Disclosure: I built this, it’s my project, and paid tiers exist. There’s a free tier and everything shown below is viewable without signing up.<\/p>\n
What it is: structured sponsorship records extracted from public podcast RSS show notes. One row per (brand, episode):<\/p>\n
brand (canonically resolved) | show | episode | publish date | promo code | promo URL + registrable domain | sponsor type (paid \/ affiliate \/ house ad) | confidence | confidence tier | first seen | last seen | the verbatim sentence the claim came from<\/p>\n
Sample rows straight out of the DB:<\/p>\n
– AG1 on Huberman Lab, 2026-07-27, evidence: “AG1: https:\/\/drinkag1.com\/huberman<\/a>“<\/p>\n – Visible on Good Hang with Amy Poehler, code HANG, 2 episodes, 14-day span<\/p>\n – Saily on Machtwechsel (German news podcast), code “Machtwechsel”, 3 episodes over 18 days<\/p>\n Method, since this sub cares about it: LLM extraction over the show-notes text, then a deterministic brand-resolution layer on top. Domain evidence merges entities first (drinkag1.com and athleticgreens.com collapse into one AG1 entity), exact normalized-name match second, and anything that is merely name-similar goes to an adjudication queue and is never auto-merged. That last rule is what keeps Dove the soap separate from Dove the chocolate. Every record retains its source sentence so any claim can be audited by hand.<\/p>\n Honest limits, up front:<\/p>\n – Show notes only. Ads that exist purely in audio and never appear in the notes are invisible to this. Transcript coverage is not built yet.<\/p>\n – The corpus is small right now: 314 episodes across 93 shows, US + DE + FR. It grows daily but this is not a historical archive.<\/p>\n – I am deliberately not publishing an accuracy percentage. I ran a held-out evaluation, then used its failures to fix the extractor, which burns that holdout. Any number I quoted today would be inflated. A fresh untouched holdout is the next task. Until then every record carries a confidence tier and only the CONFIRMED tier is presented as fact.<\/p>\n – No spend or impression estimates. This answers who advertises where, not how much they paid.<\/p>\n Free tier is 200 requests\/month, paid is $49\/$199\/$499. Keys are not self-serve yet, so the page is an early-access list rather than a checkout.<\/p>\n Two things I would actually like this sub’s read on: is a per-record evidence string useful to you, or is it dead weight next to a confidence score? And what would you want joined onto this that is missing (show category, audience estimates, historical backfill)?<\/p>\n https:\/\/podintel.github.io\/?src=datasets<\/a><\/p>\n<\/div>\n submitted by \/u\/Smart-Farmer1966 <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
