive Revolution	349	8.03	98.0	0.0	2/4	0

Data as of 2026-06-11. Apple ratings were also collected but Apple’s public API no longer returns them for any show, so that ranking is omitted entirely rather than guessed at.

The transcript gap

Transcript availability measures the percentage of episodes in each feed that carry at least one podcast:transcript tag, the Podcasting 2.0 standard that tells apps, directories, and crawlers where the text of an episode lives.

Rank	Show	Feed-declared transcripts	Episodes tagged
1	Chain of Thought (subject)	98.4%	61 of 62
2	AI Engineering Podcast	93.7%	74 of 79
3	Practical AI	15.2%	55 of 361
4 (tie)	Latent Space	0.0%	0 of 207
4 (tie)	The TWIML AI Podcast	0.0%	0 of 787
4 (tie)	Gradient Dissent	0.0%	0 of 137
4 (tie)	No Priors	0.0%	0 of 166
4 (tie)	The Cognitive Revolution	0.0%	0 of 349

One honest caveat before the argument: 0% here means zero transcripts declared in the feed, not zero transcripts anywhere. Some of these shows publish writeups or transcripts on their websites, and Latent Space pairs episodes with essays on its Substack. The metric measures what a machine can find by reading the feed, because that’s the surface that travels. A transcript page on your website helps your website. A transcript tag in your feed follows the episode into every app, directory, and answer engine that consumes RSS.

And that surface is where discovery moved. Answer engines cite what they can read. When an LLM-backed search tool assembles an answer about agent memory or GPU economics, a podcast episode is a candidate source only if its words exist as crawlable text with a stable URL. TWIML’s 787-episode catalog represents roughly a decade of interviews with the people who built modern ML, and its feed declares none of it as text. No knock on the show; the podcast:transcript tag was worth little in 2021, when discovery meant charts and word of mouth. In 2026, with a growing share of technical questions answered by machines reading text, an untagged catalog is a library without a card catalog.

Chain of Thought’s position here is a hosting-stack choice as much as a virtue: transcripts are generated for every episode and declared in the feed, and each one also gets a readable transcript page on this site. The one untagged item of 62 is the 61-second trailer. A young show topping this table proves little by itself. What the table actually shows is the rare ranking where age and budget don’t matter: any show on this list could be at 98% within a quarter.

Feed hygiene: the crawlability checklist

Feed hygiene is a 4-point checklist read from each feed’s channel metadata: channel artwork, category tags, a podcast:funding tag (where listeners and platforms find your support link), and a podcast:locked tag (an anti-hijacking declaration of feed ownership). These are the signals directories and crawlers use to classify and trust a feed.

Rank	Show	Score	Has	Missing
1 (tie)	Chain of Thought (subject)	3/4	artwork, categories, locked	funding
1 (tie)	AI Engineering Podcast	3/4	artwork, categories, funding	locked
3	All six other shows	2/4	artwork, categories	funding, locked

Nobody scored 4/4. Every show clears the table stakes (artwork and categories, which hosting platforms set practically by default), and then six of eight stop there. The two newer-generation tags, funding and locked, are declared by exactly one show each. These take minutes to add in most hosting dashboards, so the gap is attention rather than effort: nobody is looking at this surface yet, which is exactly when looking at it pays the most.

Cadence and catalog: where the tradeoffs are honest

Publishing cadence, episodes per month over the trailing six months:

Rank	Show	Eps/month (6mo)
1	Latent Space	8.85
2	The Cognitive Revolution	8.03
3	No Priors	3.77
4	Practical AI	3.28
5	Chain of Thought (subject)	2.62
6	The TWIML AI Podcast	1.80
7	AI Engineering Podcast	1.31
8	Gradient Dissent	1.15

Catalog depth, total episodes:

Rank	Show	Catalog episodes
1	The TWIML AI Podcast	787
2	Practical AI	361
3	The Cognitive Revolution	349
4	Latent Space	207
5	No Priors	166
6	Gradient Dissent	137
7	AI Engineering Podcast	79
8	Chain of Thought (subject)	62

Data as of 2026-06-11.

This is where the subject of the report sits at the bottom, and it should say so plainly: Chain of Thought is the youngest catalog in the set at 62 episodes, against TWIML’s 787 and Practical AI’s 361. Catalog depth compounds, and there is no shortcut for it. Cadence is mid-pack at 2.62 episodes per month, in a field where Latent Space and The Cognitive Revolution each ship more than eight.

Duration shows the format spread: No Priors runs tightest at an average 38.6 minutes, The Cognitive Revolution longest at 98.0, and Chain of Thought sits at 45.4.

The reason to publish these numbers next to the transcript table is the contrast between them. Catalog and cadence are won with years and headcount. The feed surface is won with configuration. A back catalog is an asset in the AI-search era only to the degree that it’s readable, which means the 787-episode library and the 62-episode one are competing on more even terms for machine-assembled answers than the raw counts suggest. That evens out the moment the bigger shows flip the transcript switch, which is partly why this report exists: the gap is real today and it won’t stay open.

Methodology

This benchmark was generated by a command-line tool that fetches public data and writes every raw number with a per-source fetch timestamp. All data was collected 2026-06-11 (07:26 UTC).

Sources:

Direct fetch of each show’s public RSS feed (the same XML every podcast app reads).
Apple iTunes lookup API for catalog depth (trackCount) and genre.

How each metric is computed:

Catalog depth: Apple trackCount. In this run, every show’s feed served its full archive (feed item counts matched Apple’s catalog counts), so feed-derived metrics cover whole catalogs.
Publishing cadence: episodes published in the trailing 183 days, divided by 183/30 (30-day months). Computed from feed pubdates.
Average duration: mean minutes across in-feed episodes that report a duration.
Transcript availability: percent of in-feed episodes carrying at least one podcast:transcript tag.
Feed hygiene: one point each for channel artwork, category tags, a podcast:funding tag, and a podcast:locked tag.
Days since last episode: from the newest feed pubdate to the collection date.

What was deliberately not collected: download numbers (private to each show; no public source has them, and estimating would make the report worthless), Apple or Spotify chart positions (scraping violates their terms), and anything behind authentication. This report benchmarks public signals only. It says nothing about audience size.

The full raw output, including every per-source fetch timestamp and per-episode data, is in benchmark.json (590 KB). The flat per-show table is the CSV download above.

Limitations

Transcript availability reflects the podcast:transcript RSS tag only. A show can publish transcripts on its website without emitting the tag, in which case it reads as 0% here. The metric measures feed-declared transcripts, not all transcripts that exist.
Feed-derived metrics cover the episodes a feed actually serves. Some hosts cap feeds to recent items, which would truncate the window; in this run no feed was truncated, but a re-run against different shows should check that flag in the JSON.
Apple’s public lookup API no longer returns rating counts or averages for any show. Ratings were collected as N/A across the board and the ranking is omitted rather than estimated.
Catalog counts can differ by a few episodes between sources depending on trailers and crawl timing. The JSON records which source each number came from.
This is a point-in-time snapshot of eight shows chosen as Chain of Thought’s peer set in AI engineering. It is not a census of AI podcasts, and a different peer set would shift the rankings.

If you run one of these shows and a number looks wrong, the raw JSON has the fetch timestamps and per-episode data to check it against. Tell me what I missed.

About Chain of Thought

Chain of Thought is an AI engineering podcast hosted by Conor Bronsdon, with guests from NVIDIA, Google DeepMind, AMD, Cisco, Databricks, and Vercel, covering agents, evals, infrastructure, and the business of AI. Every episode publishes with a full transcript on this site, which is also why this report exists: the bet that machine-readable feeds matter is one the show has already made.