Proxy Metrics: When You Can't Measure What You Actually Want

A proxy metric is a measurable stand-in for an outcome you can't observe directly — using 30-day retention to represent "customer satisfaction," or time-to-first-value to represent "onboarding quality." Choose one by testing three things: how tightly it tracks the real outcome, how hard it is to game, and how cheaply you can measure it every week, then pair it with a second metric that catches gaming.

Quick answer: A good proxy metric correlates strongly with an outcome you can't measure directly, resists gaming per Goodhart's Law, and is cheap enough to track continuously. Pair it with a guardrail metric, and log the assumption connecting proxy to outcome so you can check it later — not reconstruct it from memory months from now.

What Counts as a Proxy Metric, and When You Actually Need One

A proxy metric is any observable, frequent, cheap-to-measure signal you track instead of an outcome that's unmeasurable in real time — customer trust, long-term loyalty, or "product value." You need one whenever the real outcome can only be confirmed too slowly, too expensively, or too subjectively to inform this week's decisions.

Most of what a product team actually cares about falls into that gap. You can't measure "trust" with a sensor, and you can't know whether a feature made someone's job easier without asking — and even then, self-reported answers are sparse and noisy. Proxies exist to close that distance.

A proxy metric is not the same as a vanity metric, though the two get confused constantly. A vanity metric — page views, downloads, total registered users — is chosen because it climbs and to the right, with no outcome behind it. Eric Ries drew this exact line in The Lean Startup, arguing that a metric only earns its place on a dashboard if it's actionable, not just impressive.

A proxy metric is chosen deliberately: it stands for a specific, named outcome, and its worth is judged by how well it tracks that outcome. If you're not sure which one you're looking at, our anti-vanity metrics guide has a sharper test for telling them apart.

This distinction matters because proxies are the connective tissue of any data-driven roadmap. Get the proxy wrong, and every prioritization call built on top of it inherits the error — a failure mode we cover in more depth in our guide to data-driven product decisions.

Why Proxies Fail: Goodhart's Law, Campbell's Law, and Gaming

Proxies fail for one structural reason: once a measure becomes a target, people — consciously or not — start optimizing the measure itself instead of the outcome it was meant to represent. Two researchers, decades apart and in different fields, described this same collapse.

"When a measure becomes a target, it ceases to be a good measure." — the popular paraphrase of Goodhart's Law, restated in this form by anthropologist Marilyn Strathern in 1997.

Economist Charles Goodhart coined the original idea in 1975 while writing about UK monetary policy. Sociologist Donald Campbell reached the same conclusion independently in 1976, studying U.S. social programs: the more a quantitative indicator gets used for decision-making, the more it distorts the very process it's supposed to monitor. Campbell's Law and Goodhart's Law are, functionally, the same warning from two directions.

In product terms this shows up as gaming — sometimes deliberate, more often people rationally responding to the incentive you built. A few patterns repeat constantly:

• The metric moves, but a harder-to-fake related signal (churn, complaints, refund rate) doesn't move with it.
• People start describing their work in terms of the metric's literal definition, not the outcome behind it.
• The measurement method quietly changes right when the number starts looking bad.
• A metric improves on one team's dashboard but never shows up in the company-level number it's supposed to feed.

None of this means proxies are unusable. It means an unguarded proxy is a target waiting to be gamed — which is exactly what the framework and the guardrail approach below are built to prevent.

A Framework for Choosing a Defensible Proxy Metric

Choose a proxy by scoring candidate signals against four criteria — correlation strength, resistance to gaming, whether it's leading or lagging, and cost to measure — then picking the strongest scorer, not the most convenient one already sitting in your dashboard. Convenience is an availability bias, not a decision.

1. Write the real outcome in plain language, no metric names yet — "customers find onboarding valuable," not "increase D7 retention."
2. List three to five candidate signals that plausibly move when that outcome moves.
3. Score each candidate against correlation, gameability, leading-vs-lagging timing, and measurement cost.
4. Pick the top scorer as your primary proxy, then pick a second as a guardrail against gaming it.
5. Log the assumption linking proxy to outcome, with a date to revisit it (more on this below).

Here's what that scoring looks like applied to a real question — "do customers find onboarding valuable?" — across three candidate proxies:

The scoring only works if step one is honest. This is the same discipline behind Jobs to Be Done: a customer doesn't "increase session frequency" — they're hiring your product to make progress on a job, and the proxy should track progress on that job, not raw activity. Our Jobs to Be Done guide walks through naming that job precisely enough to build a metric around it.

For outcomes that are specifically about user experience rather than business results, Google's HEART framework (Kerry Rodden and colleagues, 2010) — Happiness, Engagement, Adoption, Retention, Task Success — gives a structured checklist for picking a proxy when the real outcome is something as subjective as "ease of use."

Amazon's leadership team, as described in Colin Bryar and Bill Carr's Working Backwards, separates controllable input metrics from output metrics for a related reason. A proxy your team can actually influence day to day is operationally more useful than one that only moves in hindsight, even if the lagging metric is closer to the "real" outcome.

Pair Every Proxy with a Guardrail Metric

A guardrail metric is a second, harder-to-game signal tracked alongside your primary proxy specifically to catch the gaming Goodhart's Law predicts. If the guardrail moves the wrong way while the proxy improves, you've caught the problem before it compounds.

Andy Grove described this in High Output Management as pairing indicators: no single number should ever be trusted alone, because any one number can be optimized in isolation without anyone noticing. A support team's ticket-closure rate needs a guardrail like reopened-ticket rate sitting next to it; an onboarding checklist's completion rate needs day-30 retention watching its back.

This same pairing logic scales up to the single top-level proxy a company rallies around — its north star metric. Choosing one number to represent overall product health is proxy selection at the highest stakes, which is why it deserves its own rigor. Our guide to choosing a north star metric covers how to pressure-test a candidate before an entire org anchors on it.

Guardrails work best anchored to a specific point in the experience, not just a company-wide dashboard number. Mapping a proxy and its guardrail onto the exact stage of the customer journey where the behavior actually happens makes it obvious when a metric is being gamed upstream of where the real outcome would ever show up.

The Real Root Cause: Log the Assumption Behind the Proxy

Most proxy metric failures don't start with a bad metric. They start with an assumption nobody wrote down — a PM decides "checklist completion correlates with retention," ships the dashboard, and six months later nobody can say why that proxy was chosen or what it was supposed to predict.

By the time the number starts drifting, the reasoning behind it has to be reconstructed from memory, usually in a retro, usually by whoever happens to still be on the team. That reconstruction is unreliable by nature — recency bias, incomplete recall, and hindsight all creep in. The fix isn't a better metric. It's a habit: the moment you choose a proxy, write down in one sentence what you're assuming it stands in for, and set a date to check it.

This is the step teams skip most often, and it's the one that would have prevented most of the failure patterns covered in this article. An assumption that's logged the moment it forms is checkable later; an assumption that lives only in someone's head is gone the day they change roles.

Prodinja's Journals are built around exactly this moment. You can log a metric hypothesis or an analytics assumption the instant you form it — including with real browser voice capture, so you're not stopping mid-thought to type — and it sits there timestamped and revisitable, instead of being reconstructed from memory three months later in a retro. The habit matters more than any tool; what a tool like this is designed to do is make the habit take five seconds instead of feeling like homework.

When to Retire or Replace a Proxy Metric

Retire a proxy the moment you can measure the real outcome directly, or the moment its correlation with that outcome breaks down — and you'll only notice the second one if you're periodically re-testing the relationship, not just watching the number climb. A proxy is a temporary tool, not a permanent dashboard fixture.

The most reliable way to test whether a proxy still means what it used to is to watch it across cohorts over time, rather than in aggregate. If the proxy-to-outcome relationship holds for last quarter's cohort but not this quarter's, gaming or a product change has likely broken the link — and an aggregate view will hide exactly that drift. Our cohort analysis guide for PMs covers running this check without a data science team.

Watch for these signals that it's time to swap a proxy out:

• A direct measurement of the real outcome has become newly available or affordable.
• The proxy-outcome correlation has visibly weakened in your most recent cohorts.
• The team can recite the metric's definition but not the outcome it was chosen to represent.
• Two internal teams are optimizing the same proxy toward different, quietly conflicting outcomes.

Key Takeaways

• A proxy metric stands in for an unmeasurable outcome — it's chosen deliberately for correlation, not because it produces a good-looking chart.
• Goodhart's Law and Campbell's Law both predict the same failure: once a measure becomes a target, people optimize the measure instead of the outcome.
• Score candidates on four criteria — correlation strength, gameability, leading-vs-lagging timing, and cost — before picking one out of convenience.
• Pair every proxy with a guardrail metric so gaming shows up as a mismatch between the two, not as a surprise months later.
• Anchor proxies to a specific journey stage and a specific job-to-be-done, not to a vague company-wide sentiment.
• The root cause of most proxy failures is an unlogged assumption — write down what the proxy is standing in for the moment you choose it.
• Retire a proxy on a schedule, testing it across cohorts, rather than leaving it on the dashboard indefinitely out of habit.