The potential locations of aftershocks, which can be large and damaging, are often forecast by calculating where the mainshock increased stress. We find, however, that the mainshock-induced stress field is often rapidly altered by aftershock-induced stresses. We find that the percentage of aftershocks that are secondary aftershocks, or aftershocks triggered by previous aftershocks, increases with time after the mainshock. If we only consider aftershock sequences in which all aftershocks are smaller than the mainshock, the percentage of aftershocks that are secondary also increases with mainshock magnitude. Using the California earthquake catalog and Monte Carlo trials we estimate that on average more than 50% of aftershocks produced 8 or more days after M ≥5 mainshocks, and more than 50% of all aftershocks produced by M ≥7 mainshocks that have aftershock sequences lasting at least 15 days, are triggered by previous aftershocks. These results suggest that previous aftershock times and locations may be important predictors for new aftershocks. We find that for four large aftershock sequences in California, an updated forecast method using previous aftershock data (and neglecting mainshock-induced stress changes) can outperform forecasts made by calculating the static Coulomb stress change induced solely by the mainshock.