Seed dormancy is thought to be a key mechanism allowing annual plants to spread extinction risk in unpredictably varying environments. Theory predicts increasing germination fractions with increasing probability of reproductive success but solid empirical evidence is scarce and often confounded with environmental factors. Here we provide an empirical test of bet-hedging via delayed germination for three annual plant species along a ‘predictability gradient’ in Israel. We excluded confounding environmental and maternal effects by raising inbred seed families and germinating them under controlled conditions. Additionally, we germinated field-collected seeds in three consecutive seasons to compare their germination with inbred families where maternal effects were removed. Risk of reproductive failure was quantified using demographic data from the field and from second-generation inbred lines raised in a rainfall gradient in the greenhouse. Our findings were consistent with bet-hedging theory in that germination fraction was negatively related to species- and site-specific risk of reproductive failure. Both field and hand-raised seeds of one species exhibited higher dormancy with increasing risk of reproductive failure across sites, and hand-raised seeds of another species showed the same pattern. The third species exhibited a rather random pattern of germination between years and sites, corresponding to the lack of site-specific risk of reproductive failure. Species-specific patterns of dormancy and risk could be related to alternative risk-spreading strategies such as high adult survival, but were also affected by phylogeny. We provide strong empirical evidence for seed dormancy being a mechanism to reduce the risk of reproductive failure in highly variable environments, but a larger number of rigorous experimental tests of bet hedging germination are needed. Specifically, the genetic basis of bet-hedging must be shown in species with different life histories, for demonstrating that dormancy is adaptive and how it is modified by other risk-spreading traits.