Skip to content
Case study

Netflix: the metric fell, but nobody stopped watching

A root-cause analysis of a 10% drop in Average View Duration in two weeks.

Traces the drop to a cold-start discovery failure on mobile, not a content problem, and sequences three fixes with a guardrail against gaming the metric.

Root-cause analysisHypothesis eliminationPrioritization scoringExperiment design
Netflix: the metric fell, but nobody stopped watching full deckOpen full screen

What happened

AVD fell 10%. Watch time didn't move.

Average View Duration dropped 10% in two weeks — mobile-first, mostly new users, global, sudden. This RCA traces it to the denominator: sessions inflated while watch time held flat. Not a content problem. A discovery problem wearing a content problem's costume.

The shape of the drop

−10%
AVD over a two-week window
Flat
total watch time — the tell
sessions per user, inflating the denominator
New
users on mobile, hit hardest

Eliminate before you commit

Five hypotheses. Four fall to the evidence.

Technical failure
Out

Watch time is flat. A real outage would drag it down too.

Content decline
Out

Existing users would feel a weak slate equally. They don't.

External competition
Out

Sessions went up. People are opening Netflix more, not less.

Algorithm regression
Kept

Fits the two-week window and hits new users hardest. Stays on the board as the fallback thread.

Cold start + discovery
Root cause

New users can't find a match on mobile. Explains every signal with no contradictions.

What the user actually lives

Same intent, two very different sessions

Today · broken
  1. Opens the app. New user, no history — generic trending.

  2. Scrolls, overwhelmed. Choice paralysis on a small screen.

  3. Exits under two minutes. Session counted.

  4. Comes back later, hits the same wall. The denominator climbs.

The fix · guided
  1. Opens the app. A three-tap visual taste setup.

  2. Personalised hero and top-3 high-match picks.

  3. Gets into a title. A sustained session.

  4. Returns straight to Continue Watching.

Tradeoffs

Three fixes, sequenced by score, not by flash

Chose

Ship Continue Watching front and centre first

over Leading with the flashy personalisation work

Making Continue Watching the first row with a push deep link turns several short sessions into one arc. It's the lowest-effort, most proven pattern, so on a reach-impact-confidence-over-effort score it ranks first by a wide margin.

Chose

A three-tap visual taste picker at signup

over Letting the recommender learn passively over weeks

A new user's first session is exactly where the data is thinnest and the stakes are highest. A skippable, visual setup gives the recommender something to work with from minute one instead of serving generic trending.

Chose

Gate the auto-play nudge behind a guardrail

over Shipping it wide because it moves the metric

Auto-playing a pick after a few idle seconds would juice AVD whether or not it helps anyone. It ships last, behind a satisfaction guardrail, and gets killed if it only improves the number.

AVD fell because new users on mobile can't find something worth watching fast enough. It's a discovery failure, not a content one.
The whole case in one line — guardrail: sessions per user must stay flat
Next · case study

Flipkart: the return that costs twice

A slow refund is annoying. A silent one is what actually breaks trust.

Continue