One implicit assumption underlying the use of digital avatars is that human-avatar-interaction mimics human-human-interaction, especially when similarity between human and avatar is high. Using a level 1 visual perspective taking (L1 VPT) task (Samson et al., 2010) in an online experiment (N = 102), we tested the hypothesis that perceived similarity in terms of anthropomorphism strengthens L1 VPT, causing higher altercentric and egocentric intrusions for human avatars as compared to robots. Intrusions were measured through reaction times and error rates. The design was manipulated within subjects: 3 (avatar: human, humanoid, roboter) x 2 (human’s and avatar’s perspective: consistent vs. inconsistent) x 2 (task: apply the avatar perspective vs. self-perspective). Against the prediction, we found comparable degrees of intrusions for human, humanoid, and robot avatars. However, while for human avatars intrusions were unrelated to the applied perspective, for robot avatars, intrusions increased under the avatar perspective as compared to the self-perspective. We suggest that for robot avatars, calculating responses is not an automatic process, but rather only takes place after the controlled, conscious process of selecting the relevant perspective. Also, the role of attentional cues like gaze direction will be discussed.