Sustainability puffery rarely raises customer satisfaction; consumers focus on performance and value. A multimodal AI that fuses seller copy, user reviews and image cues flags likely greenwashing with 81.6% accuracy on an Amazon hold-out set and 97% on a large Persian marketplace, offering platforms a scalable screening tool.

From signals to trust: Multimodal detection and consumer perception of sustainability claims in e-commerce

Seyed Mohammad Sina Mirabdolbaghi, Adel Aazami, Sebastian Kummer · August 20, 2026 · Electronic Commerce Research and Applications

openalex correlational medium evidence 7/10 relevance Summary only summary available; pdf_status=paywall DOI Source PDF

Structured author observations

Linked only from stored provider relations; the raw author line above is never matched by name.

OpenAlex

Latest observation: August 24, 2026

Seyed Mohammad Sina Mirabdolbaghi provider ID
Adel Aazami provider ID
Sebastian Kummer provider ID

Semantic Scholar

Latest observation: August 24, 2026

Seyed Mohammad Sina Mirabdolbaghi provider ID
Adel Aazami provider ID
Sebastian Kummer provider ID

Sustainability language in listings and reviews does not predict higher ratings or satisfaction, while a multimodal AI combining seller claims, reviews, and image cues can detect likely greenwashing with strong accuracy (81.6% on Amazon hold-out; 97% on BaSalam).

Citation observations

Cumulative provider counts captured on specific dates; providers are never combined.

0 cumulative citations

OpenAlex · Observed August 24, 2026

View corpus context

0 cumulative citations

Semantic Scholar · Observed August 24, 2026

View corpus context

This study pursues two interrelated objectives. First, it conducts a text-based analysis of consumer product reviews to examine whether sustainability-related language in reviews and product descriptions is associated with product ratings and customer satisfaction. Second, it develops and empirically tests a multimodal AI-based framework for detecting greenwashing in e-commerce environments, identifying misalignments between seller-generated sustainability claims and consumer-verified experience. Drawing on Signaling Theory and the Theory of Planned Behavior as interpretive lenses, the study analyzes data from two large-scale e-commerce platforms: The Amazon Sales Dataset (1465 product listings) and the BaSalam Persian-language marketplace (182,275 product-review pairs). Natural language processing, image-based color analysis, statistical testing, and supervised machine learning are employed. Results indicate that sustainability-related language in consumer reviews is not associated with higher product ratings or customer satisfaction; consumers evaluate products primarily on functional criteria such as performance, usability, and value. The analysis further shows that sustainability-mention frequency differs significantly across discount categories, with the High Discount category exhibiting the highest frequency. The proposed multimodal detection pipeline achieves an overall accuracy of 81.6 % on the Amazon hold-out set and 97 % on the BaSalam dataset, suggesting its potential usefulness for real-time consumer-facing and regulatory screening contexts. Implications for green marketing strategy, platform governance, and sustainability communication policy are discussed.

Summary

Main Finding

Sustainability-related language in product reviews and descriptions does not predict higher product ratings or greater customer satisfaction; consumers prioritize functional attributes (performance, usability, value). A multimodal AI pipeline that fuses seller claims, consumer reviews, and image cues can detect likely greenwashing (misalignment between seller sustainability claims and consumer-verified experience) with high accuracy (81.6% on an Amazon hold-out set; 97% on a large Persian-language BaSalam dataset), indicating practical potential for real-time consumer-facing and regulatory screening.

Key Points

Theoretical framing: interprets results via Signaling Theory (sustainability claims as seller signals) and the Theory of Planned Behavior (consumer intentions/behavior).
Consumer reviews with sustainability mentions are not associated with higher ratings or satisfaction; functional features dominate evaluations.
Sustainability-mention frequency varies by discount level; the High Discount category shows the highest frequency of sustainability language—suggestive of strategic usage or targeted green claims during promotions.
Multimodal detection (text + image cues + review signals) identifies misalignment between seller claims and consumer-experienced evidence, i.e., potential greenwashing.
Performance: detection pipeline achieves 81.6% accuracy on Amazon hold-out and 97% on BaSalam, illustrating robustness across languages/platform scales in this study.

Data & Methods

Data sources:
- Amazon Sales Dataset: 1,465 product listings (used for training/hold-out evaluation).
- BaSalam marketplace: 182,275 product-review pairs (Persian-language dataset).
Methods:
- Text analysis of product descriptions and consumer reviews using NLP to detect sustainability-related language and mention frequency.
- Image-based color analysis as a visual cue feature set.
- Statistical hypothesis testing to assess associations (e.g., sustainability mentions vs ratings; differences across discount categories).
- Supervised machine learning in a multimodal pipeline combining seller claims, consumer review signals, and image features to classify likely greenwashing.
Evaluation: hold-out testing on Amazon and out-of-sample evaluation on BaSalam, reported as overall accuracy.

Implications for AI Economics

Information asymmetry & signaling: Evidence that sustainability language does not raise satisfaction suggests green claims function more as marketing signals than as indicators of experienced product quality—impacting models of signaling equilibrium and reputational capital in marketplaces.
Market design & platform governance: Multimodal AI affordance enables scalable screening tools for platforms and regulators to detect deceptive sustainability claims, influencing platform policy design, listing moderation, and trust mechanisms.
Consumer welfare & competition: Automated detection can reduce exploitative greenwashing, improving welfare by aligning claims with experienced attributes and potentially shifting competition toward substantive sustainability investments.
Pricing and promotional strategies: Higher sustainability-mention frequency in high-discount categories suggests strategic behavior; platforms and economists should consider how discounts interact with signaling incentives and consumer inference.
Policy and regulation: High-performing detection models support more targeted enforcement and disclosure requirements, but raise considerations about false positives, cross-lingual generalization, and seller gaming—necessitating careful calibration, transparency, and continual model validation.
Research directions: Integrate detection outputs into structural economic models to quantify welfare gains, seller response equilibria, and optimal regulatory interventions; examine long-run dynamics as sellers adapt to automated screening.

Assessment

Paper Typecorrelational Evidence Strengthmedium — The paper uses a large Persian dataset (182k pairs) and performs hold-out and out-of-sample evaluation, lending credibility to the predictive claims; however, the Amazon training/hold-out set is small (1,465 listings), the operationalization of 'greenwashing' appears to be a proxy based on misalignment rather than validated ground truth, and analyses are correlational so causal claims about signaling or behavioral mechanisms are not identified. Methods Rigormedium — Methods combine standard NLP, simple image-cue features, and supervised ML with both in-sample hold-out and cross-platform out-of-sample tests — a solid empirical approach for detection tasks. Limitations include unclear labeling/annotation protocol for greenwashing, potential class imbalance, limited image feature sophistication (color analysis only), and absence of robustness checks for confounders or alternative explanations for the associations. SampleTwo datasets: (1) Amazon Sales Dataset with 1,465 product listings used for training/hold-out evaluation; (2) BaSalam Persian-language marketplace with 182,275 product-review pairs used for large-scale out-of-sample evaluation. Features include product descriptions (seller claims), consumer reviews (text signals and mention frequencies), and image-based color cue features. Themesgovernance adoption IdentificationObservational associations and supervised classification: statistical hypothesis tests link sustainability-language mentions to ratings/satisfaction; multimodal supervised ML (text of seller claims, consumer review signals, image color cues) trained on labeled examples with hold-out validation on Amazon and out-of-sample evaluation on BaSalam. No experimental or quasi-experimental causal identification is implemented. GeneralizabilityAmazon training set is small (1,465) relative to marketplace heterogeneity, limiting representativeness for global platforms., BaSalam results are language- and market-specific (Persian marketplace) — consumer behavior, labeling norms, and seller strategies may differ across cultures and platforms., Greenwashing label appears to be inferred from misalignment proxies rather than independent ground-truth enforcement actions, raising measurement validity concerns., Image features are restricted to color analysis; richer visual cues (logos, badges, product composition) are not used, limiting visual generalization., Product-category heterogeneity and time dynamics (seasonality, evolving seller strategies) may limit transferability across categories and time without revalidation.

Claims (7)

Claim	Direction	Outcome	Confidence & Evidence	Details
Sustainability-related language in product reviews and descriptions is not associated with higher product ratings or greater customer satisfaction; functional attributes such as performance, usability, and value dominate evaluations. Consumer Welfare	null_result	Product ratings and customer satisfaction	Reading fidelity high Study strength medium	n=1465 0.3
The frequency of sustainability-related language varies by discount level, with the High Discount category showing the highest frequency of sustainability mentions. Task Allocation	positive	Frequency of sustainability-related mentions in product listings and reviews	Reading fidelity high Study strength medium	n=1465 0.3
A multimodal pipeline combining seller claims, consumer review signals, and image cues can identify likely greenwashing, defined as misalignment between seller sustainability claims and consumer-experienced evidence. Ai Safety And Ethics	positive	Detection of likely greenwashing or claim-experience misalignment	Reading fidelity high Study strength medium	n=183740 0.3
The greenwashing-detection pipeline achieved 81.6% overall accuracy on an Amazon hold-out evaluation. Regulatory Compliance	positive	Classification accuracy for likely greenwashing	Reading fidelity high Study strength medium	n=1465 81.6% accuracy 0.3
The greenwashing-detection pipeline achieved 97% overall accuracy on the BaSalam Persian-language dataset. Regulatory Compliance	positive	Classification accuracy for likely greenwashing	Reading fidelity high Study strength medium	n=182275 97% accuracy 0.3
The results suggest that sustainability claims may function primarily as marketing signals rather than reliable indicators of experienced product quality. Consumer Welfare	mixed	Relationship between sustainability claims and experienced product quality or satisfaction	Reading fidelity medium Study strength speculative	n=1465 0.03
Multimodal AI screening could enable scalable detection of deceptive sustainability claims by platforms and regulators. Governance And Regulation	positive	Scalable detection and screening of potentially deceptive sustainability claims	Reading fidelity high Study strength medium	n=183740 81.6% accuracy on Amazon; 97% accuracy on BaSalam 0.3