Aim Estimating species richness from a series of samples is an important and widely debated issue in ecology and biodiversity conservation. Numerous tests of respective richness estimators gave insights into the precision, the limitations and the pitfalls of richness forecasting. However, few benchmark tests used almost complete empiric census data obtained at those spatial scales where richness estimation is most useful for conservation management.Location Japan.Methods We use an extraordinary dataset on the spatial distribution of Japanese plants containing complete information on the occurrence of each Japanese plant species at the 10 x 10 km(2) grid cell level. We link the estimates of four estimators representing different theoretical approaches, Chao2, rarefaction, species-area relationships (SAR) and species abundance distributions (SAD), to environmental data using a fully nested sampling design.Results Chao2 and rarefaction behaved very similar in all tests and significantly underestimated true richness below 40% sampling fraction. SAR and SAD were less precise than Chao2 and rarefaction at higher sampling fraction but also less affected by low sample size. In general, SAD provided robust estimates over the whole range of sampling fraction and 67.4% of estimates ranged within the 10% error level. Higher species spatial turnover increased and high evenness in occurrence decreased the precision of the SAD estimator. Precision of the four estimators was largely unaffected by environmental variability but increased with increasing latitude.Main conclusions Our results strongly indicate that the pattern of Japanese plant species spatial distribution is sufficiently scale invariant for richness estimators to provide precise forecasting results at the country level. The simplest process to generate such a spatial distribution is ecological drift.