Netflix’s recommendation algorithm posted a 40% jump in click-through rate the same quarter average session length fell 18%. Most teams read the gap as expected: the model improved first, and watch time will catch up once the signal compounds. This breakdown tests that assumption against two alternatives before betting on the wait.
Click-through rate measures the share of recommendation impressions that result in a play. It rose 40% the same quarter a major recommendation model update shipped. Average session length, the median minutes watched per session, fell 18% over the same window.
New training signal takes weeks to propagate. CTR responds first because it reflects the ranking change directly. Session length responds later because it depends on subscribers building trust across multiple sessions.
A 2-4 week lag between signal types is normal after a large model refresh. Rolling back now would sacrifice a real CTR gain to react to an incomplete metric. The standard read: let the model run, watch session length over the next month, and resist calling failure early.
Session length should recover during a lag, not keep falling. If the model were mid-compounding, watch behavior would hold flat or improve slowly as better matches accumulate, not drop another 18%. A metric moving the wrong direction during its own grace period is diverging, not lagging.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.