A key physical quantity during reionization is the size of HII regions. Previous studies found a characteristic bubble size which increases rapidly during reionization, with apparent agreement between simulations and analytic excursion set theory. Using four different methods, we critically examine this claim. In particular, we introduce the use of the watershed algorithm -- widely used for void finding in galaxy surveys -- which we show to be an unbiased method with the lowest dispersion and best performance on Monte-Carlo realizations of a known bubble size PDF. We find that a friends-of-friends algorithm declares most of the ionized volume to be occupied by a network of volume-filling regions connected by narrow tunnels. For methods tuned to detect the volume-filling regions, previous apparent agreement between simulations and theory is spurious, and due to a failure to correctly account for the window function of measurement schemes. The discrepancy is already obvious from visual inspection. Instead, HII regions in simulations are significantly larger (by factors of 10-1000 in volume) than analytic predictions. The size PDF is narrower, and evolves more slowly with time, than predicted. It becomes more sharply peaked as reionization progresses. These effects are likely caused by bubble mergers, which are inadequately modeled by analytic theory. Our results have important consequences for high-redshift 21cm observations, the mean free path of ionizing photons, and the visibility of Ly-alpha emitters, and point to a fundamental failure in our understanding of the characteristic scales of the reionization process.

replaced to match published version in MNRAS, conclusions unchanged