Prototype faces, created by averaging faces from several individuals sharing a common characteristic (for example a certain personality trait), can be used for highly informative experimental designs in face research. Although the facial prototype method is both ingenious and useful, we argue that its implementation is associated with three major issues: lack of external validity and non-independence of the units of information, both aggravated by a lack of transparency regarding the methods used and their limitations. Here, we describe these limitations and illustrate our claims with a systematic review of studies creating facial stimuli using the prototypes dataset 'Faceaurus'. We then propose some solutions that can eliminate or reduce these problems. We provide recommendations for future research employing this method on how to produce more generalisable and replicable results.