How accurate are step counters? Phone, watch and pedometer against a hand tally

The reference here is a person walking behind the walker with a mechanical tally counter, pressing it once per footfall. That is the whole reference. It is not a better sensor than the ones under test; it is a different kind of thing, and it is the only one available that cannot be fooled by an arm.

Nine positions were worn at once, so every device saw the same walk, the same floor and the same cadence. Nothing here is a manufacturer’s figure. All of it is ours, and the parameters are below so that somebody can disagree with them specifically.

Run parameters

The six conditions

Three of them are the walk everybody tests: a treadmill, a brisk outdoor loop, a slow indoor walk. Three of them are the walk nobody tests, which is walking while your hands are doing something else.

Figures are signed percentage error against the tally. Negative means the device counted fewer steps than were taken. Nine sessions per cell.

The cart was loaded to 18 kg and pushed with both hands on the handle. The bag was 6 kg, carried in the dominant hand, the other arm swinging freely. The rail condition is one hand on the treadmill rail at the same speed as the first column.

Read the first two columns and every product on this page looks fine. In the condition the marketing pictures show, arms swinging, sensors on a wrist are within 2.2% of a person counting by hand. There is nothing wrong with the accelerometers. The problem is where they are.

Overall, with the interval attached

Bias equals absolute error in every row, which is worth saying out loud: while walking, nothing here over-counted in any condition. Every device on every position missed steps. The over-counting happens when you stop walking, and it is further down the page.

Under M-03 we call two things different only when their intervals do not overlap. So:

• The hip-worn EX-510 and the iPhone in a pocket are a tie. 1.2% and 1.5% is not a gap we are entitled to rank, and neither is anybody quoting us.
• The iPhone and the Pixel in a pocket are also a tie. The EX-510 does separate from the Pixel. Those two statements are not contradictory, they are what overlapping intervals look like, and the honest summary is that pocket and hip are one group.
• The four wrist devices cannot be separated from each other. Garmin’s 16.2% and Fitbit’s 21.8% look like a ranking and are not one. Do not quote a wrist ranking from this page.

What the data does support is a gap between placements: pocket and hip beat wrist by more than an order of magnitude, and no interval comes close to closing it.

Where each one fails

The four wrist devices fail when the arm stops. Cart, bag and handrail cost them between 19.8% and 47.8%. This is not a defect. A wrist counter infers footfalls from arm swing, and there is no swing to read. But the failure is invisible to the wearer, it is not flagged in the app, and it lands on exactly the walking most people do most of: a supermarket, a stroller, a briefcase, a dog lead, a handrail.

The phone in your hand fails everywhere, mildly. 9.8% overall. Holding a phone stabilises it against your body’s motion, which is the motion the counter is trying to read. It was the only position that was mediocre in all six conditions rather than excellent in three and terrible in three.

The spring-lever pedometer fails at low speed and only at low speed. The Yamax Digi-Walker SW-200 was the single most accurate device at both walking speeds, 0.3% and 0.5%, and then missed 24.6% of steps at 0.70 m/s. A pendulum arm needs a vertical impulse to close a contact and a slow shuffle does not produce one. If your walking is mostly corridors, queues and crowded pavements, this is the wrong instrument rather than a cheaper one.

The pocket phones fail slightly at low speed too, 3.4% and 5.2% at 0.70 m/s, and that is their worst column. It is also the smallest failure anything on this page has in its worst column.

The 3D pedometer’s failure is not in this table. It is below.

Steps that were never taken

Six seated sessions, 60 minutes each, true count zero. Twenty minutes typing, twenty minutes washing up at a sink, twenty minutes reading a hand-held book. Median steps logged per hour, range across the six sessions in brackets.

The mechanism that rescues the EX-510 at 0.70 m/s is the same one that costs it here: an accelerometer sensitive enough to resolve a slow footfall will also resolve a scrubbing brush. The spring lever that ruins the SW-200 on slow walking is what makes it the quietest device in the set when nobody is walking. You are choosing which error you want.

For wrist devices the two errors do not cancel. A 200-step phantom hour does not repay a 40%-shortfall walk, and they do not occur on the same days for the same reasons.

What we got wrong, and what this cannot support

The first four cart sessions were discarded. We ran them with an empty cart, and an empty cart on a smooth floor gets pushed with one hand while the other arm swings, which is not the case we set out to measure. They were re-run loaded to 18 kg with both hands on the handle. The discarded sessions are not in any figure above.

One inter-observer pair disagreed by 0.63%, over the 0.5% permitted by M-05, on an outdoor session. That session was repeated. The other pairs agreed to within 0.21% and 0.34%.

Then the limits, which are real:

• One walker. Every one of the 54 sessions was walked by the same adult, with one gait, one leg length and one arm-swing amplitude. Gait varies enormously and a single walker cannot support a claim about anybody else’s absolute figure. Treat the spread between positions as more trustworthy than any single number.
• Per-session residuals are not published. Only the summary is. Until they are out, this report is checkable in principle and not in practice.
• No second lab. Nobody has reproduced this on their own walkers. That is the standard we apply to other people’s accuracy claims, and by our own log entry RP-COR-2026-01 it is one this desk does not yet meet for its own work.
• Firmware moves. These figures describe the software each device was running on 2026-06-18. A step-detection change ships without a changelog entry more often than not, so treat the date as part of the figure.

What would change our mind

A wrist device that detects the arm-occupied case and says so, rather than silently reporting a low number, would change the recommendation immediately. So would a second walker set showing the cart and rail losses are specific to this walker’s gait. We will run a multi-walker series before we publish a step figure with anybody’s name ranked against anybody else’s.

Until then the finding is a placement finding, not a brand finding: the most accurate step counter most people own is the phone already in their pocket, and the only thing that beat it was a clip-on device that costs a fraction of any watch here and does nothing else.