The test for whether a metric matters is simple: would a bad reading change what you do? If the number can only go up and never triggers an action, it's there to make people feel good, not to run the program.
Total registered users is the classic example. It only grows, it never drops, and it tells you nothing about whether the product works. Active users who came back this week is harder and more honest, because it can fall, and when it falls you have to do something.
I push teams toward metrics that can deliver bad news. Conversion that can drop. Latency that can spike. Adoption measured as people getting value, not people who touched a feature once. A metric that can hurt you is a metric you'll actually watch.
There's a subtler trap too: measuring what's easy instead of what's true. At Atlassian the easy AI adoption number counted any session that touched the feature. The true number counted sessions where someone got something out of it. Those were very different, and only the true one was worth steering by.
Pick the numbers that can ruin your week. Those are the ones telling you the truth.