This study aimed at evaluating the effect of different types of external Human-Machine Interfaces (eHMIs) on cyclists’ crossing behaviour when interacting with an automated vehicle (AV) in a shared space. Using a Virtual Reality (VR) experiment, four eHMI designs (no eHMI, textual, symbolic and lights) were systematically tested throughout a series of crossing scenarios, in which participants engaged in the role of a cyclist using a VR bicycle simulator. Each design was tested under yielding and non-yielding conditions. Cyclist behaviour was recorded through objective measures, such as crossing initiation time, gazing time and crossing intentions, while subjective perceptions were captured in a post-experiment questionnaire. The results indicate that eHMIs do influence cyclists’ crossing behaviour, however not consistently across all behavioural metrics. The textual and light eHMI significantly decreased the crossing initiation time of cyclists when the AV is yielding, while the lights increased the crossing time and gazing time when the AV is not yielding. In contrast, any form of eHMI shows a significant improvement across all subjective measures, enhancing the perceived safety, trust and decision-making of cyclists. The symbolic eHMI consistently received the highest ratings, however, overall differences were minimal. Therefore, a gap is observed between the subjective perceptions and impressions of cyclists and their actual behaviour. This study contributes to the growing literature on cyclist-AV interactions, which is still underwhelming. It forms a solid basis for further research optimizing eHMI design for real-world implications. In addition, it proves that VR is an effective tool for analyzing these interactions in realistic yet controlled environments.