Summary: On October 29, 2026, Exa added podcasts to its search index: 3.2 million episodes from 119,000 shows, each with a full transcript — about 23.7 billion words of spoken content. It’s live across all Exa endpoints with no additional charge, and it changes what your agents can retrieve: what people actually said, not just episode titles and show notes.
Why spoken audio was a blind spot
Podcasts are where founders, executives, investors, and researchers talk at length — often before anything gets written down. That made them a gap for agents: audio files aren’t something a typical search backend can rank on meaning. Exa’s bet is that transcripts plus structured metadata turn this into ordinary, queryable text, with new episodes appearing soon after they air.
What the index actually returns
Each result goes beyond a title. Per the announcement, results include show and episode titles, publisher, description, hosts and guests, publish date and duration, and the full transcript with speaker labels. Exa credits data partner Particle for contributions to the index.
The examples Exa gives are the practical kind: the newest episode of Hard Fork, what Jensen Huang said about inference demand on recent appearances, or episodes where founders discuss pricing mistakes. Query by speaker and date, get quotes back — that’s structured data your agent can cite, not a fuzzy summary it has to trust blindly.
The benchmarks, briefly
Exa benchmarked podcast search across 7 APIs, each in its best tested mode, and reports two results:
- Given 250 active shows, Exa returned each show’s newest episode in the top 5 results 98.4% of the time — the best among tested alternatives.
- On 667 questions about specific episodes, Exa answered 85.0% correctly, again the best of the tested set.
The framing matters for builders: stale or incomplete results don’t just degrade answers, they trigger costly retries in agentic pipelines. Freshness is a correctness problem, not a nice-to-have. That’s the same thread I pulled on in Exa’s reranking work — relevance alone stopped being the hard part; getting the right, current result is.
Where this fits in a build
The use cases Exa names map cleanly onto agent workflows:
- Personal assistants summarizing this week’s episode or querying the transcript by speaker.
- Investment research tracking what a CEO said about demand or competition between earnings calls, with attributed quotes.
- Sales, marketing, and recruiting surfacing a prospect’s or candidate’s recent public statements before a meeting.
Since it’s exposed through the existing search API at no added cost, integration is mostly a query-design question: decide whether you want the freshest episode or a semantic match across the corpus, and how far downstream you want to pass speaker-labeled quotes versus summaries.
One caveat: these numbers come from Exa’s own benchmarks, so it’s worth running your own probes — especially freshness on shows you actually care about — before wiring it into anything customer-facing.