What running power actually measures

The difference from cycling power, which is the whole story

A bicycle power meter contains strain gauges in the crank, pedal or hub. When you push, the metal deforms by a tiny amount, the gauges convert that deformation into a voltage, and the device multiplies the resulting torque by angular velocity. Watts come out. It is a direct physical measurement of mechanical work being done on a rigid object, and that is why a cycling watt is a portable unit: 250 W on your bike is 250 W on a different bike, on a different day, in a different country.

Nothing in running works this way. There is no crank, no rigid transmission, and no sensor in the shoe measuring the force you put into the ground. Instead a device observes motion — accelerations, cadence, ground contact time, vertical oscillation, speed, sometimes grade — and runs a model that outputs a number labeled watts.

There is also no agreed definition of what running power should even include. Should it count the energy stored and returned by your tendons? The work done against air resistance? The cost of swinging your limbs? Different developers answered differently, and those answers are proprietary. This is the reason running power values are not comparable across platforms, and it is not a bug anyone can fix with better hardware.

The two families of implementation

Foot-mounted inertial sensors. A pod on the shoe sits close to the point where force is actually transmitted and samples accelerations at high rate. It also measures ground contact time and leg spring stiffness directly rather than inferring them from the wrist. Stryd is the widely used example.

Watch-native estimates. The watch, sometimes with a chest strap contributing running dynamics, models power from pace, grade, cadence and body mass. No extra hardware, and no sensor anywhere near your foot. Garmin, COROS and Polar all publish something along these lines.

Both are models. The foot pod has better inputs; that is a reason to expect it to be better, not a demonstration that it is.

What has actually been measured

The foot pod against laboratory instruments

Imbach et al. (2020) in Sports ran an incremental test from 8 to 20 km/h on a 200 m indoor track with six recreational runners, against 500 Hz force platforms, a Cosmed K4b2 metabolic analyzer and 400 Hz motion capture. That is a genuine laboratory comparison, and it is the strongest validation this category has.

• Stryd power output against oxygen consumption: R² = 0.82.
• Stryd power output against externally measured mechanical power: R² = 0.88.
• Ground contact time and leg spring stiffness showed no significant difference from the laboratory reference — the underlying biomechanical measurements are good.

Those are respectable numbers and they support the practical claim that the metric tracks effort. But the same paper reported a proportional error in power output that grew with speed, and a systematic underestimation relative to the reference. The device is a better proxy at some speeds than others, which matters if you intend to pace by it across a range of intensities.

The authors also computed the net mechanical efficiency implied by their data at 55 ± 3%. We report that as a flag rather than a finding: an efficiency figure that high is a signal that the quantity being reported is a model output with its own internal conventions, not the physical mechanical work your body performed.

The study's limits are severe and the authors state them: six participants, a correction function estimated from that same small dataset, no maximal speeds, no slopes, no changes of direction, and proprietary algorithms they could not inspect.

Two devices, one runner, one trail — and 82 watts apart

Berzosa et al. (2024) in Sensors did something more directly useful to a buyer. Five participants ran a 2.5 km trail course with 195 m of climbing, twice, a week apart, wearing a Stryd foot pod and a Garmin fēnix 7S Solar with an HRM-PRO simultaneously.

Each device repeated itself almost perfectly. Week to week, Stryd returned intraclass correlations of 0.980 to 0.997 with coefficients of variation of 0.4-2.8%. Garmin returned 0.927 to 0.997 with 0.3-3.3%. Both are excellent. If you use one device and look for changes over time, that is the property you need, and both have it.

The two devices did not agree with each other.

Read that table twice. On the things both devices genuinely measure — speed, cadence, contact time — they agree almost exactly. On the thing both devices model — power — they diverge by an average of 82 watts on the same runner on the same run, and the disagreement is worst uphill, where a power number is most useful for pacing.

A crucial caveat, which the authors state plainly: this study had no gold standard. It cannot tell you which device is closer to the truth. It can only tell you that at least one of them is substantially wrong, and that both are wrong in a perfectly consistent way.

Reliable is not the same as valid, and this is the clearest example on the site

Reliability is whether a device gives you the same answer under the same conditions. Validity is whether the answer is correct. They are independent properties, and a device can have the first without the second — a scale that reads eight pounds heavy is perfectly reliable and completely invalid.

Running power is that scale. Intraclass correlations of 0.99 week to week, and 82 watts of disagreement between two implementations of the same concept. That is a reliable measurement of something, and nobody has established that the something is the quantity printed on the screen.

It is worth being fair about what follows from this. Reliability alone is genuinely sufficient for the main thing runners use power for: pacing consistently within a session and comparing efforts on the same device. It is not sufficient for treating your watts as a physiological constant you can carry between devices, publish, or compare with a training partner.

The right way to use it

As a within-device pacing signal. Its real advantage over pace is that it responds to grade and wind immediately, where pace requires you to translate manually and heart rate lags by thirty seconds or more. On a hilly course, that responsiveness is genuinely valuable and the poor cross-device agreement does not diminish it.

Set your zones from your own testing, on your own device. Never import zones or target watts from another platform, another athlete or a coaching plan written around different hardware. Given the measured gap, a target from the wrong device could be off by more than the width of a training zone.

Do not change devices mid-plan. The step change will look like a fitness change and will not be one. If you must switch, retest and rebuild your zones from scratch.

Be most careful uphill. That is where the two implementations diverged most — 109 W apart on average — and, unhelpfully, where runners most want the number.

Do not treat it as more objective than heart rate. Heart rate is measured, with a documented error. Power is modeled, with an undocumented one. Neither is the truth, and the metric that looks most like physics here is the one furthest from a direct measurement.