Rating Distribution

Also known as star breakdown, review distribution

An average is one number standing in for a shape, and two products with the same 4.6 can be different products. One earns it from a tall column of fives, a decent column of fours, and a handful of ones. The other earns it from half fives and a third ones and twos, which is a polarised product: it works well for some buyers and fails for others, and the reason is almost always something specific like sizing, a fit assumption, or a mismatch between what the listing promises and what arrives. The average hides that. The distribution states it in five bars.

The shape most consumer products actually produce is heavily weighted at five stars, thinner at four, sparse in the middle, and slightly raised again at one. This pattern is widely observed in review data and often called a J-curve, and the usual explanation is selection: the people who go to the trouble of writing are mostly the delighted and the annoyed, while the merely satisfied say nothing. Treat it as a common pattern rather than a rule, and read your own category before deciding your shape is unusual.

The temptation for anyone selling review software is to say the goal is more fives. It is not, and a distribution with no ones at all across several hundred reviews is the shape shoppers and platforms treat as suspicious rather than excellent. There are only two ways to produce it. One is to ask selectively, routing likely complainers somewhere private, which is review gating. The other is to remove low ratings after the fact, which is not moderation: a review is removable for what it contains, such as abuse, personal data, spam, or being about a different product, and it is not removable for being inconvenient. Both produce a clean chart and a corpus that no longer describes the product.

The middle of the distribution is the most useful part to a merchant, because two, three, and four star reviews are where the trade-off is written down in words. A five says it is good. A one often says it broke. A three tends to say what the buyer liked, what they did not, and under what conditions, which is the only kind of text that answers a hesitant shopper's actual question and the only kind that gives a retrieval system something specific to quote.

Consider a Shopify store selling a merino base layer at 4.6 stars across 210 reviews. The breakdown is 150 fives, 22 fours, 5 threes, 6 twos, and 27 ones. The 27 ones say almost the same thing in different words, which is that it ran a size small, while the fives talk about the fabric. The average said the product was good, and it was. The distribution said the sizing chart was wrong. Correcting the chart and adding a measured chest width to the listing was a change to the product page rather than to the product, and it applied to every future buyer rather than to the ones already refunded.

Two cautions on reading it. Small samples do not have distributions; five reviews at five stars is five people, and a lower confidence bound such as the Wilson score is a more honest floor for the rating than the average when the count is low. And the standard aggregateRating markup on a product page carries an average and counts, not a per-star breakdown, so anything reading only your structured data sees the number without its variance. If the shape is the part that would reassure a buyer, it has to be readable on the page as text, not only drawn in a widget.