Capture–mark–recapture methods have been extensively used to estimate abundance, demography, and life history parameters of populations of several taxa. However, the high mobility of many species means that dedicated surveys are logistically complicated and expensive. Use of opportunistic data may be an alternative, if modeling takes into account the inevitable heterogeneity in capture probability from imperfect detection and incomplete sampling, which can produce significant bias in parameter estimates. Here, we compare covariate‐based open Jolly‐Seber models (POPAN) and multi‐state open robust design (MSORD) models to estimate demographic parameters of the sperm whale population summering in the Azores, from photo‐identification data collected opportunistically by whale‐watching operators and researchers. The structure of the MSORD also allows for extra information to be obtained, estimating temporary emigration and improving precision of estimated parameters. Estimates of survival from both POPAN and MSORD were high, constant, and very similar. The POPAN model, which partially accounted for heterogeneity in capture probabilities, estimated an unbiased super‐population of ~1470 whales, with annual abundance showing a positive trend from 351 individuals (95% CI: 234–526) in 2010 to 718 (95% CI: 477–1082) in 2015. In contrast, estimates of abundance from MSORD models that explicitly incorporated imperfect detection due to temporary emigration were less biased, more precise, and showed no trend over years, from 275 individuals (95% CI: 188–404) in 2014 to 367 (95% CI: 248–542) in 2012. The MSORD estimated short residence time and an even‐flow temporary emigration, meaning that the probability of whales emigrating from and immigrating to the area was equal. Our results illustrate how failure to account for transience and temporary emigration can lead to biased estimates and trends in abundance, compromising our ability to detect true population changes. MSORD models should improve inferences of population dynamics, especially when capture probability is low and highly variable, due to wide‐ranging behavior of individuals or to non‐standardized sampling. Therefore, these models should provide less biased estimates and more accurate assessments of uncertainty that can inform management and conservation measures.