Research impact and its measurement are of increasing importance. This is particularly significant for agricultural science, which is expected to produce solutions to future challenges that will arise from population growth, climate change and ecosystem degradation. Much econometric effort has been devoted to analysis of investment in agricultural research and its effects on farm productivity. This analysis, reviewed here, has produced a consensus suggesting that returns are high, although they are achieved only after long lags. However, policy‐makers perceive the occurrence of impacts as too few, and poorly targeted with respect to their needs. An attribution gap between the outcomes of agricultural research and how they reach farmers has motivated evaluation of the process of transmission and translation of agricultural research outputs into ultimate impacts. This gap can be narrowed by Participatory Impact Pathway Analysis, implemented mostly so far in low income countries. However, it is a costly and cognitively complex approach. Content analysis of the UK's 2014 REF Impact Case Studies uncovers the mind set of researchers and their managers regarding the description of impact and how it is supposed to occur. This reveals a nascent conservatism that focuses on research that can be shown to have impact, rather than research impact itself. From the overall discussion it can be concluded that the impact evaluation of agricultural science raises more profound issues than simply efficiency or transparency. Confirmation bias threatens impact evaluation, principally by distracting from other important stories about how and why the ultimate effects occur, but also by transforming the nature of the process itself. Methodological pluralism, with greater integration and triangulation between different evaluation approaches, is a promising means of resolving these problems.