Goodhart's Law

Managing work by numbers produces one scene over and over. The number that was made a target clearly improved, while the condition that number was supposed to stand for stayed exactly where it was or got worse.

Customer support response time is the standard illustration. A target lands to bring the average down from 30 minutes to 4, and a quarter later the dashboard reads 4 minutes. What actually happened is that the support team started firing off a "we're looking into this" reply the instant a ticket arrived. Time to first reply was what was being measured, so the number came down honestly, and the time it took a customer to get an actual answer didn't move at all.

Before the effect had a name, this got handled as a people problem. That team is gaming it, or they aren't serious about the work — the kind of diagnosis that implies a staffing change would fix it. So the same thing kept recurring, one department over.

In 1975 a British economist writing about monetary policy put the general form of it into a sentence: any observed statistical regularity tends to collapse once pressure is placed on it for control purposes. The point is that metrics break because of how the system is arranged, not because of anyone's character. Having a name for it made it possible to argue about the design of a target instead of the integrity of the people chasing it.

The failure runs through three stages. The striking part is that at no stage does anyone do anything wrong.

1. You pick something measurable in place of what you actually care about. What you want to know is whether the customer's problem got solved, and that is hard to count directly, so you count response time instead. Choosing that stand-in is exactly the work a metric definition, which pins down what gets counted and how, is doing.
2. You attach a target and a reward to that number. As soon as it becomes a quarterly goal or feeds a performance review, it stops being a gauge people read and becomes a direction people move in.
3. Optimization flows through the gap between the stand-in and the real goal. Response time and problem resolution move together most of the time, but they are not the same thing. Somebody holding the target finds the place where the two come apart — the route that lowers response time and leaves resolution untouched.

Stage three does not require bad faith, which is what makes the law worth knowing. For a support agent looking at twenty open tickets and a 4-minute target, sending the acknowledgment first is a sensible move. That person is doing as well as the rules allow.

The wider the gap, the worse the break. In the Pagedeck case the average fell from 30 minutes to 4, while the share of customers who wrote back in about the same issue rose from 12 percent to 31 percent. Acknowledgments turned one ticket into two, so the real throughput got worse while the reported figure improved.

The same structure shows up everywhere. Target the number of tasks completed and tasks get sliced thinner. Target signups and the ad budget goes up. Nobody lied in either case.

The moment to raise this law is the moment a target gets attached to a number.

Goal-setting is the first place. When drafting OKRs, which split an objective from the key results that confirm it, or the KPIs a team watches continuously, write down in advance how the number you want to raise could be raised the wrong way. It matters most when a single large figure absorbs everyone's attention, as with a north star metric meant to point a whole company in one direction.

Experiment design is the second. In an A/B test, which splits traffic randomly between two versions and compares them, defining success as click-through rate alone hands the win to whichever version had the more baiting headline. Clicks go up, retention behind them goes down, and the test declares that a victory.

Performance reviews and compensation are the third and most dangerous. A number that money depends on gets optimized fastest.

There are three common defenses. The first is to pair the target with a number that watches whatever could break while you push the target up; putting the repeat-contact rate next to response time closes the empty-acknowledgment route. The second is to keep reading cases behind the number rather than the number alone — twenty support threads a quarter will tell you things the average never will. The third is to avoid leaving a target in place longer than a quarter, since the workarounds get more sophisticated the longer it stands.

"Make that a target and you'll get Goodhart. Let's put a guardrail next to it."

The number being targeted can be raised through a side door, so pair it with something that watches for that.

"B wins on click-through, but week-one retention after signup is lower for B."

The success criterion and the outcome anyone actually wants have come apart.

"Response time didn't improve — we just sent more acknowledgments."

The number fell because of a gap in how it's counted, not because anything got better.

This comes up whenever Mia sets quarterly goals. Last quarter Pagedeck made new signups the team target and cleared it. Checking a month later, a large share of those extra signups were accounts that arrived once, created a single document, and never came back.

What Mia changed this quarter wasn't the number but its partner. Signups stayed, and next to it she hung the share of accounts that create three or more documents within two weeks of signing up. Once both numbers had to rise for the goal to count, rewriting ad copy to pull in more signups stopped being an answer.

Most people read the law as a warning never to make a metric a target. The sentence does sound like that, and anyone who has been burned by a number reads it that way first.

A number that isn't a target changes nothing. A gauge nobody steers toward gets read once a quarter and forgotten. What the law actually says is not that you shouldn't set targets, but that setting one changes the nature of the number, so you should map the side doors and attach a partner before you do it.

Taken purely as a warning, teams overcorrect. They either stop putting targets on anything, or they only pick numbers too safe to be gamed. Numbers that safe are usually numbers nobody can move, and the dashboard fills up with vanity metrics — figures that look good and support no decision.