Surveys to map the distributions of marine benthic assemblages require sampling strategies that yield information at spatial scales and taxonomic resolutions relevant to research questions. Allocation of sampling effort, however, involves trade-offs between factors including the number of sites sampled, the sampling methods used, the taxonomic resolution achieved, and the primary purpose of the survey. We used data sets from extensive Ocean Survey 20/20 (OS 20/20) benthic surveys of Chatham Rise and Challenger Plateau, New Zealand, to evaluate different approaches to survey design. To check that the survey data would be useful for informing design criteria for other locations, we first compared the variance structure of each data set at a range of spatial scales. Next, we used multiple regression, for univariate measures, and Canonical Correspondence Analysis (CCA), for multivariate community data, to evaluate the proportion of variance in the test data sets that was explained by a suite of oceanographic and seabed variables. These variables were used in three forms: (1) the unclassified data, (2) after non-hierarchical k-means clustering, (3) after hierarchical average-linkage clustering. We also evaluated the design strata that were actually used for the OS 20/20 surveys, and a classification based on variables derived from multibeam echosounder (MBES) acoustic data alone, and then re-evaluated the proportion of explained variance after incorporation of the spatial distance between samples as a predictor variable. We used power analysis to estimate the number of samples needed to detect differences between classes reliably, and evaluated the effect of increased numbers of samples on the precision of estimates for a suite of biodiversity metrics. Finally, we constructed correlograms to evaluate rates of change in assemblage similarity with spatial distance across Chatham Rise only. For most data sets, variances did not differ significantly between or within locations but the magnitude of variances differed with both sampling method and spatial scale. Variance in the OS 20/20 faunal data was best explained by regression against the unclassified environmental variables (24.1% to 48.7% explained depending on metric) but classes derived from the k-means clustering method performed similarly (26.3% to 42.7%). Hierarchical clustering and MBES classes performed poorly by comparison. Increases in the power to differentiate between habitats and the precision of assemblage metric estimates were observed with all increases of sampling effort up to the limit possible using the OS 20/20 data. Spatial distance explained 11% of the total variance, whereas environment alone explained 29 % and only 1% was explained by interaction of the two. We conclude that the Chatham-Challenger OS 20/20 data provide a valuable test set for evaluating different approaches to survey design and provide insights into how broad-scale biodiversity surveys in New Zealand might be planned in future. The results demonstrate that increasing the density of sampling (decreasing sample lag) is likely to be the most effective way of improving the accuracy of biodiversity maps. In practice, this would entail sampling with a smaller number of methods. Of the methods evaluated here, data sets from video and epibenthic sled were the most informative, primarily because they combined the smallest sample lag with fine taxonomic resolution. While environmental data can serve as a sound basis for initial survey design, our analyses show that a proportion of sampling effort should also be allocated purely on the basis of spatial distance between sites, or by reference to characteristics of the area that are not captured by the environmental data.