Who actually writes reviews?

Plot any large platform’s ratings and the same silhouette appears: a wall of five-star reviews, a spike of one-star reviews, and a valley where the threes should be. The canonical study named it in 2009: Amazon ratings are J-shaped, mostly 5-star ratings, some 1-star ratings, and hardly any ratings in between, which is why its authors concluded the average is a poor proxy of product quality. Platforms have since confirmed the shape from the inside. Trustpilot’s own transparency report counts 73% of its reviews at five stars and 14% at one star, which leaves 13% for everything between.

The mechanism has a name in the literature: polarity self-selection. A 2020 study of 280 million reviews across 25 platforms found a high degree of polarity on most of them, driven by the higher tendency of consumers with extreme evaluations to provide a review. Writing costs minutes, and mild satisfaction rarely pays them. A product that quietly does its job produces a customer with nothing to say.

The silence is enormous and measurable. Nielsen’s old rule for online communities held that 90% of users read without ever contributing. Pew found 82% of Americans read reviews while roughly one in ten nearly always writes one. Even on Airbnb, where every single guest gets a review prompt by email, 68% of trips produced a guest review in the control group of a large field experiment: a third stayed silent with the form already open.

Selection produced the shape. A second force has been moving it upward for years. An NBER study titled Reputation Inflation watched a large online labor market for six years and saw the share of five-star ratings climb from 33% to 85%, concluding that such systems sow the seeds of their own irrelevance. The same paper collects the sibling numbers: the median eBay seller holds 100% positive feedback, and nearly 90% of UberX trips measured in Chicago carried a perfect five.

Buyers have priced this in without reading a single paper. Northwestern’s Spiegel Research Center found purchase likelihood peaks between 4.0 and 4.7 and then falls as scores approach a perfect five, and Trustpilot’s own report notes that over half of consumers in the UK, US, and France believe that a less-than-perfect review score is more authentic. Buyers, in other words, have started correcting for the instrument.

The middle exists. It just talks in different rooms, and it talks when something extracts it. Pay people to review, as one Marketing Science experiment did, and the induced reviews were more likely to contain information about lower quality experiences than same-starred organic ones: the lukewarm truth was there all along, waiting for something to ask for it. That experiment paid people to ask. A conversation asks for free: in a thread, a shrug gets a follow-up question, a complaint collects seconding replies and workarounds, and ambivalence survives contact with the format. Scores count ballots. Threads hold hearings.

Which is why a score and a conversation about the same product so often disagree, and why the disagreement is itself information. The Trustpilot guide documents the shape on a single company: a bank marketing the 4.7 Excellent Trustpilot shows for it, while the snapshot examined there sorts into 80% five-star and 7% one-star, the middle nearly empty and the one-star column telling a story the average arithmetically buried. The record behind that claim carries the testimony itself.