Evaluation of energy consumption and device lifetime in battery-powered wireless sensor networks (WSN) is almost exclusively based on measurements or simulations of the total charge (i.e. total mA-h) consumed by the device. In reality, batteries are complex electro-chemical systems and their discharge behavior depends heavily on the timing and intensity of the applied load. However, there is very little empirical data or reliable models available for the kinds of batteries and loads that are typically used in WSNs. The effect of battery dynamics on sensor lifetime is therefore not well understood. We characterize CR2032 Li coin cells using carefully controlled synthetic loads and a wide range of WSN-typical load parameters. Our results are the first to quantify in-depth the discharge behavior of primary batteries in the WSN context. We report that in some common cases, observed lifetimes can differ from predicted ones by almost a factor of three. Furthermore, loads with similar average currents -- which would be expected to have similar lifetimes -- can vary significantly in the amount of capacity they can utilize, with short duration loads generally faring better. The results show that energy evaluation based on ``total mA-h consumed'' has significant limitations. This has important implications for the design and evaluation of WSN applications, as well as for important practical problems in network dimensioning and lifetime estimation.