Real-time optimisation (RTO) systems face challenging scenarios in industrial practice, such as incomplete and corrupted process information, uncertain large-scale mathematical models and numerical optimisation issues. Proper design of RTO structure and robust diagnosis tools are keys for good performance, although they are neglected in commercial RTO software and not fully solved in the technical literature. This article reviews the concepts behind the two-step RTO approach and suggests performance metrics. It also points out the large list of structural decisions and the consequences of intuitive, experience-based RTO design choices. The discussed vulnerabilities are illustrated with some simulations and real industrial implementations. (c) 2012 Canadian Society for Chemical Engineering