How YouTube started using Google Brain's AI to improve video recommendations in 2015, which now drive 70% of videos' watch time
in lieu of netflix — and it's all because of the F E E D https://www.theverge.com/... Nilay Patel / @reckless : Blown away by the amount Google optimizes YouTube using AI in ways big and small. http://www.theverge.com/... http://twitter.com/...
Context & Ripple Effects
This retrospective closes the loop on a decade-long arc: the InnerTube platform overhaul was already rebuilding YouTube's recommendation engine and search from scratch in 2015, and wiring in Google Brain's models turned that rebuilt infrastructure into an algorithmic front door. The payoff is the number in the headline — 70% of watch time now flows through AI-driven suggestions rather than search or subscriptions.
That makes this story the origin point for everything in the current coverage cycle: Stratechery's argument that YouTube's AI-enhanced video catalog is more monetizable than text-based Search results, the feed's role in pushing Shorts toward 200B daily views, and the emerging tension where one in five videos suggested to new users is AI-generated 'slop' produced by the very system the feed rewards.
First-order effects
- Creators' incentive structure flips: with 70% of watch time routed by recommendations, optimizing for the feed — thumbnails, watch-time retention, upload cadence — matters more than subscriber counts or channel branding.
- Google gains a compounding data asset: every viewing session trains the recommenders further, deepening the moat between YouTube's feed and any rival's catalog.
Second-order effects
- Netflix's subscription-first model faces a discovery-side disadvantage — Nielsen's October numbers show YouTube averaging 6.3M daytime viewers at 11 a.m. versus Netflix's 2.8M, evidence that algorithmic free feeding beats curated catalogs for habitual viewing.
- Advertisers follow the feed: as Stratechery's analysis argues, AI-enhanced video monetizes better than Search text, shifting ad budgets toward YouTube and pressuring platforms whose inventory isn't recommendation-driven.
Third-order effects
- If the pattern holds, the feed becomes the industry's default content-distribution interface — extending into auto-dubbing every upload into every language, per YouTube VP expectations, so the same recommender can route one video globally.
- The system's success creates its own quality problem: when suggestion volume outpaces human review, AI-generated 'slop' fills the supply, forcing YouTube to spend its AI advantage policing the incentives its own algorithm created.
The trend: Recommendation algorithms are absorbing the discovery layer of media entirely, turning platforms like YouTube from content hosts into AI-ranked attention markets.