This paper develops methods of investigating the existence and extent of cointegration in fractionally integrated systems. We focus on stationary series, with some discussion of extension to nonstationarity. The setting is semiparametric, so that modelling is effectively confined to a neighbourhood of frequency zero. We first discuss the definition of fractional cointegration. The initial step of cointegration analysis entails partitioning the vector series into subsets with identical differencing parameters, by means of a sequence of hypopthesis tests. We then estimate cointegrating rank by analysing each subset individually. Two approaches are considered here, both of which are based on the eigenvalues of an estimate of the normalised spectral density matrix at frequency zero. An empirical application to a trivariate series of oil prices is included.