Counter Metric

Give a team one number to raise and the number usually goes up. What happens next is the problem. Every metric has a fastest route upward, and that route almost always eats something standing next to it.

Pagedeck learned this once. Signup completion went on the board as the quarterly target, the proposal that won was sending up to three onboarding emails a day, and completion did move — 62 percent to 71 percent. A month after the win was reported, the email unsubscribe rate had jumped from 0.4 percent to 1.8 percent. Once somebody unsubscribes there is no path back to them, so there was nothing to undo. The loss was larger than the gain, and it stayed invisible for a month because nobody had that second number on a screen.

A dashboard carrying only the target can't show that trade. So the practice settled into naming the number to protect at the same time as the number to raise, as a pair. The protecting half is the counter metric. Like the barrier along the edge of a road, it does nothing at all until something crosses the line.

A counter metric is not a new calculation. It's an existing measurement promoted into a specific role, and the promotion follows a fixed order.

1. Pin down what you're raising. One sentence, like signup completion this quarter. Choosing what to protect requires knowing what will be pushed.
2. Have the team say out loud the laziest way to move that number. Send more email, bury the cancel button, drop required fields from signup and collect them later. This isn't assuming bad faith — it's looking at where the road runs. Teams under pressure take the short route, and thirty minutes of discussion surfaces all of them.
3. Pick a number that already measures what those shortcuts damage. More email maps to unsubscribe rate, a buried cancel button to support ticket volume, deferred fields to first-week drop-off. The pairing is worth arguing over, because a counter metric that sits downstream of a different mechanism will stay flat while the damage happens somewhere you aren't looking.
4. Write the line. Record the current value and the tolerance together: unsubscribe rate sits at 0.4 percent, so crossing 0.8 percent reverses the change. Write it before the experiment or release starts. Deciding afterward means deciding against whatever the result happened to be.
5. Put both numbers on the same screen. If you run an A/B test — randomly splitting traffic between two versions and comparing the outcomes — the results table shows the target and the counter metric side by side.

The verdict is mechanical. Target up but counter metric over the line means the change gets reverted. Opening a debate about whether this one deserves an exception is the same as never having drawn the line.

A good counter metric is one that actually moves. Attach a number that never budges and you get the feeling of safety without the function. If the value barely changed over the past year, this release won't change it either, and something else should hold the slot.

Numbers that respond too slowly are weak in the role for a different reason. Quarterly churn is accurate and tells you four months from now whether this release was a mistake. Use something that moves earlier instead. A metric that moves ahead of the outcome is a leading indicator, and in churn's slot it might be return visits during the week after release. The trade is that early signals carry more noise, so the tolerance around them has to be wider than it would be on a slow, settled number.

Two or three is the right count. One leaves a direction unwatched; six means something is always over its line and nothing ever ships. From the shortcut list in step 2, pair lines only with the two or three routes the team would realistically take.

The last test is the record. If no change got reverted by these lines last quarter, either the lines are loose or nobody is reading them. The number only carries weight in a team that has reverted something once. After that happens, nobody turns up with the counter metric row left blank on an experiment plan.

"The counter metric for this test is unsubscribe rate at 0.8 percent."

Sharing, in advance, the line that reverses the change no matter what the result says.

"Conversion went up but we broke the counter metric. Rolling it back."

The target was hit and the price crossed the line that was set for it.

"That one can't be a counter metric. It hasn't moved in a year."

It isn't sensitive enough to stop anything.

Every experiment writeup at Pagedeck has a table, and the table always has four rows. The top row is the number being raised, the next two are counter metrics, and the last is the sample size. When Ravi ran a version that removed one step from the signup flow, the top row rose while the second row, first-week drop-off, went over its line.

What came out of that meeting wasn't scrapping the idea. It was a question about why removing that step showed up a week later. The answer turned out to be that the skipped field was the one that decided which workspace a new user landed in. Mia says a version of the same thing often: an experiment that trips a counter metric isn't a failed experiment, it's an experiment that found something.

The most common misreading treats a counter metric as a number that must never move. The language of protection invites that. In practice it comes with a tolerance attached. Unsubscribe rate going from 0.4 to 0.5 percent hasn't crossed anything; it's inside the expected range. Revert on every wobble and the team learns that changing nothing is the safe play, at which point the target metric stops moving too.

Believing that more watched numbers means more safety is just as common. More eyes should mean fewer misses. But metrics fluctuate on their own, so with six lines drawn, one of them always reads oddly. Each time, a discussion follows, and within a few weeks the team starts ignoring the warnings. A team with six lines and a team with none end up in the same place. Safety comes from honoring the lines, not from how many were drawn.