YouTube Custom Feeds: viewers now write the algorithm

YouTube just handed the recommendation engine to the viewer. At Made on YouTube in New York on September 23, the platform announced Custom Feeds: you type a description of what you want to watch, Gemini builds a feed around it, and that feed gets pinned as its own tab on your homepage. YouTube Custom Feeds turn discovery from something inferred about you into something you state in a sentence — and that changes what your video metadata has to do.

What actually happened

Custom Feeds work like a saved prompt. YouTube's own examples are shaped like real intents: video podcasts sized for a 30-minute train commute, or relaxing commentary to unwind with. You can keep several feeds at once, and they roll out to US users on web, mobile and TV this fall.

Studio got the bigger pile of features. Ask Studio, the Q&A layer over your channel data, expands to iOS and Android and now reviews unpublished drafts — feedback on pacing, structure and storytelling before you ship. New Insights and Research destinations surface what's working on the platform and which videos are outliers. Thumbnails get channel-matched generation, dynamic variants served to different audience segments, and a background job that revisits older videos to suggest refreshed art. Video A/B testing is coming: up to three cuts of the same video, tested for which hook holds. YouTube says creators have run more than 40 million A/B experiments since thumbnail testing launched in 2024.

Elsewhere: a Gemini editing assistant in Shorts and the YouTube Create app for the tedious work — reordering frames, syncing music to the beat. Opt-in comment moderation that learns an individual creator's style. Likeness detection adding voice to face detection, reaching the mobile app later this year.

Why prompt-based discovery matters for your business

If you publish video as a channel for a real business — product walkthroughs, install guides, service explainers — this is the same shift that already hit search, arriving on the platform where your demo videos live.

For a decade the job was to guess the keyword. Now a person writes "honest breakdowns of small-business POS systems from someone who's actually used one" and a model decides whether your video is that. Keyword stuffing does not answer that request. A title, description and first 30 seconds that state plainly who the video is for and what it settles — that answers it.

Two practical moves. First, write your metadata the way a customer describes their problem out loud, not the way a keyword tool ranks it. The retrieval layer is a language model now; it reads for meaning and it is not fooled by a comma-separated tag dump. Second, use the draft-feedback and A/B tools for the mechanical parts — pacing notes, thumbnail variants, hook tests — and keep the substance human. The reason your walkthrough beats a generated one is that you have actually replaced that part, priced that job, eaten that mistake. A model can retime your cuts. It cannot supply the scar.

The risk worth naming: every one of these tools deepens your dependence on one platform's judgment about your content. Publish to YouTube, but keep the masters, the transcripts and the customer-facing versions on infrastructure you control.

Sources: YouTube Blog, TechCrunch.