The aim of the Response Surface Methodology (RSM), originally designed for chemical, physical and biological applications and experiments, but then extended also to simulations in the industrial field, is the construction of reliable response surfaces characterized by high adherence to the experimental data describing the reality being studied [1]-[2]-[11]. In order to achieve this result the most important scholars in this fields focus their attention on experimental projects capable of providing regression meta-models, that fit well the initial experimental data and have a sort of stability of the width of the confidence intervals on the average response and the prediction intervals [11]. At the same time it is also necessary to choose experimental projects allowing a good estimate of the Experimental Error divided into the two components called pure error and lack of fit by the literature on this subject. One of the key assumptions of the conventional RSM approach is that the Experimental Error strictly linked to the system under examination is fixed and it cannot be controlled by experimenters. Nevertheless, for applications in complex industrial plants, it is not possible to carry out the experimental phase directly on the real system and the object of study needs to be transferred into a simulation model [3]-[8]. Therefore, this assumption becomes meaningless. The transfer of the real physical model into a simulation model implies a substantial change in the nature of the Experimental Error, which, compared to the one in the conventional experiments, changes from fixed to dependent on the length of the simulation run and hence time-variant [4]-[6]-[7]-[9]-[10]. Therefore, it is no longer enough to try to improve the quality of the response surface using the traditional RSM concepts of Optimal Variance, Orthogonality and Rotatability, as these are strictly dependent on the value of the experimental error. In the applications on the discrete and stochastic simulation models experimenters must hence try to reduce, each time it is possible, the magnitude of the simulation model’s outgoing Experimental Error extending the simulation run and consequently the computation time. Only once this has been done will it be possible to fine-tune the study examining the response surface to find the aforementioned properties whose influence on the quality of the outgoing surface is substantially lower compared to the reduction of the Experimental Error. This article illustrates an application to an industrial case that not only shows the validity of what is affirmed, but also highlights other limits linked to the use of the conventional RSM approach to simulation experiments on complex industrial plants. These limits challenge the efficacy of linear regression meta-models as descriptors of the relation between the independent and dependent variables considered.