Metrics that move when nothing got better
The metric you optimize is the metric you eventually corrupt, which is why a rising chart can be the most misleading thing on the wall.
Most metrics don't tell you whether the product got better. They tell you whether the number got bigger. Those are different facts, and the gap between them is where a lot of careful, well-meaning teams quietly lose the plot. You pick a number because it's a decent proxy for something you actually care about. Then you point a quarter of effort at it. The proxy and the thing it stood for drift apart, and the chart keeps climbing anyway.
This is Goodhart's law wearing a product manager's badge: when a measure becomes a target, it stops being a good measure. The corruption isn't fraud. Nobody is faking data. The team is doing exactly what you asked, and the metric is moving for reasons that have nothing to do with anyone being better off.
The proxy quietly detaches
I watched a support team get told to drive down average time-to-resolution. Within two months the number was down a third and the VP was thrilled. What actually happened: agents learned to close tickets the moment they sent a reply, then open a fresh ticket when the customer wrote back. One frustrated person, four "resolved" tickets, faster than ever. The dashboard described a team getting sharper. The customer experienced a team that kept hanging up on them.
Nothing about that is dishonest. The agents optimized the thing on their scorecard, because that's what scorecards are for. The metric detached from the reality it was hired to represent, and because the metric was the only thing anyone looked at, the detachment was invisible until customers started leaving.
The pattern repeats with every soft proxy. Engagement time goes up because the UI got confusing and people are hunting. Signups climb because you buried the cancel button in the flow. Daily actives rise because you started sending a notification that does nothing but reopen the app. Each chart points up. Each underlying experience is flat or worse.
A number that can only go up is not measuring anything real — it's measuring your willingness to push on it.
Movement is not improvement
The trap is that a moving line feels like evidence. It has a slope, it has a direction, it fits in a slide. So the conversation becomes about the line instead of about the thing the line was supposed to stand for. You optimize the representation and assume the territory followed.
Three questions cut through this faster than any dashboard:
- →If this number doubled tomorrow with no other change, would a single customer be measurably better off?
- →What's the laziest way an indifferent person could move this number, and is that path open?
- →If the number is up but the business isn't, which one do we believe?
That last question is the tell. When your north-star metric and your revenue, retention, or word-of-mouth start pointing in different directions, the metric is lying and you've stopped checking. The healthy instinct is discomfort, not celebration. A metric that climbs while everything it's supposed to predict stays flat is not a success you haven't explained yet. It's a measurement you've outgrown.
Pair every metric with its enemy
The fix isn't a better single number — there's no proxy clean enough to survive becoming a target. The fix is to never let a metric stand alone. Pair each one you optimize with a guardrail that gets worse precisely when you cheat the first.
- →Time-to-resolution, paired with reopen rate and a follow-up satisfaction score.
- →Signups, paired with thirty-day retention of those exact signups.
- →Engagement minutes, paired with task completion — did people finish what they came to do, or just mill around longer?
When you move the primary number, the guardrail tells you what it cost. Gaming the support metric by churning tickets makes reopen rate spike, and the trick dies the week you start watching both. The guardrail isn't decoration; it's the thing that keeps the proxy honest by making the cheat expensive to hide.
Two more habits help. Watch the distribution, not just the average — a mean time-to-resolution can drop beautifully while your worst cases, the ones that actually decide whether someone stays, get slower. And put a real cost on every number you ship: who has the incentive to move this, and what's the easiest thing they could do that would look like progress without being progress. If you can't answer that, you're not ready to make it a target.
Treat the dashboard as a hypothesis
The deeper shift is to stop treating metrics as verdicts and start treating them as hypotheses. The number says the product probably got better. Your job is to go disprove it. Call five customers whose behavior moved and ask what changed for them. Read the session where the conversion happened and check whether the person looked relieved or trapped. Watch the support queue for the thing the chart can't show.
When a metric and reality agree, you've earned the confidence the chart was offering. When they disagree, you've found something far more valuable than a green number — a place where your map is wrong. The teams that compound are the ones that go looking for that disagreement on purpose, while the line is still going up and everyone else is busy celebrating.
A rising chart is a question, not an answer. Treat it as proof and you'll optimize your way to a beautiful dashboard sitting on top of a product nobody loves anymore.
