Abstract For the 2016/2017 central Italy earthquake sequence, various studies reported spatiotemporal variations in the Gutenberg–Richter b‐value, including its unexpected increase before the Mw6.5 Norcia mainshock. Here, we investigate this specific observation with a combination of machine learning techniques. Before the mainshock, we identify two seismicity volumes with asynchronous activity, disparate fault orientations, and distinct b‐values. Throughout the sequence, both structures maintain their distinct b‐values; only when combined, the b‐value apparently increases before the mainshock (and drops afterward). Our observation suggests that a temporal b‐value variation in a large volume, or the entire extent of this sequence, originates from individual structures with distinct b‐values being active at different times. More generally, temporal b‐value variations may also reflect changes in the spatial distribution of seismicity besides other underlying processes. Our findings highlight that b‐value interpretation must acknowledge the structural heterogeneity, such as associated fault segments and their orientations.