More than 40% of participants in a study of online product ratings used decision patterns that differed from a simple average, giving retailers a reason to examine the spread of their reviews as well as the headline score.
The research, published in PLOS One on July 14, 2026 and publicized by Paderborn University in September, found that some participants concentrated on negative ratings. Others treated four- and five-star reviews as broadly positive, one- and two-star reviews as negative, and paid less attention to the middle category.
The average remained the best-fitting approach for approximately half the sample. The result concerns differences among people reading the same information.
What the experiment actually measured
Researchers from Paderborn University and Ludwig Maximilian University of Munich studied 107 participants recruited from a student pool. They ranked groups of three products using rating distributions without seeing product names or other product characteristics.
The experiment took place in December 2018. Publication in 2026 does not make the observations a survey of today’s entire online-shopping population.
Participants could receive actual products, with their rankings affecting the probability of receiving them. The arrangement gave them a reason to express their preferences. One group saw rating bars alone; another also saw numerical percentages and average scores.
The authors found little evidence that the extra numerical information materially changed the decision patterns. The university’s account of the research says some participants focused especially on one-star experiences, while others used a positive-versus-negative approach.
These were controlled product-ranking decisions. The experiment did not measure an increase in retailers’ revenue, compare the profitability of alternative website designs or establish which real-world products were objectively better.
Two products can share an average and a different downside
A simple numerical example shows what an average removes. Suppose two products each have 100 reviews. Every reviewer gives Product A four stars. Product B receives 75 five-star reviews and 25 one-star reviews.
Both average four stars. Product A’s total is 100 multiplied by four. Product B’s is 75 multiplied by five, plus 25 multiplied by one. Each total is 400, divided by 100 reviews.
| Rating measure | Product A | Product B |
|---|---|---|
| Average | 4.0 stars | 4.0 stars |
| Four- or five-star reviews | 100% | 75% |
| One-star reviews | 0% | 25% |
A customer trying to avoid a seriously disappointing purchase may care about the last row. Another may want to understand why Product B delighted so many buyers but disappointed others. Neither question can be answered by the average alone.
The written comments matter here. A low rating could concern a product fault, poor fit, late delivery or an expectation the product was never designed to meet. Counting negative reviews does not establish which explanation applies.
Review count and context belong beside the score
The Australian Competition and Consumer Commission’s guidance on online reviews says businesses should display the number of reviews behind a rating. A score based on five customers contains far fewer observations than one based on 2,000.
A large count still does not turn reviews into a random sample of all buyers. People choose whether to leave feedback, platforms have collection rules, and commercial relationships or incentives can affect what appears. A highly precise average can summarize a selective set of experiences.
For a buyer, reading the recent negative reviews and the recent positive ones can expose recurring themes. Check that comments concern the same model, version or service being considered. An old review of an earlier product cannot establish the quality of its replacement.
The US Federal Trade Commission’s consumer guidance recommends comparing sources, checking dates and looking at reviewers’ other activity. It also cautions that appearance alone cannot reliably distinguish genuine reviews from fabricated ones. A plausible three-star comment can be false too.
Better presentation does not mean hiding criticism
A retailer responding to the research could show the full star distribution, review count and sorting options clearly. It could also let customers inspect comments about fit, durability or delivery, where those categories accurately describe the feedback.
Such designs need testing. The Paderborn experiment provides a reason to investigate them, not a measured conversion improvement to insert into a business case.
The FTC’s guide for review platforms recommends explaining how reviews are collected, processed and displayed, and how overall ratings are calculated. It advises equal treatment of positive and negative reviews and warns against editing feedback to change its message.
The agency’s questions and answers on its consumer-review rule also explain restrictions on purchased reviews and review suppression. Paying for a five-star consumer review is not made acceptable by disclosing the payment.
For a business testing a more detailed review display, the outcome should extend beyond the number of people clicking “buy”. Cancellations, returns, complaints and purchases customers keep can show whether the information helped people choose appropriately.
Our guide to using click-through rate in marketing decisions explains the same measurement problem: an interaction is an intermediate result, not proof of a satisfactory or profitable sale. A comparison group is needed to separate a display’s effect from changes in traffic or products.
The commercial reason to read one-star reviews is not simply to defend a score. If repeated comments identify a fault or a misleading description, addressing that problem may do more for the customer than changing the way the average is displayed.
