AbstractWe present an intercomparison of analyses of the main FGGE level IIb dataset with three advanced analysis systems. the aims of the work are to estimate the extent and magnitude of the differences between the analyses; to identify the reasons for the differences; and, finally, to estimate the significance of the differences for forecast skill. We restrict ourselves primarily to a consideration of the extra‐tropical analyses. the subject of tropical analyses merits separate treatment.We discuss objective evaluations of analysis quality, such as fit to observations, statistics of analysis differences, and mean fields. In addition we place substantial emphasis on subjective evaluation of a series of case studies which were selected to illustrate the importance of different aspects of the analysis procedures, such as quality control, data selection, resolution, dynamical balance, and the role of the assimilating forecast model. In some cases the forecast models are used as selective amplifiers of analysis differences to assist in deciding which analysis was more nearly correct in the treatment of particular data. In part II we consider the overall impact of the analysis differences on forecast quality, and the implications for predictability.In general the analysis systems draw reasonably well to the data, although each system has its own characteristics in this regard. the root mean square differences between the analyses are of the expected order of magnitude, although there are clear differences in the closeness of agreement between different pairs of analyses. Systematic differences arising from particular components of the assimilation suites can be identified.The discussion of the case studies highlights those areas where differences of approach to the analysis problem have led to significant differences in the analyses. Some of the case studies suggest strongly that analysis differences in the vicinity of active baroclinic zones are of particular importance. In order to validate these suggestions in one case, we present an experiment where one analysis is transplanted, locally, into another, to show that large differences in the medium‐range forecasts are attributable to localized differences in the analyses.The results presented here, together with the forecast verification results of part II, suggest that in some cases uncertainties in the analysis are a major contributor to the loss of forecast skill in the medium ra