---
title: "The first review decides"
slug: "first-review-decides"
date: 2026-10-28T00:00:00+00:00
description: "''One early rating can still affect the average years later. Online stars are not a measurement of quality, but a process that confirms itself.''"
---

The beer is called Grain Belt Nordeast. On BeerAdvocate, its first review gave it a 4.38. On RateBeer, shortly afterwards, it got a 2.7: the same beer, two different starting points.

Years later, the average stood at 3.49 from 467 reviews, versus 2.59 from 115. The difference never went away.

Behind that difference is a natural experiment that researchers at EPFL took advantage of. Thousands of beers appear on both sites, rated by different people. Usually the first reviews resemble each other. Sometimes they don't, and then it's chance which site got the high one and which the low one. No one had seen the other site.

That let them measure something you normally never measure: what happens if the first judgment had turned out differently.

They compared 7,424 beers. In the extreme group, high on one site and low on the other, the same beer on the site with the high first review received a second rating that averaged 22 percent higher. After twenty or more reviews, there was still about a 10 percent difference in rank.

## Why this happens

An average should become more reliable as more reviews come in. Here the number only gained self-confidence.

The explanation is simple. People use earlier reviews as information: if everyone likes it, it must be good. That reasoning is rational, and it reinforces the signal instead of measuring it. What the first review measures is one person on one day, with a headache or in a good mood. It is a poor sample of the product's quality. After that, that one person determines what the second sees, and the second what the third sees.

The same mechanism is everywhere a first judgment is visible before there are enough other judgments. Under a forum thread, the first reply sets the tone for the rest. In the app store, the first star decides whether someone installs the app at all, and the first bid in a negotiation works as an anchor, even if everyone knows it was chosen at random.

## The solution

The researchers propose a solution that is simple: hide all reviews until a product has at least ten. The sites currently hide only the average score; this is about the reviews themselves. Then you collect ten independent first impressions, and the eleventh reader sees an average that is closer to reality.

You temporarily give up transparency for that. For a site that lives on reviews, that is difficult, because reviews are also entertainment and community. Statistically, the trade-off works out well.

As far as I'm concerned, one practical rule follows from that. For a product with fewer than ten reviews, you look at the first reviewer. An average of two is a coin toss with a little star attached. Wait, or look for something with enough history.

*Source: [Of sheep and beer — Gael Lederrey and Bob West, EPFL Data Science & AI Lab](https://dlab.epfl.ch/2017-08-30-of-sheep-and-beer/); paper: [When Sheep Shop](https://arxiv.org/abs/1802.06578).*