Dogs are highly successful at detecting disease using olfaction; however, performance varies among dogs. Training detection dogs represents a substantial investment of time and resources; thus, identifying behavioral markers of a successful detection dog and factors that affect dog performance is needed. In a previous study, one of four dogs was able to distinguish between nasal and saliva swabs from healthy cattle and cattle undergoing an inflammatory response at a rate greater than chance, with 73 % accuracy. The present study evaluated how dogs’ accuracy changed throughout training and testing, whether dogs’ responses were related to illness signs present in cattle, and whether dogs’ accuracy was related to their own behavior on a per-trial and per-sample basis. Trial number, or the order of a detection trial within a 20-trial training or test session, was slightly positively correlated with accuracy for one dog. Sample number, the order in which the dogs encountered samples, was slightly positively correlated with per-session accuracy for all dogs. Dog accuracy declined across training sessions that used the same sample (F4381 = 3.48, P = 0.01), suggesting a deterioration in sample quality over time. Among the clinical illness signs measured in cattle, the strongest association demonstrated that greater changes in cattle body temperature were positively associated with dog accuracy (r = 0.32, P < 0.01). During training and testing, positive (sickness-model), negative (healthy) and blank (unscented) samples were presented at equal rates. Dogs spent more time sniffing and made more visits to investigate the positive samples than negative samples, and spent more time sniffing and made more visits to negative samples than blank samples, indicating that less time was required for dogs to rule out unscented stations than to distinguish between cattle swabs. Visits to the negative sample and visits to the positive sample were negatively (r = −0.581, P < 0.01) and positively (r = 0.761, P < 0.01) correlated with accuracy, respectively. Increased latency to search was negatively correlated with accuracy (r = −0.10, P < 0.01). This represents another potential behavioral marker of successful detection dogs.