Ask YouTube: AI-Powered Conversational Video Discovery
Learn how Ask YouTube’s conversational discovery works: the preparation and question paths behind it, the components it adds to YouTube’s design, and the decisions that keep its answers grounded, fast, and safe.
The YouTube design so far helps viewers find videos. Search matches keywords from titles, descriptions, and transcripts, and returns a ranked list. That works well when viewers know what to look for. It works less well when they have a question, such as “How does a load balancer decide which server gets a request?”, and the answer sits ten minutes into a forty-minute video.
Conversational discovery closes that gap. Viewers ask in natural language, refine the request with follow-up questions, and receive a short answer that links to the exact moments in videos that support it. YouTube offers this experience as Ask YouTube, both in search and on the watch page, where viewers can ask about the video they are watching.
In this lesson, we look at how the feature works, which components it adds to our design, and the decisions that keep it accurate, fast, and safe.
Disclaimer: This is a proposed architecture for learning the feature. YouTube documents the product behavior, not its internal implementation.
How conversational discovery works
The feature answers from evidence, not from a model’s memory. When a viewer asks a question, the system first finds the transcript passages that address it, then asks a language model to write an answer using only those passages. Every claim in the answer points back to a video and a timestamp. ...