Should a podcast publish its own llms.txt?

Yes, if the show has its own domain or a docs-style site. Audio is the one content type an AI crawler cannot read at all: it can fetch the MP3, but it cannot hear it. Everything an assistant can quote — the show description, episode titles, guest names, timestamps and the transcript — lives on the episode page. An llms.txt for a podcast is therefore less an index of pages than an index of transcripts. In our October 27, 2026 check of 30 podcast hosts, only five served a real Markdown file, and none of them was an individual show: they were a directory (Apple Podcasts), a player (Pocket Casts), a hosting platform (Fireside), an industry news site (Podnews) and an editing tool (Descript).

Can a show hosted on Spotify, Apple or Buzzsprout publish llms.txt?

Not at the platform root — you do not control those hosts. You publish at your own path if the show has a custom domain, or at the path prefix you do control, such as a podcast page on your company site. Apple Podcasts serves its own file at podcasts.apple.com/llms.txt and it describes the Apple product, not your show. The practical pattern that works for most creators is a show site on a custom domain that lists episode pages, transcripts and the RSS feed, with the platform links treated as syndication rather than as the canonical home.

Should the transcript go inside llms.txt or behind a link?

Behind a link, one per episode, near the episode URL. A 60-minute episode transcript runs 8,000 to 12,000 words and roughly 11,000 to 17,000 tokens; a fifty-episode back catalogue inlined into a single file is several hundred thousand tokens and no client will read it. Keep the index file small enough to be fetched in one request and list transcript URLs so a client can pull the two or three episodes it needs. If you want the full text in one machine-readable bundle, that is what llms-full.txt is for, and it still needs a size budget.

Podcasts · October 2026

llms.txt for Podcasts: Episode Pages, Transcripts and Show Notes

Audio is the one format an AI crawler can download and still not read. For a podcast, the indexable asset is the episode page and the transcript behind it. Here is what 30 podcast hosts actually serve at /llms.txt, and the show-level structure worth publishing.

What 30 Podcast Hosts Serve Today

We requested /llms.txt from 30 hosts on October 27, 2026 — the major directories and players, the large hosting platforms, and a handful of audio-adjacent publishers — following redirects to the final URL.

Response	Count	Hosts
`200`, `text/plain` or `text/markdown` — a real file	5	podcasts.apple.com (7.4 KB), podnews.net (2.8 KB), descript.com (2.0 KB), pocketcasts.com (1.6 KB), fireside.fm (1.1 KB)
`200`, but HTML behind it	1	www.spotify.com — `text/html`, 47 KB, after redirecting to a locale path
`403` (edge or WAF block)	8	libsyn.com, mixcloud.com, castbox.fm, podchaser.com, redcircle.com, rss.com, podcastindex.org, listennotes.com
`404`	15	iheart.com, npr.org, wnyc.org, prx.org, simplecast.com, buzzsprout.com, megaphone.spotify.com, acast.com, podbean.com, overcast.fm, tunein.com, transistor.fm, podcast.co, creators.spotify.com, soundcloud.com
`503`	1	audible.com

Five real files out of 30 is a lower adoption rate than we measured on developer docs and on news publishers, and the shape of the gap is different. Podcast infrastructure sits behind catch-all routing: overcast.fm/llms.txt returns a 350-byte 404 page, while tunein.com/llms.txt answers 404 with 327 KB of HTML, and transistor.fm/llms.txt with 82 KB. The status code is honest, the body is not, which is why a client that ignores status codes can end up indexing a template.

The Five Files That Exist Do Five Different Jobs

podcasts.apple.com — a 146-line product description: launch date, platforms, listener features, creator features, subscriptions, regional availability. It is Apple explaining Apple Podcasts, not an index of shows. Useful to an assistant asked "what is Apple Podcasts", useless to one asked "find episodes about llms.txt".
pocketcasts.com — the most interesting of the five. It indexes an MCP server, not pages: an endpoint, a .well-known/mcp.json manifest, seven public tools, and an OpenAPI spec. This is the read layer pointing at the action layer — the file tells a client where to search a catalogue rather than listing catalogue pages.
fireside.fm — a hosting platform using the file as a support router: pricing, API docs, help center, and one distribution guide per destination (Apple, Spotify, Amazon, Podcast Index, iHeartRadio, TuneIn).
podnews.net — an industry news site, served as text/markdown.
descript.com — an editing and transcription tool, the closest thing in the sample to a content index.

What a Single Show Should List

A show file has one job: tell a model which pages carry quotable speech. Six buckets cover it.

# The Long Index

> A weekly interview show about technical writing and developer tools.
> 112 episodes, 2019-2026. New episodes every Tuesday.

## Start here
- [About the show](https://example.fm/about): Host, format, audience
- [Episode archive](https://example.fm/episodes): All 112 episodes
- [RSS feed](https://example.fm/feed.xml): Canonical distribution feed

## Episodes
- [Ep 112: llms.txt in production](https://example.fm/ep/112): 48 min, transcript included
- [Ep 111: Schema for AI answers](https://example.fm/ep/111): 51 min, transcript included

## Transcripts
- [Ep 112 transcript](https://example.fm/ep/112/transcript): Full text, speaker-labelled
- [Ep 111 transcript](https://example.fm/ep/111/transcript): Full text, speaker-labelled

## Guests and topics
- [Guest list](https://example.fm/guests): 96 guests with titles and employers
- [Topic tags](https://example.fm/topics): 14 recurring themes

## Contact
- [Sponsorship](https://example.fm/sponsor): Rates, audience size, placement
- [Press kit](https://example.fm/press): Host bios, logo, show art

Two rules matter more than the layout. First, every link needs a description — a bare URL list tells a model nothing about which page answers which question, as covered in link hygiene. Second, the canonical home should be your own domain, with Spotify and Apple treated as syndication. If your only web presence is a platform subdomain, you cannot publish at the root and there is nothing to index.

Transcripts: Link Them, Do Not Inline Them

A 60-minute interview transcript runs roughly 8,000 to 12,000 words. That is 11,000 to 17,000 tokens for one episode. A 100-episode back catalogue inlined into a single index file is over a million tokens — larger than most context windows and pointless to fetch.

So keep llms.txt as the map and put the text behind one URL per episode. Ship the transcript as plain text or Markdown rather than a PDF, keep speaker labels, and include the episode page as well because that is where the summary, chapters and guest links live. If you genuinely need a single machine-readable bundle of everything, that is llms-full.txt, and it still needs a budget — the same size and token limits apply, just with a bigger ceiling.

Networks and Back Catalogues

A network with twenty shows on one domain should not write one flat list of 4,000 episode URLs. Structure it by show: a section per show, three to six links each (show page, archive, latest episode, transcript index), and one network-level section for about, advertising and contact. Let the per-show archive page carry the depth. If each show has its own subdomain, each subdomain can serve its own file — the most specific file wins, so the network root file should still list the shows rather than assume the client walks the links.

Check Delivery Before You Wait for Traffic

Podcast hosts are unusually prone to three delivery faults: a catch-all page returning 200 with HTML, a CDN that blocks unfamiliar user agents with 403, and a locale redirect that answers a text file with an HTML page — Spotify's llms.txt request in our sample ended at /jp/llms.txt as 47 KB of HTML, exactly that failure. Run three commands:

curl -sI https://example.fm/llms.txt | head -5
curl -s -o /dev/null -w "%{http_code} %{content_type} %{size_download}\n" https://example.fm/llms.txt
curl -s https://example.fm/llms.txt | head -3

You want 200, text/plain or text/markdown, and a first line starting with #. If you get HTML, the routing is wrong before the content is. Build the file with the generator, then run it through the checker to confirm every link resolves and the Markdown parses.

What Not to List

Do not list the MP3s. A model can fetch an audio file but cannot read it, and a list of episode media URLs wastes the budget on links nothing can use. Keep out private or subscriber-only feed URLs, unlisted bonus episodes, ad-server URLs, and any transcript of a guest who asked to be off the record. If a show is paid-only, list the public episode pages and stop there.