Vocal removers seem almost magical the first time they work well, and completely broken the next time you try a different song. The inconsistency isn't a bug, it's a direct consequence of how the underlying trick actually works: it depends entirely on how the original song was mixed, not on any understanding of what a human voice sounds like.
The center-channel cancellation trick
Most commercial vocal removers use a technique called phase cancellation, or center-channel extraction. In a typical stereo mix, the lead vocal is panned dead-center, meaning it's mixed at equal, identical volume into both the left and right channels. The trick inverts the phase of one channel and adds it to the other; anything identical in both channels cancels out to silence, while anything panned to one side (guitars hard-left, a synth hard-right) survives because it isn't identical across channels. Since the vocal is usually the most prominent centered element, this often removes it convincingly. You can hear this directly with the Vocal Remover, which applies this cancellation automatically.
Why it fails on certain songs
The trick fails wherever the assumption breaks down. A mono recording (both channels genuinely identical everywhere, not just for the vocal) cancels out entirely, leaving silence rather than an instrumental. Live recordings and older masters often pan the vocal slightly off-center or apply different reverb to each channel, so cancellation only partially removes it, leaving a thin, phasey ghost of the voice. And because bass, kick drum, and snare are also frequently mixed dead-center for punch, they get cancelled right alongside the vocal, which is why a "vocal-removed" instrumental often sounds oddly hollow in the low end. Checking whether a track is genuinely stereo first with the Audio Metadata Viewer or converting with the Stereo-Mono Converter can help you understand why a specific file behaves the way it does.
If a track isolates poorly, splitting it into individual left/right channels with the Audio Channel Splitter can sometimes reveal which channel carries more of the instrumental information worth keeping.