How Trust Signals in Product Reviews Predict Recommendation Behavior: A Behavioral Study Using E-Commerce Data

The rise of e-commerce has made online product reviews a cornerstone of the U.S. digital economy, essentially shaping consumer’s purchasing behavior. However, the integrity of this marketspace is threatened by inauthentic or misleading review content, creating a challenge for consumers who rely on these reviews to make informed decisions. This study investigates how specific language patterns within review texts, termed "trust signals," can predict a consumer's recommendation behavior more effectively than traditional sentiment analysis methods. Firstly, baseline predictive models like Random Forest and XGBoost were developed using numerical data such as product ratings, achieving high accuracy but demonstrating that the models rely heavily on the star rating alone. To address this limitation and analyze the unstructured text, a robust Large Language Model (Mistral-7B) was fine-tuned on a large dataset of e-commerce reviews for a women’s clothing retail brand. The fine-tuned model was able to predict recommendation behavior with over 92% accuracy based solely on the review text. Furthermore, an n-gram analysis identified that specific phrases related to product fit ("true size," "runs small") and quality ("soft comfortable," "fabric soft") were some of the most powerful trust signals. These findings demonstrate that advanced machine learning models can be trained to detect complex, context-rich hints in language, providing a more sophisticated and reliable method for understanding the key elements that builds consumer trust. This research offers a valuable methodology for enhancing the transparency of the digital marketplace, which is vital for the nation's continued economic stability.