Current assessments of research impact have been criticized for capturing what can be easily counted not what actually counts. To empirically examine this issue, we approached measuring research impact from two directions, tracing forwards from research and backwards from policy, within a defined research-policy system (childhood obesity prevention research and policy in New South Wales, Australia from 2000 to 2015). The forward tracing research impact assessment component traced a sample of 148 local research projects forward to examine their policy impacts. Of the projects considered, 16% had an impact on local policy and for a further 19%, decision-makers were aware of the research, but there was no evidence it influenced policy decisions. The backward tracing component of the study included an analysis of research use across three policy initiatives. It provided a more nuanced understanding of the relative influence of research on policy. Both direct uses of specific research and indirect uses of research incorporated as broader bodies of knowledge were evident. Measuring research impact from both directions captured the diverse ways that research was used in decision-making. Our findings illustrate complexities in the assessment process and in real-life policymaking trajectories. They highlight the role that timing of assessment plays in perception of impacts and difficulties attributing longer-term impacts to specific research. This study supports the use of models where politics and complex system dynamics shape knowledge and its influence on decision-making, rather than research being the primary driver for policy change.