Voice-mode answers from ChatGPT and Perplexity are pulling podcast titles directly into responses to 'recommend a podcast about X'-style queries. The retrieval is happening at the episode level, not the show level.
The metadata schema we're testing across four shows: a one-sentence episode promise (under 140 characters), three topic tags drawn from a controlled vocabulary, and a guest credential string. All three are getting cited verbatim in voice answers.
The transcript itself matters less than we expected. The model isn't summarizing the episode; it's matching the episode's promise to the listener's question. A precise, narrow promise outperforms a 'we discuss everything about X' description by roughly 4x in our pulls.
The full schema, the controlled vocabulary we've started with, and the four-show test set are in the appendix.
"Your podcast description is now an SEO asset. Treat it like one."