A fundamental challenge in modern mantle geochemistry is to link geochemical data with geological and geophysical observations. Most of the early geochemical models involved a layered mantle and the concept of geochemical reservoirs. Indeed, the two layer mantle model has been implicit in almost all geochemical literature and the provenance of oceanic island basalt (0113) and mid-ocean ridge basalt (MORB) [van Keken et al., Annu. Rev. Earth Planet. Sci. 30 (2002) 493-525]. Large-scale regions in the mantle, such as the 'convective' (i.e. well-stirred, homogeneous) upper mantle, sub-continental lithosphere, and the lower mantle were treated as distinct and accessible geochemical reservoirs. Here we discuss evidence for a ubiquitous distribution of small- to moderate-scale (i.e. 10(2)-10(5) m) heterogeneity in the upper mantle, which we refer to as the statistical upper mantle assemblage (SUMA). This heterogeneity forms as the result of long-term plate tectonic recycling of sedimentary and crustal components. The SUMA model does not require a convectively homogenized MORB mantle reservoir, which has become a frequently used concept in geochemistry. Recently, Kellogg et al. [Earth Planet. Sci. Lett. 204 (2002) 183-202] modeled MORB and OIB Sr and Nd isotopic compositions as local mantle averages of random distributions of depleted residues and recycled continental crustal material. In this model, homogenization of the MORB source region is achieved by convective stirring and mixing. In contrast, in the SUMA model, the isotopic compositions of MORB and OIB are the outcome of homogenization during sampling, by partial melting and magma mixing (e.g. [Helffrich and Wood, Nature 412 (2001) 501-507]), of a distribution of small- to moderate-scale upper mantle heterogeneity, as predicted by the central limit theorem. Thus, the 'SUMA' acronym also captures what we consider the primary homogenization process: sampling upon melting and averaging. SUMA does not require the participation of distinct (e.g. lower mantle) reservoirs to explain OIB compositions. The complementary methods outlined in this work and in e.g. Kellogg et al. and Helffrich and Wood, provide a strong link between geochemical data and potential geophysical observables, such as the degree of partial melting, size of the melting regime, the characteristic length scales of the chemical heterogeneity and bathymetry, seismic scattering and anisotropy. We support the view that statistical distributions of lithologic components and sampling theory should replace the concept of distinct, isolated geochemical reservoirs, and extensive convective stirring prior to sampling. (C) 2003 Elsevier B.V. All rights reserved.