What an Online Switch Tester Can and Cannot Measure
A browser receives 2 events per key, one when it goes down and one when it comes back up, and everything an online switch tester reports is derived from those two. That is enough to measure registration, repeats and timing to the millisecond, and it is not enough to measure what people actually choose switches for: a 45 gram spring and a 67 gram spring produce identical events, and so does a switch that clicks and one that does not.
Two different objects
The confusion starts with the name. A physical switch tester is a small block holding four to nine loose switches, sold so you can press them and decide which one you like. It is not connected to anything and it measures nothing at all.
A web page called a switch tester is testing your assembled keyboard. It sees the finished output of a switch, a keycap, a matrix, a controller and a cable, and it cannot tell you which of those five produced the behaviour on screen. When a key misbehaves, the page has shown you that something is wrong, not what.
What the browser can see
The measurable list is short and genuinely useful.
- Registration: whether a key produces an event at all, and which key the system thinks it was.
- Position against character:
KeyboardEvent.codegives the physical location andKeyboardEvent.keygives the character produced, which is how a test can draw several layouts correctly. - Repeats: a second press with no release in between, and whether the operating system marked it as auto repeat.
- Timing: how long a key was held, and the interval between one release and the next press.
- Simultaneous presses: how many keys the board reports at once, which is where the six key limit and rollover become visible.
Held together, those five cover the faults that matter in daily use: a key that does not fire, a key that fires twice, a key the system reads as a different key, and a combination the board refuses.
What no web page can measure
Force is invisible to software. The controller reports a closed circuit, not the pressure that closed it, so spring weight leaves no trace in the data. No page can tell a light switch from a heavy one, and a page that names a weight is repeating something it was told.
Travel and actuation point are equally hidden. The event arrives at the moment the contact closes, and how far down the stem that happened, whether at 1.2 mm or 2 mm, is not in the data. A page claiming to measure actuation depth from a browser is describing something it cannot see.
Sound and feel are outside the channel altogether. A tactile bump, the pitch of a bottom out, the difference between a lubricated switch and a dry one, the sound of the case: none of it reaches the browser. Nor does switch type. A page that names your switches is reading a claim you typed into it, or guessing from the device name the operating system reported.
Two further gaps are worth naming. Latency measured in a browser includes the operating system, the browser itself and the display, so it is a figure for the whole chain rather than for the switch. And durability, the rating printed on a datasheet in tens of millions of presses, is a manufacturing claim that no session of any length can confirm.
Using each tool for its own job
Buy on the properties a web page cannot see, and verify with the properties it can. Feel, weight and sound come from pressing real switches, borrowing a colleague's board, or reading a review that used a force curve machine. Registration, repeats and rollover come from the keyboard test, which is also where a second hand board should be checked before money changes hands.
That sequence catches the two common mistakes. One is choosing switches from a description of sound written by a shop. The other is concluding a board is faulty from a page that only reads characters and therefore misreports every key that produces nothing on its own, including the dead keys that European layouts rely on. A key that seems sluggish rather than broken is often just dirty, and cleaning it is the cheaper experiment.
Once the hardware checks out, the useful test is a real one: does the board stay pleasant for a few hundred words? Sonnet 33 answers that nicely. William Shakespeare saw it printed with the rest of the sequence in 1609, and its 437 characters end almost every line in punctuation, which tells you more about a keyboard than any specification sheet does. John Keats gives the longer answer: To Autumn, printed in 1820, holds 494 characters together with semicolons, so the little fingers stay busy from the first line to the last. And The Village Blacksmith, which Henry Wadsworth Longfellow published in 1841, keeps a short regular line and a steady beat across 371 characters, which is the pleasantest way to listen to a board you are still deciding about. All three are waiting in the typing test, each with its author, year and source on its own page.