We analyze one million steps to find the best device.

We analyzed 1,024,917 manually counted steps across 293 public reports, including videos and written reviews. Each count came from the person doing the test. We checked counting methods, device readouts and duplicate results. Coverage favors devices with more published tests, so newer models can have less evidence. Here are the results.

Collection ongoing

We’re checking more public tests and will add verified reports in batches.

18 of 276 device labels qualified for the walking ranking. Each ranked model had at least 5 walking tests, 3 testers and 10,000 manually counted steps in total. Each test needed at least 500 steps. How we ranked

Give equal weight to
Open any device to see its tests.
Walking comparison. Lower median difference is better. 18 devices
Original tests
12025 to 202647,69848,0540.52%
2202626,08626,1870.55%
32024 to 202533,50033,5780.59%
42020 to 202117,90616,2250.60%
52025 to 202636,25436,2740.68%
62025 to 202641,89242,1400.70%
72023 to 202550,90051,0550.71%
8202669,00068,9450.95%
9202646,57546,0671.07%
102024 to 202528,36128,4521.27%
112025 to 202626,60026,2881.28%
122020 to 202111,00010,6461.30%
132024 to 202548,94448,4161.66%
142021 to 202613,80014,3022.40%
152025 to 202618,00418,3442.54%
162017 to 202017,00017,4083.27%
172022 to 202511,00011,3513.61%
182020 to 202315,00015,0784.30%

The step columns sum each device's included tests. Shared walks appear in each device's row but count once toward the 1,024,917-step collection.

Device explorer

What stands out

8 of 18 ranked models had a median counting difference below 1%. Garmin Venu 4 led the walking ranking at 0.52%, followed by Fitbit Edge at 0.55% and Apple Watch Ultra 2 at 0.59%. Fitbit Sense ranked #18 for walking at 4.30%. Changing which tests we include changes the order.

A good run can hide a bad walk. In one reviewer's Suunto Race 2 test, it missed 3,728 of 10,000 walking steps, then just 8 of 5,000 jogging steps. Same tester and model, very different result.

More steps does not mean more people. Four testers supplied about 71% of the reference steps. Giving each tester equal weight shifts several results. This is a ranking of the tests we collected, not a universal winner.

Browse the full dataset and original reports

How we counted, ranked and checked the data

What the numbers represent

The collection contains 659 credited walks or trials, 1,082 device observations and 81 tester identities or source-level groups. The 276 device labels include historical models and variants. Vora compiled the published results and ran the analysis. We did not perform the physical tests, independently recount videos or use private member data.

The ranking rules

Walking trials need at least 500 manual steps. A ranked model needs five results, three tester identities and 10,000 reference steps. We chose these cutoffs after collection. The 18 ranked models contribute 178 device observations. The full dataset also includes results from named third-party apps, counts derived from reported chart errors, and separate running, stairs and restricted-arm tests.

Each test's error is the absolute difference between the device and manual count, divided by the manual count. We take the median of those percentages. A tracker showing 4,950 or 5,050 for a manually counted 5,000-step walk differs by 1%. Totals cannot distinguish missed steps from false steps that cancel them out.

Known reposts and reused walks were merged. Multiple devices can record one walk, but that walk contributes its reference steps once to the total. The first release covered 999,717 steps across 287 reports.

Source checks and uncertainty

We required explicit manual references and exact device results, checked extraction and known duplicates, and retained source URLs and uncertainty notes. We did not give creators numerical credibility scores or weight them by popularity. Manual counts can still be wrong.

We compared ordinary medians with means, step-weighted errors, equal-tester summaries and leave-one-tester-out results. The walking comparison CSV includes the results after removing Reddit self-reports and additional reports with documented quality concerns.

Conditions, firmware and readout timing vary. These are reported count totals, not controlled head-to-head tests or all-day validation. “Every device” spans all recorded activities and readouts, so those rows are unranked. This snapshot reflects the data reviewed through November 18, 2026.

Disclosure: Vora has a partnership with WHOOP. WHOOP results appear in the full data and are evaluated using the same thresholds.

Download the data

Sources & References

  1. Full step count dataset and original reports
  2. Complete source list (293 public reports, CSV)
  3. All 1,082 device observations and original source URLs (CSV)
  4. Michael L Hicks. Michael Hicks: ten-watch walking and jogging test (2025-12-22)
  5. tfk / the5krunner. the5krunner: 9,000-step Air, CIRQA and Ultra 4 test (2026-11-13)

All research discussed in this article is summarized in our own words. We link to original sources for full access. This content is for informational purposes and does not constitute medical advice.

Step countingWearablesApple WatchGarminFitbit

Bring your wearable data into Vora

Vora brings wearable data together with workouts, food, sleep and recovery to surface things worth noticing.

Related Articles

Wearables

Your training load can change without a new workout

A load chart can fall as old workouts age out, or a ratio can rise as its baseline changes. Reproduce both examples and compare three official definitions.

Read more
Health Technology

Your watch and food app disagree. They owe you an explanation.

A workout total, active calories and a food-app adjustment can all be correct without matching. The problem is making people discover the difference after they have already planned their day around the numbers.

Read more
Wearables

Why your sleep trackers disagree about the same night

An 85 is not the same across Apple, Oura, WHOOP and Garmin. Compare what the scores mean, separate missing data from different definitions, and use a downloadable log to investigate your own nights.

Read more