We evaluated the accuracy and precision of three population estimation methods (mark–resight, distance sampling and five-minute bird counts) for two populations of South Island robin (Petroica australis australis) of known size in the Eglinton Valley, Fiordland, over 5 years (March and August, 2005–2009). The performance of these population estimators was compared to known robin abundance derived from simultaneous territory mapping of individually marked birds. Mark–resight methods performed well with Bowden's estimator generating accurate and precise population size estimates and trends very similar to those obtained from territory mapping. Distance sampling estimates displayed significant positive bias and poor precision even though we could identify the general population trends derived from territory mapping at Knobs Flat and Walker Creek. Five-minute bird counts (and associated generalised linear mixed models) performed well when the assumption of constant detectability was met, and poorly when it was not. Such failures prevented robust inference and confounded longer-term trend analyses. As robins are attracted towards stationary observers, we recommend that they be counted from line transects rather than points. Whenever monitoring objectives demand accurate and unbiased estimates of population abundance, the monitoring methods used should explicitly account for incomplete detectability wherever possible.