Simple survey methods for small mammals, such as indices of trap captures per unit effort, are often the only practicable means of monitoring populations over the long term and at landscape scale and the only source of valuable historical data. They include two fundamental assumptions about the target populations (uniform distribution and equal detectability). Concern has often been expressed that, if these assumptions are violated, conventional density indices could give misleading results. Site occupancy analysis (SOA) can detect significantly uneven distribution of local populations (from variation in probability of occupancy) and reliability of indices of abundance (from variation in detectability) without requiring enumeration. We use this method to examine standardised capture records from long-term population surveys of non-commensal house mice (Mus musculus), ship rats (Rattus rattus), Norway rats (Rattus norvegicus) and stoats (Mustela erminea), sampled in four representative temperate forest habitats in New Zealand. Best fit models generated by SOA were consistent with (1) constant or random probability of occupancy for stoats and dynamic equilibrium probability of occupancy for most populations of mice and rats; (2) widespread site-specific variation in probability of detection, especially substantial in rats and correlated with habitat covariates; (3) direct correlations between detectability and density index in mice and rats sampled at 50 m intervals over 3 days, probably because the effects on the density index of variation in numbers available to be caught (population size) were much larger than the effects of changes in catchability (individual behaviour); (4) declines after 6 days in detectability of stoats and rats sampled at 3–400 m intervals over 10 days, attributed to a local trap-out effect. Longer-term variations in the density index were consistent with observed changes in reproductive parameters and age structure that are known to follow variations in real numbers. We conclude that violations of the assumptions of uniform distribution and equal detectability, while real, were not sufficient to prevent these data from providing information adequate for (1) short-term population assessments (2) long-term, low-level monitoring and (3) preliminary modelling.