It is well known that computer simulation models are used to make projections of future climate change. It is less well known that some of the most-used “observational” datasets in climate science are composed entirely of simulation output. I will explain how such datasets are produced (via a practice known as data assimilation) and consider whether they are really so different from conventional observational datasets. I will argue that the differences are not as great as one might suspect: in principle, these datasets can be high-quality measurements of atmospheric properties, despite their genesis in simulation. In arguing for this conclusion, I will consider several conceptions of measurement as well as the roles of models in measurement. I will also argue, however, that data assimilation is a special case – most simulation studies do not deliver measurements of real-world systems, at least not if we adopt a reasonably demanding conception of measurement.