Appropriate reliance on AI, following system recommendations when correct and overriding them when incorrect, is the central behavioural objective of human–AI interaction research. Despite the importance of this objective, researchers draw on a heterogeneous set of behavioural, self-report, performance-based, and process-level measures, with little principled guidance on their suitability and no shared reporting standard to enable cross-study comparison. This paper presents a systematic review of how reliance is measured across 67 HCI papers published between 2014 and 2026. We develop a taxonomy of four measurement categories and evaluate each against seven methodological dimensions grounded in measurement theory and human factors research. Sixty-one percent of included papers combine two or more measure types, yet the majority report divergent signals across instruments. We identify five structural problems limiting cumulative progress: construct ambiguity across definitions, the conflation of trust and reliance as interchangeable constructs, static measurement of an inherently dynamic construct, an ecological validity gap driven by crowdsourced single-task designs, and measurement fragmentation that prevents cross-study synthesis. We translate this diagnosis into a prescriptive measurement framework providing measure selection guidance by research question type and a minimum reporting standard applicable to all reliance studies.