No universally accepted definition of atopic dermatitis (AD) exists in epidemiological studies, and no objective test that can unequivocally confirm the diagnosis. The heterogeneity of AD is now broadly accepted, but there is no consensus on what the best approach is to disaggregate this complex condition, and how best to identify different AD endotypes. One approach is to use data-driven techniques to uncover the unobserved (ie latent) structure in the data set(s) to identify homogenous groups of individuals in the heterogeneous AD population. Since AD usually starts early in life, and its clinical features may progress, remit or relapse over time, analyses assessing trajectories related to their presence or absence over time may help us understand AD heterogeneity. However, studies that used unsupervised methods such as latent class analysis to discover AD ‘phenotypes’ have reported inconsistent findings. If we are to use data-driven analyses to uncover patterns of AD with different long-term consequences, then the discovered ‘phenotypes’ must be consistent and reproducible. We therefore need to understand the reasons for the inconsistencies between different studies. We suggest that in addition to capturing a simple presence or absence of symptoms, more detailed clinical features, such as the severity and medication used, should be assessed in longitudinal studies. Ultimately, the objective is to move from ‘deep phenotyping’ to more informative genetic studies, followed by functional studies to understand mechanisms. In this way, the identification of novel therapeutic targets could be facilitated.