We present a complete empirical case study of fundraising decisions that demonstrates the importance of in-context field experiments. We first design novel matching-based fundraising appeals. We derive theory-based predictions from the standard impure altruism model and solicit expert opinion about the potential performance of our interventions. Both theory-based predictions and descriptive advice suggest improved fundraising performance from a framing intervention that credited donors for the matched funds (compared to a typical match framing). However, results from a natural field experiment with prior donors of a non-profit showed significantly poorer performance of this framing compared to a regularly framed matching intervention. This surprising finding was confirmed in a second natural field experiment, to establish the ground truth. Theoretically, our results highlight the limitations of both impure altruism models and of expert opinion in prediction complex "warm glow" motivation. More practically, our results question the availability of useful guidance, and suggest the indispensability of field testing for interventions in fundraising.