Hyperscanning has been hailed as a game-changing method which will allow us to understand the neuroscience of multi-person social interactions and create ‘second person neuroscience’. Here, I present a critical review of fNIRS hyperscanning studies, examining what they can and cannot tell us about social neuroscience. A key problem is that many current hyperscanning methods cannot distinguish a pure pattern of interpersonal coherence from effects driven by a common input. Limited data on participant behaviour during testing sessions compounds this problem, because it is not clear what behaviours might mediate the brain coherence patterns that are reported. Some studies respond to this problem by retreating from strong cognitive interpretations of hyperscanning data and measuring how overall levels of coherence differ with individual factors (such as age, gender, social relationships between people or clinical diagnosis). Here, I suggest that there is a better way to analyse and interpret hyperscanning studies. By tracking behaviour in detail, and analysing behaviour and brain activity patterns together, it will be possible to define what types of action, perception and mutual prediction arising in the interaction of two or more people can lead to coherent brain signals and why. This approach acknowledges that social brains are embodied and that coherent brain activity arises from the social behaviour of two people in an interaction. By integrating our study of the social brain and social behaviour, we will be able to strengthen the science of both.