BACKGROUND: Virtual Reality exposure therapy (VRET) is an efficacious treatment of phobias and allows for automatic data collection during standardized yet naturalistic exposure paradigms, yet next to no research has explored the clinical potential of such data. GOALS: To explore the predictive potential of using gaze proxy data collected during VRET to predict treatment outcomes. METHODS: Gaze focus proxy data from n=29 participants undergoing gamified, self-help VRET for spider phobia were extracted, compiled, and modeled. The VRET session featured eight levels with increasingly frightening spiders, each with an approach task requiring participants to keep looking at the phobic object for a specified time. Relative spider gaze focus was defined as time spent looking at each spider at each level (derived from head movement and overlap) divided by total time in level. High- versus low-improvement was defined using median-split on subsequent improvements on an in-vivo behavioral approach task. RESULTS: During the final three levels of the exposure session, relative spider focus time was initially lower among high-improvers (p=.039) and the decrease was lower over levels than among the low-improvers (p=.029). This suggests that non-improvers experienced a fear level mismatch during the final exposure phase. There were no differences in gaze patterns during other phases of the session. CONCLUSIONS: These preliminary results suggest that gaze proxy data automatically collected during VRET, even when rudimentary, can be used to predict treatment outcomes, and may thus be used to automatically personalize the exposure design of VRET self-help applications during actual use, to increase efficacy.