Abstract

In the contemporary social media environment, organizational crisis communication rarely ends after a single response. Crisis events often evolve through event exposure, a first response, continued public fermentation or new information, and a subsequent response by the focal actor. Existing crisis communication research, however, has primarily examined single-response strategies and has paid less attention to how response sequence, response timing, and later information shocks jointly shape public judgment. This study proposes a real-comment-calibrated LLM-Agent framework for simulating multi-round crisis response paths in counterfactual experimental settings. Public agents evaluate each crisis and response along eight dimensions: perceived severity, responsibility attribution, perceived sincerity, perceived transparency, anger, moral disgust/contempt, trust, and diffusion intention. Huaxizi and Zhang Xiaoquan were used as calibration cases, and the Xibei prepared-food controversy was used for external validation. Four formal cases-Baidu Qu Jing's short-video controversy, the East Buy 'little essay' incident, the Three Sheep/Meicheng mooncake false-promotion controversy, and the Pangdonglai cold-noodle hygiene incident-were then tested under 36 two-round response paths generated by denial, explanation, and repair strategies combined with immediate and delayed timing. The results show that the second-round response has a stronger influence on final public judgment than the first-round response. The best-performing paths mostly end with immediate repair, whereas the worst-performing paths concentrate on second-round denial. The findings suggest that responsiveness is not equivalent to responsibility: immediate repair can improve trust and reduce diffusion intention, but immediate denial may be interpreted as an attempt to evade responsibility. The study contributes to crisis communication research by conceptualizing crisis response as a path-dependent dynamic process and by demonstrating the methodological value of calibrated LLM-Agent counterfactual simulation.

Download

Share

COinS