Background The measurement of blood glucose in critically ill patients is still performed in many ICUs with glucose meters and capillary samples. Several prevalent factors in these patients affect the accuracy of the results and should be interpreted with caution. A weak recommendation from the Surviving Sepsis Campaign (SSC) suggests the use of arterial blood rather than capillary blood for point of care testing using glucose meters. Aims and objectives To analyse the agreement between arterial, central venous, and capillary blood samples of glucose values measured by glucose meter in critically ill patients and study potential confounding factors. Design Prospective cross‐sectional study in a general intensive care unit (ICU). Patients needing insulin treatment (subcutaneous or intravenous) and blood glucose control were included. Methods Standardized collection of blood samples and measurement of glucose values with a glucometer. Agreement was studied by the Bland–Altman method and stratified analysis of disagreement‐survival plots was used to study the influence of haematocrit, pH range, SOFA score, capillary refilling time, intravenous insulin infusion, and lactic acid. Results A total of 297 measurements from 54 patients were included. The mean arterial blood glucose was 150.42 (range 31–345 mg/dL). In the graphical analysis, there is a poor agreement both in capillary and venous central to arterial samples, but in opposite direction (underestimation of capillary and overestimation of central venous). Factors associated with a reduction in the agreement between arterial and capillary samples were elevated lactate, poor capillary refilling, and hemodynamic failure. Patients without hemodynamic compromise have an acceptable agreement with values for absolute differences of 16 mg/dL for a disagreement of 10%. Conclusions In critically ill patients, the measurement of blood glucose with a glucose meter should be performed with arterial samples whenever possible. Capillary samples do not accurately estimate arterial blood glucose values in patients with shock and/or vasoactive drugs and underestimate the values in the range of hypoglycemia. Venous samples are subject to error because of potential contamination. Relevance to clinical practice This study adds support to the recommendation of using arterial blood rather than capillary or venous blood when using glucose meters in critically ill patients, especially in those with hemodynamic failure.