Reconstructing past changes in global mean surface temperature (GMST) is one of the key contributions that palaeoclimate science can make in addressing societally relevant questions and is required to determine equilibrium climate sensitivity (ECS). Previous work has suggested that the temperature of the deep ocean (Td) can be used to determine GMST with a simple Td-GMST scaling factor of 1 prior to the Pliocene. However, this metric lacks a robust mechanistic basis, and indeed, such a relationship is intuitively difficult to envisage given that polar amplification is a ubiquitous feature of past warm climate states and deep water overwhelmingly forms at high latitudes. Here, we interrogate whether and crucially, why, this relationship exists using a suite of curated data compilations generated for key deep-time climate intervals as well as two independent sets of palaeoclimate model simulations. We show that models and data are in full agreement that a 1:1 relationship is a good approximation. Mechanistically, both sets of climate models suggest that i) increasingly seasonally biased deep water formation, and ii) a faster rate of land versus ocean surface warming are the two processes that act to counterbalance a possible polar amplification-derived bias on Td-derived GMST. Using this knowledge, we interrogate the quality of the existing deep ocean temperature datasets and provide a new Cenozoic record of GMST. Our estimates are substantially warmer than similar previous efforts for much of the Paleogene and are thus consistent with a substantially higher-than-modern ECS during deep-time high CO2 climate states.