Protein photosensors from all kingdoms of life use bound organic molecules, known as chromophores, to detect light. A specific double bond within each chromophore is isomerized by light, triggering slower changes in the protein as a whole. The initial movements of the chromophore, which can occur in femtoseconds, are tightly constrained by the surrounding protein, making it difficult to see how isomerization can occur, be recognized, and be appropriately converted into a protein-wide structural change and biological signal. Here we report how this dilemma is resolved in the photoactive yellow protein (PYP). We trapped a key early intermediate in the light cycle of PYP at temperatures below -100 degrees C, and determined its structure at better than 1 A resolution. The 4-hydroxycinnamoyl chromophore isomerizes by flipping its thioester linkage with the protein, thus avoiding collisions resulting from large-scale movement of its aromatic ring during the initial light reaction. A protein-to-chromophore hydrogen bond that is present in both the preceding dark state and the subsequent signalling state of the photosensor breaks, forcing one of the hydrogen-bonding partners into a hydrophobic pocket. The isomerized bond is distorted into a conformation resembling that in the transition state. The resultant stored energy is used to drive the PYP light cycle. These results suggest a model for phototransduction, with implications for bacteriorhodopsin, photoactive proteins, PAS domains, and signalling proteins.