Reasoning is computationally expensive. This is especially true for reasoning on the Web, where data sets are very large and often described by complex terminologies. One way to reduce this complexity is through the use of approximate reasoning methods which trade one computational property (eg. quality of answers) for others, such as time and memory. Previous research into approximation on the Semantic Web has been rather ad-hoc, and we propose a framework for systematically studying such methods. We developed a workbench which allows the structured combination of different algorithms for approximation, reasoning and measuring in one single framework. As a case-study we investigate an incremental method for instance retrieval through ontology approximation, and we use our workbench to study the computational behaviour of several approximation strategies.