> This is is actually a terrible trap, as product usage tells you next to nothing about which features are useful. I suspect it' a big part of why a lot of software has turned to shit last decade or so.
This seems exactly backwards to me. No one who has ever managed a deployed product with good failure telemetry would say that it tells you "next to nothing". Telemetry detects failures that mere testing simply never can and never will.
And no one who lived through earlier ages of consumer software would ever say that it's "turned to shit". Software is getting better. It's getting better steadily and inexorably, and it's been doing that for half a century. It may not be doing what you want, but it's doing it with higher quality than people in the 90's or 00's would have ever dreamed. When was the last time[1] you saw a major consumer application (the TikTok or Instagram client, or your web browser, or VSCode) crash hard and fail on you? This was a daily (or worse) event for most of software history!
And certainly robust telemetry practice is one of the big drivers of this quality revolution we've seen.
[1] Heh, cue the peanut gallery of everyone wanting to post crash reports. That's on me for poor framing, I guess. But seriously, everyone: Both anecdotally and via actual studies, consumer-visible software is getting much better.