Rollover, ghosting, and why your keyboard drops the third key
Keyboards do not have a wire per key. That would need over a hundred connections, so keys are wired as a grid of rows and columns and the controller scans it, energising one row at a time and reading which columns respond. A key at the intersection of row three and column five is identified by being the thing that connects them. This is efficient and it has a specific failure.
Hold three keys that share rows and columns in the wrong arrangement, and the current finds a path through the other two that is indistinguishable from a fourth key being pressed. The controller reports a key nobody touched. That phantom is called ghosting, and the cheap defence against it is blocking: when the controller detects an ambiguous pattern it refuses to report the third key at all. Blocking is why a key you are definitely pressing does nothing, and it is far more common than ghosting because most manufacturers choose to drop input rather than invent it.
The fix in hardware is a diode at every key, which makes current flow one way only and removes the phantom path. Keyboards advertising N-key rollover have those diodes. Keyboards advertising six-key rollover usually do not; six is the number of simultaneous keys the standard USB keyboard report can carry, and is a protocol limit rather than a matrix one. This is why the distinction matters to anyone gaming or playing a software instrument, and why it is invisible to anyone typing prose, where three simultaneous keys is already unusual.
There is a second thing this tool tests that has nothing to do with rollover, which is layout. Browsers expose two different identifiers for a keystroke. One is the character produced, which depends on the layout, so the same physical key gives "a" on QWERTY and "q" on AZERTY. The other is the physical key position, which is the same everywhere. Testing has to track the physical position or the diagram lights up the wrong key for anyone not on a US layout, and where the browser can report your actual layout the printed legends are relabelled to match rather than showing you someone else keyboard.
Keystrokes are read to drive the display and are never stored or transmitted, which is worth stating plainly on a page of this shape, since a keyboard tester is exactly what a keylogger would look like.