Ratings, Returns, and the Information Value of Public E-Commerce Signals: A Digest of Our Research

Why we study e-commerce signals

Star ratings and reviews are today's dominant quality signal for online purchases. But how much information does that signal actually carry? Six of our studies examine what publicly visible product-page data really says about quality and dissatisfaction risk — and what it no longer can. Below we summarise the main quantitative results and link to the full manuscripts.

1. Rating inflation and the price-quality paradox on Amazon

An analysis of 2.1 million product listings from the Amazon Reviews 2023 corpus. The mean rating is 4.23/5 and fewer than 1% of products fall below 3.0 — the usable scale has compressed to roughly a single star of range (3.8–4.8). Review attention concentrates extremely on a small share of products ("winner-take-all"), and the link between price and visible review-quality signals is weak. The study thereby reframes the price-quality paradox: not as a single-product anomaly, but as a property of the entire signal distribution.

Why it matters: the five-point scale is functionally a one-point scale. Average ratings therefore cannot serve as a comparison tool — and recommendation systems that ingest them as features are working with a collapsed measuring instrument.

Full text: Rating Inflation on Amazon (Google Drive)

2. The predictive ceiling: what returns reveal about information asymmetry

This study tests the upper bound of how accurately publicly visible product-page signals can predict return rates, using machine learning on listing-level data with SHAP explainability. The result: predictive performance has a structural ceiling. That is not a modelling failure but a finding about the market itself: public information is too thin to carry real dissatisfaction risk. The study thereby separates the information value of public signals from the richer proprietary data most of the return-prediction literature relies on.

Why it matters: information asymmetry in e-commerce is measurable. A buyer simply cannot tell from a product page what awaits them — which is the argument for external data signals like a repairability score.

Full text: The Predictive Ceiling of Public Signals (Google Drive)

3. Predicting returns from engagement signals: evidence from Alza.cz

This study (co-authored with Radim Dolák and Lucie Waleczek Zotyková, Silesian University) tests whether customer-engagement signals — average rating, recommendation rate, review volume — together with price and sales volume explain return rates across six categories on Alza.cz, the largest Czech online retailer. The result: some signals, especially ratings and recommendation rates, are associated with lower return rates — but the strength and even direction of these relationships vary substantially across categories. Pooled and category-specific analyses thus map the practical limits of using customer-facing metrics as proxies for return risk.

Why it matters: e-shops and consumers treat ratings as a universal indicator. Our data shows its information value is category-conditional — what works in one category does not hold in another.

Full text: Predicting Product Return Rates from Customer-Engagement Signals (Google Drive)

4. Repairability as a signal: replaceability cues versus returns

A study of ~31,000 products with real return rates tests whether visible replaceability cues (typically the wired vs. battery-dependent distinction) are associated with lower return propensity. It is one of the first empirical tests of right-to-repair-adjacent signals against actual post-purchase outcomes — and a demonstration that cheap, storefront-visible cues can be evaluated without proprietary engineering data. The study also clarifies the limits: a single visible cue is no substitute for a full repairability score.

Why it matters: right-to-repair policy assumes repairability cues correspond to better outcomes. Our data shows how to test that assumption empirically.

Full text: Repairability Cues vs Returns (Google Drive)

5. When reputation fails: high ratings, high returns

On a dataset of more than 31,000 products across 12 categories from the largest Central European e-shop, this study profiles products that combine high ratings with elevated return rates. On average, higher ratings do mean fewer returns — but the mismatch is substantial and concentrates systematically among novelty products, premium-positioned items, and goods that depend on personal fit (experience goods). Category-adjusted models show this is not noise: reputation failure has a predictable structure.

Why it matters: ratings are not a measure of realized quality but an imperfect attention-and-expectation signal. For some product types they fail predictably — and that is exactly where an objective complementary signal is needed.

Full text: When Reputation Fails (Google Drive)

6. Why products vanish: catalog churn and signal stability on Alza.cz

Research typically treats product listings as stable observational units. Our three-wave panel study (13 months, nine appliance and electronics categories on Alza.cz) shows the opposite: six-month delisting hazards are large and strongly heterogeneous across categories. In category-stratified hazard models, the most consistent predictor of delisting is price — not stars, review volume, or return rate. And among listings that survive all waves, signals differ sharply in temporal stability: price and ratings remain relatively stable, while return rate is considerably less reliable over time.

Why it matters: any quality score built on e-commerce data must account for products continuously leaving the market and for some signals "aging" faster than others. This study is the methodological foundation for how we update our database.

Full text: Why Products Vanish: Catalog Churn on Alza.cz (Google Drive)

The takeaway

Public e-commerce signals — stars, reviews, price — carry less information than they appear to. Rating inflation has collapsed the scale, return prediction has a structural ceiling, reputation fails systematically rather than randomly, and signals are not even stable over time. Both consumers and e-shops need an external, objective data signal. That is precisely the role of our repairability score and e-shop API.

A note on the linked texts: the links lead to full-text author manuscripts (preprints) shared via Google Drive. Several are currently under peer review — where a published version exists, please cite that version.