Why the loudest reviewer is rarely the typical customer

Review scores cluster at the extremes, with a thin middle. This is not because products are either excellent or terrible. It is because a customer who found something adequate has no reason to spend fifteen minutes writing about it, while a delighted or aggrieved one does, and the record is assembled from whoever had a motive.

What everyone sees

A rating average is read as a summary of customer opinion. It is a summary of submitted opinion, which is a sample selected on the strength of reaction. The average of that sample bears no fixed relationship to the average of the population, and the direction of the gap changes with the composition of the extremes.

What is actually happening

Hu, Pavlou and Zhang documented the J-shaped distribution of online product ratings and traced it to two selection stages: who chooses to buy, and who then chooses to report, both of which favour strong opinions. Dellarocas and Wood examined the silence in feedback systems and showed that non-reporting is systematically related to experience, so the missing observations are not missing at random and correcting for them changes the estimated distribution substantially. Both analyses reach the same structural point. The visible ratings are the output of a filter whose criterion is intensity, and no amount of averaging inside the sample recovers what the filter removed.

Why it stays hidden

The selection hides because volume feels like coverage. Two thousand reviews look like a survey, and large samples are trusted. The number of observations says nothing about whether the sampling was biased, and a large biased sample estimates the wrong quantity with more precision than a small one.