Although observers know about the law of reflection, their intuitive understanding of spatial locations in mirrors is often erroneous. Hecht et al. (2005) proposed a two-stage mirror-rotation hypothesis to explain these misconceptions. The hypothesis involves an egocentric bias to the effect that observers behave as if the mirror surface were rotated by about 2° to be more orthogonal than is the case. We test four variants of the hypothesis, which differ depending on whether the virtual world, the mirror, or both are taken to be rotated. We devised an experimental setup that allowed us to distinguish between these variants. Our results confirm that the virtual world — and only the virtual world — is being rotated. Observers had to perform a localization task, using a mirror that was either fronto-parallel or rotated opposite the direction of the predicted effect. We were thus able to compensate for the effect. The positions of objects in mirrors were perceived in accordance with the erroneous conception that the virtual world behind the mirror is slightly rotated and that the reconstruction is based on the non-rotated fronto-parallel mirror. A covert rotation of the mirror by about 2° against the predicted effect was able to compensate for the placement error.