Methodology
How the Reaction Test Measures You
The random wait
After you start, the screen stays red for a random 2 to 5 seconds. The five waits for a set are drawn at the start from a seeded random number generator (mulberry32), seeded from your browser's secure random source. A duel link carries the seed, so your friend gets the same waits in the same order. If a try is voided or you false-start, the retry draws a new wait from the same stream; the five planned waits stay as they were.
Starting the clock: the painted frame
Browsers draw the page in frames, usually 60 to 360 a second. Code that runs just before each frame gets a timestamp through requestAnimationFrame. Once the wait is over, the first frame callback whose timestamp has passed the target switches the panel to green. That's one class change that only alters the background colour, with no transition, no layout change, and text that was already on the page.
We then wait one more frame. MDN defines the timestamp passed to a frame callback as the end time of the previous frame's rendering, so the timestamp of the next callback is the earliest moment the green frame can have been handed to the display. That's the onset, and it's where the clock starts. Starting the clock when the code decides to change the colour, as many tests do, can put the start a frame or more before anything changed on screen.
We also watch the frame rate. If any frame in the 200 ms before the green took more than 1.5 times the usual interval, we tell you the browser hiccuped. If the green frame itself arrived more than one frame late, we void that try, because we can't be sure when it reached the screen.
Stopping the clock: the event timestamp
When you press, the browser creates an input event with a timeStamp recording when it received the input. It uses the same clock as performance.now(), so the two can be subtracted directly. We read that timestamp first, before any other work, so a busy page can't add to your time. Your reaction is press timestamp minus onset.
We listen for pointerdown, which fires the moment a mouse button goes down or a finger touches, and for key presses (Space, Enter or any letter, ignoring auto-repeat). We don't use click, which fires on release. If a browser gives an unusable timestamp, we fall back to the time our handler ran.
Precision: browsers round their clocks for privacy, to 0.1 ms in Chrome and 1 ms in Safari and Firefox. Firefox's fingerprinting protection can round to about 17 ms; if we detect that, we warn you. We show whole milliseconds.
How exact is the arithmetic? We test it with scripted presses at known moments on simulated 30, 60, 120, 144 and 240 Hz screens: the time shown is always within 0.6 ms of the real gap between the onset frame and the press timestamp (whole-ms rounding plus Chrome's 0.1 ms clock steps). That covers the part a browser can see. Your screen and your mouse, keyboard or touchscreen add their own delay on top, typically several to tens of milliseconds, and no web page can measure it. So a result here is your reaction plus your hardware, timed to the millisecond.
The sound test uses the audio clock instead. The beep is scheduled on the Web Audio clock and converted to the same timeline with getOutputTimestamp.
Scoring: median of five clean tries
- Median, not mean. Your headline is the middle of your five valid times, so one blink doesn't wreck the score. Mean, best, slowest, spread and a consistency rating are shown too.
- False starts (pressing while it's red) don't count and get a fresh wait. They're listed on your result.
- Under 100 ms is anticipation. It's the same 0.100 s line World Athletics uses for sprint false starts. Those tries are shown and not counted.
- No press within 2 seconds voids the try, and so does switching tabs mid-try.
What we can't measure
Two parts of the chain sit outside the browser's view: the screen and the input device. In the lab, Woods et al. (2015) measured 11.0 ms of display delay and 6.8 ms of mouse delay on their equipment. Anwyl-Irvine et al. (2021) found keyboard response lag averaging around 80 ms on their test rigs, and display timing errors averaging 11.50 ms in Chrome and 30.02 ms in Safari. Human Benchmark puts typical desktop and laptop lag at about 30 ms. Touchscreens weren't tested in those studies, and phones usually read slower.
We don't subtract a guess. Your result names your device, input type, measured refresh rate and browser, with a note on how much delay that can add, and your personal bests are kept separately for desktop, phone and tablet. Check your own setup on the input lag test.
How this compares with other tests
Human Benchmark's reaction test, the best-known one, records the start with new Date() when its React state switches to green and the press inside its mouse and touch handlers, and reports the mean of five. Both stamps are taken in code rather than from the frame and the event, and a mean is sensitive to one slow try. That doesn't make their numbers wrong, but it means scores from different tests aren't directly comparable. Compare yourself on one test and one device.
Changelog
- 2026-11-18: Launch: classic test, F1 lights out, sound and sprint start, choice, go/no-go, duel links, same-screen showdown, live race rooms, ruler drop calculator, input lag check and the daily trainer.
Sources
- MDN, requestAnimationFrame()
- MDN, Event.timeStamp
- MDN, performance.now()
- MDN, pointerdown event
- MDN, AudioContext.getOutputTimestamp()
- Woods et al. (2015), Factors influencing the latency of simple reaction time, Frontiers in Human Neuroscience 9:131
- Anwyl-Irvine et al. (2021), Realistic precision and accuracy of online experiment platforms, web browsers, and devices, Behavior Research Methods 53:1407-1425
- Human Benchmark, reaction time statistics
- World Athletics, 100 metres discipline page