Answers are grounded in this paper and cite their sources.
Background Online reviews combine star ratings, written evaluations and platform source cues. This study examines the association between rating–text inconsistency and review helpfulness across purchase-verification groups. Methods This observational study uses 479,493 reviews from six Amazon Reviews’23 categories over 2019–2022. A supervised text model was trained on product-disjoint data. Its calibrated test mean absolute error was 0.474 rating points. Helpful-vote counts were modelled using Poisson pseudo-maximum likelihood with product-clustered confidence intervals. Rating indicators, product fixed effects and exploratory specification checks were examined. Results The model-based discrepancy was positively associated with helpful votes in the pooled model (incidence rate ratio = 1.37, 95% confidence interval [1.25, 1.50]). The unverified-review slope was not significant with product and rating fixed effects (incidence rate ratio = 1.025, p = .822). Directional contrasts changed when endpoint ratings were excluded. The verification interaction was negative in the full-sample model ( b = −0.387, p < .001). It was not significant in the product fixed-effects model ( b = 0.117, p = .346). In the two-to-four-star restriction, the pooled M1 association was not statistically significant (IRR = 1.059, p = .067). In the post hoc human-coding diagnostic, model-based and human-implied absolute discrepancies showed limited correspondence (Spearman correlation = 0.234). Conclusions The findings concern a model-based discrepancy measure. Directional patterns and verification moderation depend on the specification and rating range. The human-coding diagnostic indicates limited correspondence with independently coded evaluations. Broader validation is needed before interpreting the measure as perceived inconsistency. The estimates describe observational associations.
No takes yet. Share an insight, caveat, or question.
Nebi Seren (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: