Product · October 21, 2026 · 2 min read
Why every answer comes with a timestamp
AI answers are only useful if you can check them. Here's how we make each claim link to the second it was said.
An AI summary of a podcast is easy to make and hard to trust. Did the host really say that? Was it a guest, or a joke, or taken out of context?
So we made a rule early: every claim in an answer has to point to a moment you can play.
How it works
Transcripts are split into small passages, and each passage keeps the start time of every sentence in it. When you ask a question:
- We find the passages most likely to answer it.
- The model sees those passages numbered, with their timestamps inline.
- It must cite each claim like
[2 @ 14:05]: source 2, at fourteen minutes and five seconds.
Before the answer reaches you, we check every citation. The source number has to exist, and the timestamp has to fall inside that passage. Citations that fail the check are dropped instead of shown.
Tap to hear it
In the answer, each citation is a small play chip. Tap it and the episode starts at that second, straight from the show's own audio file. You hear the words in the host's voice, with all the context around them.
That's the point: the answer is a shortcut to the episode, not a replacement for it.
When timing is off
Some shows insert ads dynamically, which can shift timestamps by the length of an ad break. When we know a show does this, the player says so and gives you quick ±30 second buttons.