medium convictionactive · updated 2026-08-21T00:00:00.000Z
YouTube recs → regretted watches
Recommendations (not search or subscriptions) drive a large share of YouTube viewing; homepage and Up Next are built from watch-history and current-video signals; Mozilla's crowdsourced regrets land mostly on recommended videos. YouTube's own borderline-demotion numbers do not close that gap.
The chain
1
Recommendations drive a significant share of YouTube viewership — more than subscriptions or search.
From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: "Recommendations drive a significant amount of the overall viewership on YouTube, even more than channel subscriptions or search."
2
Homepage ranking is primarily watch-history-based; Up Next is primarily the video currently playing — both engagement-inferred, not user-drawn.
From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: "Up Next uses the video currently playing as the main signal; homepage primarily uses watch history."
3
Videos people later regret are more often recommended than searched for (Mozilla), which is a different metric than YouTube's demoted-borderline watchtime.
From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: "71% of regret reports were recommended videos, and recommended videos were 40% more likely to be regretted than searched-for videos."
What would falsify this
- Step 3: A later fetched study with a disclosed sampling frame finds recommended videos no more likely to be user-regretted than searched videos, on a metric comparable to Mozilla's.
- Step 1: YouTube publishes (and we fetch) an official split showing search + subscriptions overtaking recommendations as the majority of viewership.
Contradictions / tensions
- YouTube (Goodrow, 2021): a growing number of independent researchers conclude recommendations are not steering viewers toward extreme content; recommended borderline watchtime is significantly below 1% after 2019 demotion (70% U.S. drop on non-subscribed recommended borderline). Different metric than Mozilla regret.
- Mozilla's sample is crowdsourced and self-defined; it is not a random panel of all YouTube users.
- Liu et al. (PNAS 2025) find limited short-term *polarization* from rabbit-hole-like perturbations of real YouTube recs. Polarization is not Mozilla-style regret and not on-intent centeredness. From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs.
- Roth et al. (2020) find topological confinement in *non-personalized* YouTube suggestion graphs — a different object than user-regretted recommended videos. Canonical for that spine: youtube-suggestion-graph-to-confinement.
Implications
- A user-owned interest graph would be aimed at Mozilla-style off-intent drift (and ordinary entertainment rabbit holes), not only at YouTube's classifier 'borderline' bucket.
- Watch history is the homepage hinge and is closed in the Data API — see youtube-data-access-constraints.
Companies
Concepts
Open questions