Steps are the number your tracker gets right. Everything downstream, less so

The step count on a fitness tracker is close to the truth for most people most of the time. The figures the same device derives from those steps — active minutes, calories burned — are far looser, and the raw step count itself falls apart when someone walks slowly or in short bursts. Two 2026 validation studies draw that line clearly, and it matters because the derived numbers are the ones people use to judge whether they have "done enough" [s1][s2].

This is the device-accuracy companion to our look at what the step-count data actually predicts about health: here the question is not what a step is worth, but whether the device counted it right.

Steps agree; calories do not

A free-living study in Japan had 36 office workers wear consumer devices alongside a research-grade ActiGraph monitor over three weeks [s1]. For step counts, agreement was reasonable: the Apple Watch and Oura Ring landed within 10% of the reference (mean differences of 2.12% and −6.24%), while a Fitbit overestimated steps by 18.00% [s1]. Correlations for step count were strong across devices (r = 0.84–0.92) [s1].

The derived metrics were another matter. Moderate-to-vigorous physical activity was underestimated by 46.22% on one device and 11.64% on another, though the Fitbit was close on that measure (0.62%) [s1]. Physical-activity energy expenditure showed the largest discrepancies of all: the Fitbit overestimated it by 139.19% and the Apple Watch by 25.91%, while the Oura Ring underestimated it by 16.87% [s1]. The study found proportional bias — errors growing at higher activity levels — in the Fitbit's energy figure and the Apple Watch's active-minutes figure [s1]. A calorie readout that runs more than double the truth is not a rounding error; it is the same energy-expenditure ceiling behind the calorie overestimates in budget fitness bands.

Where even the step count breaks

A step count's reliability also depends on how a person moves. A laboratory study validated a Fitbit Charge 6 against video observation in 14 adults with lung cancer, chosen because slow, fragmented gait stresses step-counting algorithms [s2]. Across 126 walking trials the device undercounted by only a small absolute margin — a bias of 1–3 steps — but the relative error was large and depended on how long the person walked: a mean absolute percentage error of 50.6% for five-second bouts, falling to 16.3%–18.7% for 15-to-30-second bouts [s2]. Overall agreement was low, with an intraclass correlation of 0.21 [s2].

Walking speed mattered even more. Undercounting was worst below 0.6 metres per second — a bias of roughly 13 steps, a 56.6% error — and was minimised near a normal pace of 1.0–1.2 metres per second [s2]. In plain terms, the device counts a healthy adult's brisk walk well and a frail or recovering person's slow, stop-start walk badly — which is precisely the population for whom the number is often being used clinically.

The lab study also tested whether the device could classify a minute as active or sedentary, and whether it invented steps during non-walking activity [s2]. Both are where wearable "active minutes" come from, and both are error-prone at the margins: the device's overall step agreement was weak (intraclass correlation 0.21), and its accuracy depended heavily on whether the person was walking long enough and fast enough to look like walking to the algorithm [s2]. A device that reads a healthy adult's gym session well can misread a hospital corridor shuffle, which is the setting where a clinician might most want to trust it.

Why the derived numbers drift

The pattern across both studies is that error compounds as you move from raw counts to inferences. A step is a discrete event a sensor can detect fairly directly; a "moderate-to-vigorous minute" requires the device to judge intensity, and a calorie requires it to model metabolism — each layer adding assumptions. That is why the free-living study found strong step correlations (r = 0.84–0.92) but much weaker agreement for active minutes and energy expenditure across every device [s1]. The proportional bias it detected — error growing at higher activity — means the numbers are least reliable for the most active users, exactly when the totals look most impressive [s1]. The direction of the error also differs by brand and by metric, so one device overcounts calories while another undercounts active minutes, with no single correction a user could apply.

What it means for a user

Trust the step count as a day-to-day trend for ordinary walking, and treat the calorie and "active-minutes" figures as rough at best — the evidence shows energy-expenditure error can exceed 100% and varies by device [s1]. Be especially wary of a slow walker's step count: if gait is slow or movement comes in short bursts, the device may undercount substantially, so a low reading can reflect the algorithm rather than the person [s2]. As with the other consumer metrics, the same device measured consistently is more informative than the absolute number, and no two brands should be assumed to agree [s1].