Agent-based models (ABMs) have become a valuable tool in social epistemology for addressing a fundamental question: How should scientists communicate? Yet, ABMs often yield conflicting results—some suggest that high levels of communication decrease group accuracy, while others find it beneficial. Why do models differ so dramatically? We argue that these discrepancies arise from a simple fact: different models use different conceptions of ”communication,” and model qualitatively different phenomena. To demonstrate the effects of such differences, we integrate three paradigmatic conceptions of communication from the literature—direct evidence sharing, belief averaging, and testimony exchange—into a unifying simulation. This allows us to test them under identical conditions. Our findings suggest that communication is generally beneficial. However, the effects of communication vary significantly by which conception one adopts, even under identical conditions. This lack of robustness highlights a critical issue: outcomes depend heavily on how communication is modeled.