product sense · standard
"How would you improve YouTube's recommendations?"
How would you improve YouTube's recommendations?
The failure mode on this question is treating watch time as the unambiguous North Star and proposing algorithm tweaks to increase it. YouTube’s own product team rejected that framing in 2025 when they overhauled the recommendation model to weight satisfaction signals: post-watch surveys (“Was this video worth your time?”), comment sentiment, and long-session retention. A video with slightly lower watch time but high satisfaction is now promoted aggressively. A video with high views but poor satisfaction is suppressed. Candidates who pitch “more watch time” in a Google interview are arguing against the company’s stated direction.
The question interviewers are actually asking is whether you can hold a real tension: maximizing engagement today can destroy the product long-term if it drives regret, rabbit holes, and burned-out viewers who churn. YouTube’s algorithm drives 70% of all platform viewing. Recommendations are not a feature; they are the product. At that scale, the choice of which satisfaction signal to optimize carries more weight than any individual feature proposal.
Start with a clarifying question
Ask which surface before proposing anything. Home feed, Up Next sidebar, and Shorts feed have different user intents and different algorithm architectures. YouTube fully decoupled the Shorts recommendation engine from long-form in late 2025, so an answer that conflates the two will read as out of date. Scope to the Home feed for logged-in users in lean-back sessions: it accounts for the largest share of algorithm-driven viewing and is where topic lock-in and the regretful-click problem are most acute.
Structure a strong answer
strong
"Before I propose anything: which surface are we optimizing? Home feed, Up Next, or Shorts? And what user state matters most? I want to scope to the Home feed for logged-in users in lean-back sessions, because it is the highest-impact surface and has the clearest unresolved pain.
My goal is not to maximize watch time. It is to maximize sessions users return from feeling satisfied. That framing is grounded in YouTube's own 2025 shift to satisfaction-weighted discovery, where satisfaction proxies now directly shape ranking. A product that captures 40 minutes today from regretful, momentum-driven scrolling is not the same product as one a user opens tomorrow by choice. Those are different businesses.
Three user segments to name: the Habitual Daily Viewer who knows what they want but gets trapped in a topic loop; the Discovery Seeker who wants to find new creators but gets served more of what they already watch; and the Mindless Scroller who defaults to YouTube for downtime and is most vulnerable to the regret cycle. I would anchor the rest of my answer on segments one and three because both surface the satisfaction tension explicitly.
Four pain points worth naming: (1) topic lock-in: watch one cooking video, get a cooking-heavy feed for three weeks; (2) recency bias over quality: the algorithm surfaces new uploads from followed creators even when those uploads are weaker than their archive; (3) no user vocabulary: there is no lightweight mechanism to say "I am done with this topic" without clearing full history; (4) the regretful click: clickbait that generates views but leaves viewers unsatisfied, which YouTube itself calls out as an active penalty signal in the post-2024 model.
Solutions in priority order. First, a satisfaction decay signal: after N consumed videos in a topic cluster, reduce confidence in that cluster and probe adjacent topics. This uses the post-watch survey infrastructure YouTube already ships and requires no new surface or user behavior. Second, session-intent detection: if a user opens the app and watches a 45-minute documentary, the next session's Home feed should reflect a lean-back context, not default to viral clips. Third, explicit topic control: a "seen enough of this" signal on any recommendation card, or an extension of the promptable "Your Custom Feed" chip YouTube tested in November 2025, where users type a text prompt to reshape their Home page. That feature exists in testing; shipping a simplified version is a product decision, not a feasibility problem. Fourth, creator quality signal over creator loyalty: decouple subscription from automatic promotion and surface subscribed creator uploads only when recent satisfaction scores clear a threshold.
Metrics. Primary: 7-day return rate segmented by post-session satisfaction survey score. Are users who report high satisfaction actually coming back at higher rates? Secondary: topic diversity index per user per week, to measure whether the feed is expanding or contracting over time. Guardrail: ad revenue per satisfied session, not per raw session, to hold the business case.
One trade-off to name explicitly: solutions one through three may reduce total watch time in the short term because they interrupt momentum-driven scrolling. That is the same bet YouTube's own team made in 2025, and it is the right one. I would run this as an A/B test with the guardrail metric gating any rollout."
weak
"I would improve the recommendation algorithm to increase watch time by better matching users to content they like, using signals like watch history and subscriptions." This fails on three counts. First, it treats watch time as an unambiguous goal, which YouTube abandoned in 2025 in favor of satisfaction-weighted metrics. Second, it ignores the engagement/satisfaction tension: more watch time from regretful, clickbait-driven sessions actively harms long-term retention. Third, it shows no product intuition about the actual YouTube landscape: no mention of Shorts versus long-form being decoupled, no awareness of the promptable feed in testing, no acknowledgment that "regretful clicks" are a named concept YouTube explicitly penalizes. Interviewers at Google probe this tension directly because it is the hardest judgment call in the question.
The rabbit-hole nuance
Many candidates reach for anti-rabbit-hole features as their centerpiece. Know the research context: Brookings and UPenn’s CSS Lab have both published work challenging the radicalization-via-rabbit-hole narrative, finding limited evidence that the algorithm systematically drives users toward extreme content. A strong candidate names this complexity rather than asserting rabbit holes are the obvious problem to solve. The more defensible and measurable pain point is regret and topic lock-in. The radicalization claim is contested and harder to instrument. Conflating them reads as surface-level preparation.
The PM judgment
The interviewer is checking whether you can reject a metric that is easy to optimize when optimizing it works against the user and, eventually, the business. In 2026, feasibility is largely solved: YouTube has the ML infrastructure, the data, and the compute to build almost any recommendation variant. The challenge sits entirely on the viable/lovable axis. YouTube can capture more minutes today, or it can build a product people actively choose to open tomorrow. Those are not always the same thing, and knowing which one to argue for is the distinction between a strong answer and a forgettable one.
For the engagement-versus-satisfaction conflict as its own interview topic, see two metrics in conflict: engagement vs. revenue. For the notification parallel (a feature that drives time-on-site while satisfaction stays flat), see notifications up, time on site flat. And improve Instagram Reels covers the same two-sided market reasoning in a Meta context.