Decoding virtual influencer endorsement using machine learning: The role of virtual influencer, posting, and disclosure characteristics

Abstract

Virtual influencers are promising marketing tools, which have increasingly provided endorsements on social media platforms. However, salient empirical research remains insufficient, and how to effectively promote users' social media engagement with virtual influencer endorsements is unclear. This study draws on expectation violation theory to develop a social media engagement determinants model for virtual influencer endorsement and explores how virtual influencer characteristics, posting characteristics, and disclosure characteristics impact users' social media engagement. Using data collected from 9665 virtual influencer endorsement posts on Instagram, we employed 10 different sets of machine learning algorithms to train and test the model. Our analysis reveals that LightGBM is the best-performing algorithm for predicting social media engagement. Combining the LightGBM and Shapley Additive Explanations (SHAP) models, we find that anthropomorphism is the most important predictive factor and is negatively related to social media engagement, followed by text length, AI identity declaration, brand disclosure, and item name disclosure. In contrast, basic information cues, persona, purchase channel disclosure, content dynamism, and co-creators’ followers have significant positive effects on social media engagement. Our findings provide theoretical implications for influencer endorsement research and offer practical implications for marketers to optimize virtual influencer endorsement strategies in terms of virtual influencer selection and design, posting design, and information disclosure.