As artificial intelligence (AI) advances, its role in online mental health therapy is attracting growing attention. This study uses a quantitative 2x2 factorial design to examine how AI transparency, theory of change (ToC), therapy advice style, AI acceptance, and type of mental health issue affect user perceptions of AI-driven mental health chatbots. Combining quantitative analysis with sentiment and emotional text mining, the research explores how these factors shape perceived empathy, satisfaction, and treatment outcomes. Results show that users aware of interacting with AI report more positive experiences, especially with an emotional ToC. Emotional advice fosters deeper engagement, while rational advice generates more positive sentiment. Emotional tone and dynamics also differ by topic, with depression discussions showing greater intensity. These findings highlight the need to tailor chatbot communication to user expectations and emotional needs, informing the design of more personalised mental health technologies.