According to the literature we should expect the city size distribution to either conform to the Zipf's law or to log-normality. We check the city size distribution for thirty-one European countries and for two U.S. censuses and find very scant evidence of adherence to Zipf's law (even for the upper-tail) or to log-normality for all cities. Exogenous location characteristics may determine the city size distribution. But even measuring them is a daunting task. Facing this situation we turn to a natural "lab". This lab is made of a unique data set that contains the universe of city size and location in the region of Bukhara in the 9th century A.D. The morphological homogeneity of the region approximates very well the absence of exogenous characteristics. Cities have developed for twelve centuries (from 3 BC to 9 AD) without sizeable perturbation and in a situation of relative isolation from the rest of the world.  The technology was constant over these centuries and homogeneous across the entire area; which allows us to exclude stochastic growth.  We show that a simple model of location choice à la McFadden explains extremely well the size distribution of cities. We also shows that the transport technology determined the agglomeration of urban settlements. Lastly, we find evidence of persistence of population density from the 9the century to present.