YouTube Adds Prompt-Defined Custom Feeds Built on Gemini

YouTube announced custom feeds, a feature that lets users describe in their own words the kinds of videos they want recommended, at its Made on YouTube event on September 23, 2026. TechCrunch

Users create a feed by typing a description into a prompt box on YouTube. Google's Gemini AI model then builds the feed around that request. Each feed is pinned to the top of the user's home page in its own tab.

The custom feeds will not replace YouTube's main recommendations feed. The main feed stays. YouTube said support for creating multiple custom feeds will roll out across YouTube on web and mobile starting next month.

As examples, YouTube suggested a feed of video podcasts for a 30-minute train commute or relaxing commentary videos to unwind with. Emily Moxley, YouTube VP of Product Management for Viewer AI, said the company's corpus contains over 20 billion videos.

The announcement came at YouTube's annual Made on YouTube event, held in New York City on September 23, 2026. The Alphabet-owned company also unveiled artificial intelligence creator tools and shopping updates at the event. Reuters

The broader context here is a shift from inferred intent to declared intent. Most production recommenders learn from implicit signals such as watch history, session behavior and collaborative patterns. A prompt-defined feed inverts that flow. The user states the job in natural language, and the system must map that statement onto candidate generation, ranking and refresh over time. For engineers, the familiar tensions apply: precision against recall, freshness against relevance, and constraint satisfaction against serendipity.

In this author's view, the more durable change is inspectability. A feed defined by a sentence the user wrote is easier to reason about and revise than one shaped by months of clicks. Worth flagging is how brittle natural language can be as a control mechanism. Narrow prompts can starve the candidate set, while broad prompts can collapse back toward the default ranking. Support for multiple pinned tabs points toward one plausible resolution, with users maintaining several narrow lenses rather than trying to tune a single master feed.

Looking at what this means for builders and creators, explicit feeds create distribution surfaces tied to stated need rather than inferred taste. That rewards catalog depth, consistent metadata and clear formatting, because a system must match language to inventory before it can rank. It also changes evaluation. When the user has declared a task, such as filling a commute or winding down, click-through and raw watch time become weaker proxies. Task completion, repeat visits to a specific tab, and prompt edits and refinements become more informative. If users adopt that behavior, recommender work moves closer to constrained retrieval, instruction following and transparent controls. That would give viewers a steering wheel they have not had before, without removing the main feed for those who prefer to scroll.