This paper presents a methodology for the assessment of replication accuracy using a multiscale approach. It focuses on the quantification of 'metrological replication accuracy', which is a direct comparison of the topography of a surface and its replica. This methodology uses an analysis of variance and takes into account the variability of the experimental results. To assess replication accuracy, the differences in either one or several roughness parameter values of a surface and its replica are assessed over different scales, using the F-value as an indicator. In order to get a multiscale comparison, the surfaces are examined using 17 cut-off lengths with three types of filters (high-pass, low-pass and band-pass). This methodology enables us to identify the scales at which the topography features are best replicated. Transfer functions of roughness parameters are also observed to quantify the differences between the studied surface and its replica. This methodology is applied to the study of the replication of ground titanium alloy specimens having rough to mirror-like surfaces. It enables us to show that the studied replica material enables the accurate reproduction of rough surfaces but gives inaccurate results for the replication of smooth surfaces.