The devil is in the details! Effect of differentiated platform governance on online review manipulation

Online reviews play a critical role in shaping consumer behavior, pricing, and sales. However, the growing prevalence of manipulated reviews undermines trust, damages brand reputation, and distorts competition. While prior research has explored antecedents and consequences of review manipulation, it has not compared the effectiveness of different platform governance strategies. This study addresses that gap by examining how differentiated governance—targeting manipulation intensity, manipulation duration, and manipulation of perceived product quality—affects the probability of future manipulation. The authors develop a mixed-strategy game-theoretical model of interactions between platforms and vendors to derive propositions and testable hypotheses. They then empirically test these hypotheses using Amazon data over 25 weeks in 2020. The goal is to identify which governance levers reduce manipulation and under what conditions, thereby informing more effective platform policies.

Prior studies show review manipulation harms platforms (reputational damage, customer churn), vendors (long-term sales decline despite short-term gains), and consumers (misleading decisions, trust erosion). Antecedents include business status (independents vs. chains, early-stage firms) and competitive dynamics (inferior firms manipulate more; manipulation increases when ratings and volume are low). Information manipulation theory delineates quantity, quality, relation, and manner manipulation. Mapping to reviews, the study focuses on three dimensions: intensity (extent of misleading content), duration (time fraudulent reviews remain visible), and perceived product quality manipulation (changing quality revealed via reviews). Platforms employ filters, penalties, and display rules, but the differential effectiveness across manipulation dimensions remains underexplored. The review highlights the need for differentiated governance aligned with manipulation dimensions to better deter vendor behavior.

The study combines an analytical model and empirical analysis. Analytical model: A two-player game (platform vs. vendor) with platform strategies {positive governance, negative governance} and vendor strategies {manipulate, not manipulate}. Penalty upon detection is modeled as F = λ1*m + λ2*mD, where m is manipulation intensity and mD is manipulation duration. Platform loss from undetected manipulation is modeled as L = φ(qR − qC)^2, where qR is perceived quality from reviews and qC is consumers’ experienced quality. The vendor side uses a Hotelling framework: manipulation effort e increases perceived quality from q to q + e, with manipulation cost μe^2. Consumer utility incorporates quality, price, and misfit costs; demands with and without manipulation are derived. Solving the mixed-strategy Nash equilibrium yields optimal probabilities of governance (x) and manipulation (y) as functions of penalties and losses. Comparative statics produce three propositions: (1) stronger governance on intensity reduces manipulation probability; (2) stronger governance on duration reduces manipulation probability; (3) governance related to perceived quality manipulation has a non-monotonic effect on future manipulation depending on whether qR is below or above qC, implying an inverted U relation. Empirical design: Data were scraped from Amazon.com over 25 weeks (July–December 2020). Initial data: 304,626 reviews for 965 products; after cleaning (removing products without continuous manipulation signals and very short reviews), the final panel includes 50,771 reviews for 160 products; 6,062 were flagged as manipulative (~11.94%). Measures: Dependent variable ManDegree is the percentage of manipulated reviews for a product-week relative to the product’s total reviews across the study period. Independent variables proxy platform governance using properties of previously flagged manipulative reviews in the prior period: (a) ManIntensity via recommendation entropy (Shannon entropy of recommendation terms per Fresneda & Gefen 2020); (b) ManDuration as cumulative time flagged manipulative reviews stayed visible; (c) ManQuality as sentiment (Stanford Sentiment Treebank classifier) of flagged reviews, with a quadratic term to capture nonlinearity. Controls include star rating, review volume, price, sales rank, and product weeks on Amazon. Model: Dynamic panel data specification with lagged dependent variable and one-period lags of governance measures; estimated via system GMM (Blundell-Bond). Arellano-Bond and Hansen tests were used to check for higher-order autocorrelation and instrument validity; both were satisfactory.

Analytical model: • Governance targeting intensity (λ1) and duration (λ2) reduces the equilibrium probability of manipulation; • Governance addressing perceived quality displays a sign change depending on qR − qC, implying an inverted-U relationship with manipulation probability. Empirical results: • ManIntensity (recommendation entropy of flagged manipulative reviews) is negatively associated with future manipulation probability (e.g., β = −0.265, p < 0.01). • ManDuration (time flagged reviews remained visible) is negatively associated with future manipulation probability (e.g., β = −0.102, p < 0.01). • ManQuality (sentiment) shows an inverted U-shape with future manipulation probability: linear term positive (β = 0.332, p < 0.01) and quadratic term negative (β = −0.074, p < 0.01). Correlation patterns align with regressions (e.g., intervention intensity vs. manipulation r ≈ −0.76; duration vs. manipulation r ≈ −0.62, both p < 0.05). Robustness: Results hold under alternative dependent variable constructions, alternative sentiment methods (VADER, TextBlob, SentiWordNet), consumer-focused review subsets (top recent/helpful/critical), and across subperiods (Jul–Sep vs. Oct–Dec 2020). Data scope: Final dataset covers 160 products, 50,771 reviews over 25 weeks; 6,062 flagged as manipulative (≈11.94%).

The findings directly address the research gap by showing that treating review manipulation as multidimensional enables more effective governance. Penalizing vendors in proportion to manipulation intensity and the duration manipulated reviews remain visible deters future manipulation. Governance focusing on perceived quality requires nuance: moderate increases in perceived quality can initially encourage manipulation, but excessive divergence from consumers’ own assessments triggers distrust and stronger platform responses, reducing manipulation. This explains the inverted U-shape and suggests platforms should adapt governance based on the qR − qC gap. The results advance the literature beyond generic platform controls, highlighting that governance levers differentially shape vendor strategies. Practically, platforms should calibrate detection, penalties, and removal timing by dimension, and dynamically adjust oversight when perceived quality sentiment departs significantly from experiential quality.

This study is among the first to examine differentiated platform governance across manipulation intensity, duration, and perceived product quality, integrating a game-theoretical model with empirical evidence from Amazon. It shows that intensity- and duration-focused governance effectively lower manipulation probability, while governance tied to perceived quality exhibits an inverted U-shaped relation with future manipulation. The contributions include formalizing how penalties and platform losses map onto manipulation dimensions and demonstrating their heterogeneous impacts in practice. Future research can extend the model to multi-vendor competitive settings, incorporate additional manipulation tactics (e.g., burstiness, helpfulness gaming), and validate results across diverse platforms and product categories to enhance generalizability.

Empirical evidence is from a single platform (Amazon), potentially limiting generalizability to platforms with different governance mechanisms. The analytical model considers a single platform–single vendor game and omits inter-vendor competition effects. The empirical focus is on three manipulation dimensions (intensity, duration, perceived quality); other tactics such as burstiness and helpfulness manipulation are not explicitly modeled. These constraints suggest caution in extrapolation and motivate future studies across multiple platforms, competitive market structures, and broader manipulation dimensions.