A classic Sunday evening in a running group chat: “The watch adds hundreds of metres without mercy. We ran a 32 km long run — three kilometres' difference from my friend. Is it calibration?” The replies split into “same here”, “switch on multi-band” and “GPS throws it off”. All three are true in their own way, and none of them answers the main question: how much do watches actually lie, and in which direction? That has been measured, and more than once.

Eight watches, three terrains

The most careful study is by Gilgen-Ammann and colleagues from the Swiss Federal Institute of Sport in Magglingen (JMIR mHealth and uHealth, 2020). Eight models from Apple, Coros, Garmin, Polar and Suunto, 36 measurements for each — in the city, in the forest and on the track, walking, running and cycling, over reference distances from 404 to 4,297 m.

The mean absolute error across all conditions was 3.2 to 6.1% depending on the model. Only the Polar receivers came in under 5% overall.
Systematic error — from +3.7 m to −101 m, with a spread of ±196–231 m.
In the city and in the forest every watch under-read the distance (p < 0.04), by up to 9% in the worst case.
On the track six of the eight over-read the distance, but accuracy there was the best.
Running produced more error than walking and cycling on most devices.

The direction is the main finding. Watches are not “thrown off at random”: under trees and between buildings they cut corners, and in the open they pad the distance.

Ten watches and the track that turned out hardest

Jacko and colleagues (Sensors, 2024) tested ten then-current models in triathlon scenarios.

4,000 m on a 400-metre tartan track (six repeats): an error of 0.8 to 12.1% between models.
A hilly 3.4 km cross-country course: 0.2 to 7.5%.
36.8 km on the road by bike (four repeats): 0.0 to 4.2%.
The pool is a separate story: over 200 m of individual medley with a stroke change every 25 m the error reached 91.7%.

The paradox worth remembering: the track is the worst place for GPS distance, worse than rough terrain. The receiver records a point once a second, and on a bend you cover a noticeable chunk of the arc in a second. Joining the points with straight lines means cutting the corner (under-reading); filtering and smoothing the “noise” means adding zigzags (over-reading). Which effect wins depends on the algorithm of the particular manufacturer, which is why on the same lap two models err in different directions.

The urban canyon from the satellite's point of view

A preprint by Antoń and colleagues (Research Square, May 2026) compared six smartwatches against a Leica GS16 geodetic receiver as the reference. In dense urban development the horizontal position error was 6.7–13 m, and up to 9 m vertically: the signal arrives not directly but reflected off the façades. Under open sky the same watches held 1.5–3 m, the best result being 1.45 m. The study has not been peer-reviewed, but the figures agree with what the two studies above showed: between buildings the receiver loses accuracy several times over.

Where the “extra three kilometres” come from

Three kilometres in thirty-two is 10%. That is more than any GPS error in the validation studies under normal reception. So the culprit is almost certainly not the satellites, but what happens when there are no satellites.

All modern watches switch to accelerometer estimation when the signal is lost — they count steps from arm swing and multiply by stride length. The default stride length is a crude model based on height, which for a particular person can be off by 10–20%. A tunnel, a bridge with a deck overhead, a narrow street between tall buildings, dense forest — and for several minutes the watch is “running” off your arm rather than the satellites. This is what most often lies behind the “hundreds of metres”.

The second candidate is different settings on two wrists: GPS only versus all systems, single-band versus multi-band, auto-pause, 3D distance. Two people run side by side, but their devices are solving different problems.

The third is what the arguing parties forget: your friend is not a reference either. If he ran in the forest with a 5% under-read and you ran along an open embankment with a 2% over-read, the “three-kilometre difference” is made of two errors, and whose is closer to the truth is unknown.

What to do

Switch on all satellite systems and multi-band, if the watch supports it. Multi-frequency reception cuts out reflected signals — precisely the error that dominates in the city.
Calibrate your stride length: run 10 laps of the track in a mode where the watch uses the accelerometer (or with GPS switched off) and enter the correction. This is the only way to fix the “tunnel” kilometres.
On the track, do not trust GPS distance. Count laps with the Lap button, and intervals by the markings.
A normal error on an open route is 1–2%. Over a marathon that is 400–800 m, and it does not mean the course was measured wrong.
Compare not with your friend but with a reference: a certified distance, a measured kilometre, the track. That is the only way to tell whose watch is closer to the truth.
Switch off auto-pause in the city. At traffic lights the receiver “drifts”, and auto-pause triggers sometimes but not always, adding metres and breaking your pace.
On-screen pace in the forest reads slow. If the watch shows 5:10 under the canopy, you are most likely running closer to 5:00. Go by heart rate and feel, and check pace on open sections.

Honest limitations

Models go out of date faster than papers. In the 2020 study there was no mass-market multi-band yet; data on a particular brand cannot be carried over to its next generation. Only the patterns are robust: under-reading under the canopy, over-reading on the track, a collapse when the signal is lost.

The reference is not perfect either. Distances were measured with a wheel and from maps; a reference error of a few tenths of a per cent is baked into the results.

The 2026 preprint has not been peer-reviewed, and its measurements were taken at walking speed.

No study tested the accelerometer mode separately. Our reasoning about the “extra hundreds of metres” when the signal is lost is based on the mechanism, not on published figures.

The bottom line

Eight sport watches (2020): mean absolute distance error of 3.2–6.1%; in the city and the forest — under-reading by up to 9%, on the track — over-reading.
Ten watches (2024): over 4,000 m on a track an error of 0.8–12.1% between models, on the road — 0–4.2%.
In dense urban development position accuracy drops to 6.7–13 m against 1.5–3 m under open sky.
A 10% difference from your running partner is not GPS but the accelerometer during signal loss or different settings. It is cured by stride-length calibration and multi-band.
The normal 1–2% on an open route is 400–800 m over a marathon; work out your splits from the course markings.

Sources: Gilgen-Ammann R., Schweizer T., Wyss T. “Accuracy of Distance Recordings in Eight Positioning-Enabled Sport Watches: Instrument Validation Study”. JMIR mHealth and uHealth, 2020;8(6):e17118. DOI: 10.2196/17118 · Jacko T., Bartsch J., von Diecken C., Ueberschär O. “Validity of Current Smartwatches for Triathlon Training: How Accurate Are Heart Rate, Distance, and Swimming Readings?”. Sensors, 2024;24(14):4675. DOI: 10.3390/s24144675 · Antoń K., Nosidlak T., Abazeed A., Maciuk K. “Accuracy of satellite positioning using GNSS smartwatches”. Research Square, preprint, 2026. DOI: 10.21203/rs.3.rs-9446244/v1 · Lepley A.S., Davis F., Melvin A.C., Zhao Z.Y. “Consumer Smartwatch Technology in Health and Performance Research: Validity, Limitations, and Real-World Applications”. Sensors, 2026;26(14):4486. DOI: 10.3390/s26144486