How an article is made

Every Podread article starts from the public YouTube captions of the source video, fetched through our transcript service. The captions go through a sequence of OpenAI calls that clean the speech-to-text output, plan a magazine-style outline, then write each section back into prose.

The same pipeline handles 30 article languages: the model can write directly into the target language from the source transcript, or translate a finished article in a single second pass. Every finished article is cached on Cloudflare R2 so the next reader, in any language, gets it instantly without re-spending model credits.

Quality controls

Three gates keep low-value content out of the public index.

First: every video must be at least 5 minutes long and carry a transcript of 500+ words before we even call the language model. Shorter clips are rejected up-front so we never spend on something that can't yield a real article.

Second: the model rates every article it writes from 0 to 100 for density, structure, and source-citation quality. Articles scoring under 40 are automatically marked unlisted — they're still reachable by direct link for the person who generated them, but they don't enter the discover feed, the sitemap, or AI-crawler-facing indexes.

Third: a signed-in user can generate at most 5 articles per day from the same podcast channel. This prevents any single show from being machine-fanned across all 30 languages in one sitting, which would look like exactly the kind of bulk-translated near-duplicate content search engines penalise.

Every public article carries a visible AI disclosure footer, in the reader's UI language, confirming the article is AI-assisted and verified against the source transcript.

Who wrote what

The podcast host remains the experiential authority — their words are the load-bearing content of the article. They are attributed as the article's author (a Person entity in our structured data), with links back to their channel and any public profile we can verify.

The editorial role — cleaning, outlining, summarising, translating, fact-checking against the source — is attributed to Podread, the publisher. Crucially, the human host is not the literal writer; the writing pipeline is the AI. We encode this split explicitly so search engines and AI assistants can cite both the speaker and the publisher correctly.

Takedown and corrections

If you are a podcaster or a rights-holder and you would like any article based on your video removed, edited, or kept unlisted, email [email protected] from a publicly verifiable address (your channel's contact email, your domain, or a DM from your verified profile). We act within 7 days.

The same address handles factual corrections. We would much rather fix a wrong claim than leave it in the index — write to us with the article URL and what's wrong, and we will update or unlist it.

Get in touch

[email protected] handles everything: a podcast you would like to see processed, a bug report, a partnership idea, a takedown request, or just feedback on the product. Reply within a week, usually within a day.