The reason engagement and satisfaction diverge

A recommendation system improves on every measure it tracks while users report that the results are worse. Both observations can be accurate. Clicks and dwell time are recorded automatically; whether the item was wanted has to be asked, and what is asked is a fraction of what is logged.

What everyone sees

The divergence is read as a values problem — platforms choosing profit over experience. Incentives contribute. But the same gap appears in systems with no commercial motive, which indicates a measurement problem underneath: the available signal is a poor proxy for the objective, and optimisation exploits every gap between them.

What is actually happening

Kleinberg, Mullainathan and Raghavan formalise the difficulty of inferring what users want from what they do, showing that behavioural traces are consistent with multiple underlying preference structures, so a system trained on behaviour cannot distinguish a satisfied user from a captured one. Ekstrand and Willemsen argued that purely behavioural evaluation is insufficient for recommender research and that subjective measures capture distinct and non-redundant information. Together they describe an ordinary optimisation outcome: the objective and its proxy diverge under pressure, and the system moves toward whichever items maximise the proxy.

Why it stays hidden

The gap hides in the reporting cycle. Behavioural metrics arrive continuously, in volume, and are trusted; satisfaction is sampled occasionally, in small numbers, with wide intervals, and is discounted as noisy. The stronger evidence is therefore the weaker measure, and when the two disagree the decision goes to the one with more decimal places.