Fitness tracker accuracy: steps, distance, GPS, and calories tested

A fitness tracker is a bundle of estimates. Some are very good. Some are bad enough that you should stop quoting them to your friends. The trick is knowing which is which, because the watch displays all of them with the same confident two-digit precision and never tells you that one number is a measurement and another is the output of a regression.

This piece is the consumer-facing version of our broader wearable sensor accuracy writeup, which covers HR, HRV, and sleep staging. Here we focus on the numbers you look at on the watch face during a workout: steps, distance, GPS traces, workout heart rate, calories, swim metrics, and the VO2max readout. We have tested most current devices over the last year. Where our tests and the published studies agree, we say so. Where they don’t, we name the disagreement.

Steps: the one number that mostly works

The pedometer was the first piece of consumer fitness electronics and is still the most reliable. A wrist-worn accelerometer samples motion at 25-100 Hz, filters non-step frequencies, and counts peaks. Walking on flat ground at a normal cadence produces a clean periodic signal. Counting it is not a hard problem.

Our bench against a hand counter on six current trackers over a 1,000-step walk: every device was within 30 steps of truth, five were within 15. Apple Watch Series 9 and Garmin Vivoactive 5 tied for closest. This matches Case et al. (2015) in the Journal of the American Medical Association, which found wrist-worn devices within roughly 5% of the criterion count for steady walking. Almost a decade of subsequent firmware has narrowed that gap rather than widened it.

Failure modes are predictable. Pushing a stroller is the worst case we have measured: both hands on the bar, the wrist barely moves, and step counts come in 30-50% low across every device we tried. Carrying a grocery bag in one hand swings the loaded arm differently from the free arm — undercounts of 15-20% on the loaded side, slight overcounts on the free side. Holding a phone to your ear suppresses the arm-swing signal on that wrist with similar results.

Phantom steps from car rides are still a thing. Not on the Apple Watch, which has used vehicle-motion detection since watchOS 5, and not on current Garmins. We have seen it on a 2025-model Xiaomi Mi Band on a rough Pennsylvania county road, where a 40-minute drive added 1,800 steps to the daily total. The bands and watches that get this right are in our best fitness trackers roundup.

Distance: where the wrist alone is bad

A fitness tracker computes distance in one of two ways. Wrist-only: stride length times step count, with the stride length either a height-based default or a learned average from past GPS workouts. Or GPS, from the watch or a paired phone.

Wrist-only distance is wrong by 5-15% in our testing. Stride length is not constant. It changes with pace, surface, fatigue, and whether you are on a hill. A learned average is closer than a default but is still an average applied to instantaneous motion that does not match the average. For indoor treadmill runs without manual calibration, errors of 10% in either direction are common. We have seen a Fitbit Charge 6 report 5.2 km for a measured 5.0 km treadmill run, and the same watch report 4.6 km for the same run a week later. The treadmill did not change.

GPS distance is much better, and the gap between single-band and multi-band GPS is the most important fitness-tracker hardware story of the last three years.

GPS: single-band vs L1+L5

Consumer GPS until roughly 2022 was almost universally single-frequency, running on the L1 civilian signal at 1575.42 MHz. It works. It is also vulnerable to multipath: the signal reflects off buildings, cliffs, and forest canopy, and the receiver sees both the direct path and the reflection, biasing the position fix. In open sky, a single-band wrist GPS is accurate to within a few meters. In Manhattan or on a tree-covered trail, errors of tens of meters are routine and the trace looks visibly wrong on a map.

Multi-band receivers (often labeled “dual-frequency” or “L1+L5”) pick up a second civilian signal at 1176.45 MHz. L5 is engineered for civilian use, has a wider bandwidth, and is much harder to confuse with a reflection. The Apple Watch Ultra and Ultra 2, Garmin Fenix 7 and 8, Garmin Forerunner 955/965, and the Coros Vertix 2 all run multi-band. The difference shows up most where single-band gets confused.

The practical effect: on the C&O Canal towpath near Washington under closed tree canopy, a Garmin Forerunner 45 (single-band) and an Apple Watch Ultra 2 (multi-band) measured the same out-and-back run as 9.8 km and 10.06 km respectively. The true distance, from a wheel-measured trail mile marker, was 10.04 km. Single-band was 2.4% low; multi-band was within 20 meters.

If you are buying a watch primarily for running, and your runs are anywhere with tree cover, hills, or buildings, multi-band is worth the price increment. On a track or in open suburban streets, single-band is fine and the extra battery life of the older silicon is a real benefit.

Calorie burn: the worst number on the screen

The calorie figure on your fitness tracker is, with a small handful of exceptions, not trustworthy in absolute terms. It is sometimes useful as a relative number to compare your own workouts to each other, on the same device, when nothing else has changed. As an actual energy-expenditure measurement that you should use to decide what to eat, it is bad.

The most cited piece of evidence is Shcherbina et al. (2017), “Accuracy in Wrist-Worn, Sensor-Based Measurements of Heart Rate and Energy Expenditure in a Diverse Cohort,” in the Journal of Personalized Medicine, volume 7. The Stanford group, working with Euan Ashley’s lab, tested seven wrist devices (Apple Watch, Fitbit Surge, Basis Peak, Microsoft Band, Mio Alpha 2, PulseOn, and Samsung Gear S2) against laboratory-grade indirect calorimetry across walking, running, and cycling. Median heart-rate error across devices was 2-5% for most activities, and the Basis Peak logged the lowest resting-HR error of the group — a result we discuss in basis peak heart-rate accuracy.

The energy-expenditure result was the headline. None of the seven devices met the 10% threshold the authors set as the bar for clinical usability. Errors ranged from 27.4% on the best device (the Fitbit Surge) to 92.6% on the worst (PulseOn). The Apple Watch came in at 40.4%. The same device that was reading heart rate within 5% of an ECG reference was producing energy estimates off by 40%, because the conversion from “we know your HR and motion” to “you burned X calories” requires assumptions the device cannot validate.

The reason is structural. Energy expenditure on a wearable is a function. Inputs: heart rate, motion intensity from the accelerometer, age, weight, sex, height, and your resting heart rate if the watch has learned it. The function combines those into a MET estimate and multiplies by weight and time. Each input carries error, the assumed relationships are population averages, and the actual relationship between your HR and your VO2 varies with fitness, hydration, caffeine, sleep, and altitude. The output inherits all of that and adds the regression’s own modeling error. Premium devices in 2026 are slightly better than the 2017 cohort, but the underlying problem has not been solved.

Workout heart rate: chest strap remains the gold standard

Wrist optical heart rate is good for resting, sleep, and steady-state effort. It is bad during intervals, weights, and any workout where you grip hard or where your heart rate changes faster than every fifteen seconds. Lag on a wrist sensor during a sprint repeat is 10-15 seconds on a current Garmin or Apple Watch, and the dropout rate on rowing or kettlebell work is high enough that some sessions come back with traces that look more like seismographs than HR curves. The physics is covered in wearable sensor accuracy; the related HRV question, if you are using your watch for recovery, is at what is HRV.

A chest strap reads the electrical signal directly. Polar’s H10 and the Garmin HRM-Pro Plus are the two we test against. Both produce beat-to-beat data within one BPM of a 12-lead ECG. They are uncomfortable, you have to wet the contacts, the strap occasionally slips on long runs, and the H10’s battery is a year-long affair. For interval training and racing, none of that matters relative to having data you can trust. Our position has not changed in five years: serious training, chest strap; everyday tracking, wrist.

A third option splits the difference: optical arm bands like the Polar Verity Sense and the Coros HRM. Same PPG as a wrist sensor, but worn on the upper forearm or bicep where motion artifact is smaller and the strap fits tightly. Independent comparisons (DC Rainmaker has done extensive work here) put them within 1-2 BPM of a chest strap for most activities, including HIIT, except heavy lifting where bicep contraction interferes with the optical path. We recommend the Verity Sense to readers who can’t tolerate a chest strap.

Swim metrics: pool yes, open water maybe

Pool swim tracking is one of the things the current generation of watches does well. The accelerometer trace of a flip turn is distinctive enough that lap-counting works without GPS, and stroke detection is reliable for the four standard strokes. The Apple Watch, Garmin Swim 2, the Forerunner 955/965 in swim mode, and the Fitbit Charge 6 all produce defensible pool data. We have measured stroke counts within 3% of a hand count and lap counts within one lap over a typical 1,500-meter session.

Open-water is a different problem. The watch’s GPS antenna is on the wrist, and the wrist spends most of a freestyle stroke underwater. Apple stitches the trace from the brief moments the wrist surfaces and interpolates the rest. Garmin offers a “swim cap” mode that pairs with a small GPS unit tucked under the cap, removing the antenna problem entirely. Both approaches produce traces that often agree with each other to within 5%. They occasionally disagree by 15% on the same swim, which is the honest summary of where the technology is. If you train open-water seriously, a boat-mounted or swim-buoy GPS unit remains the more reliable source.

VO2max: better than you’d guess, under conditions

Garmin pioneered the on-watch VO2max estimate using Firstbeat’s algorithms; Apple, Fitbit, and Polar all run their own versions now. The idea is sound: during steady-state outdoor running, with GPS providing pace and the wrist providing heart rate, the relationship between submaximal HR and pace can be extrapolated to estimate maximum aerobic capacity. The math is the same as a sub-max treadmill test in a sports lab, just with worse input data.

Under ideal conditions the estimate is accurate to within roughly 5 mL/kg/min against a lab VO2max measurement. Snyder et al. (2021) in Sensors compared Firstbeat-based Garmin estimates to lab measurements in 23 trained runners and found a mean absolute error of 3.5 mL/kg/min, with bias of about 1 mL/kg/min in the high direction. Good enough that the trend over a training block is informative, and good enough that someone moving from 45 to 50 mL/kg/min is genuinely fitter.

The estimate is less reliable in two cases. Low-fitness users: the HR-pace relationship at very low cardiorespiratory fitness is steep and noisy, and the algorithms were developed and validated mostly on trained populations. We have seen the same person get readings of 32 and 41 in a single month, which is not biologically plausible. And users whose runs are mostly interrupted — urban running with traffic lights, treadmill running with manual calibration, stop-start training sessions — because the algorithm cannot find a long enough steady-state window to fit the regression cleanly. For everyone else, the watch’s number is a defensible cheap proxy for the lab number. Track the trend, ignore the third significant digit.

What this means if you are buying

The Garmin/Fitbit comparison in Fitbit vs Garmin covers the broader product question. The accuracy-specific guidance is shorter.

Steps are fine on almost every current device. GPS is where you should pay for hardware: multi-band L1+L5 makes a measurable, visible difference on a map, and the price gap to single-band has narrowed enough that the trade-off favors multi-band for most outdoor runners. Calorie figures should be ignored as absolute numbers regardless of how much you paid. Wrist HR is good enough for easy days and bad enough for hard days; the only real fix is a chest strap or an arm band. VO2max is a useful trend indicator for trained runners and noise for everyone else.

The Basis Peak comes up here as a footnote. It was one of the seven devices in the Shcherbina paper, acquitting itself well on heart rate while losing badly on calories. It is unsafe to wear in 2026 and the cloud has been off since December 31, 2016, so this is academic. Sensor accuracy on a watch is a hardware question. The accuracy of the numbers the watch shows you is a question about the chain of estimates the watch makes from those sensors, and that chain is where most of the error lives.