You are confusing two separate issues.

What you say is true for automatic detection. In small systems you work be checking everyone manually and making sure they can call you.

You push a change, then a couple hours later you get an angry call from a single customer that represents 70% of all your company's income: you broke them. You check, they're right: what you thought was a fluke was actually a big problem starting. Customers have lost ~half a million by now, and their rate is about half a million every hour their system isn't working because of the problem in your system.

Now what's a better scenario here? Flip a flag and call it a day? Make a PR that undoes the change (if you're lucky your know what PR to roll back, if you're really lucky you just flip a config/variable in the code somewhere, but that's following the advice that your say doesn't apply here). You then force push the PR and push an emergency release (as oncall you get to break the glass, lucky you that you were oncall when you pushed your PR, otherwise you'd have lost precious time coordinating with another engineer, or worse debugging code you're unfamiliar with or having to get permissions and support to push the fix). Finally the release gets rolled out aggressively. This whole thing could easily be an hour. Meanwhile you just simply flip a feature flag and turn it off everywhere. Better yet you press a big red button and ask the must recent changes are undone, no need to fix it.

Next time, you use feature flags. Not because you want A/B samples, but because you want to first send a change to everyone expect the whale and then go from there. And I'd you see an issue you undo it quickly. Hell you realize that your own company is a big user of the code, so you first release only to internal users within the company: congrats you've built a poor man's canary.

The large companies you say, and the system with enough data to be smooth is great for automated detection. Here the problem is not that different, except now you lose $500k every ten seconds, instead of every hour. This justifies investing work into reacting 5 second earlier.

But let's be clear, you still want an easy way to undo any change you do, because it's really painful when you fuck up. Smaller products have less leeway to fuck up.