Several internal multiple prediction methods have been proposed in the past and are still actively being researched. Of those, the Jakubowicz method is currently the most attractive and applied one, as it only uses surface data in a convolution and correlation process, and is therefore computationally relatively cheap. This allows its commercial application on 3D data. But due to several assumptions in the prediction process, it has to rely on adaptive subtraction to correct for potential errors. These errors are differentiated and investigated in more detail. Some of them derive from theoretical shortcomings, such as the incorrect implementation of transmission operators, or when neglecting the structural complexity of a 3-dimensional earth in case of 2D or 1D application. The others are introduced because of the unknown properties and simplifying assumptions that have to be made when working with field data. The most important issues are the unknown source characteristics, noise, and the required spatial sampling. These are also of significance for any other method based on the convolution of wavefields, such as Marchenko-imaging. Any advancement that is made regarding these problems, potentially leads to improved results in their application on field data.