Analytics · 9 min read · 2026-10-18
Goodhart's Law: When Your SaaS Metric Becomes a Target
A product team sets an activation target for the quarter and hits it: the activation rate climbs from 34% to 41%, the dashboard is green, and everyone moves on. Two quarters later retention and revenue have not budged, and someone finally asks why a 20% jump in activation produced nothing downstream. The answer is in the definition change nobody logged: midway through the quarter "activated" was quietly loosened from "completed the core workflow" to "clicked into the workflow," and the number went up because the bar came down. The metric did exactly what Goodhart's Law predicts a metric does once it becomes a target.
This is an anti-pattern teardown of the single most corrosive failure mode in a metrics-driven org: the proxy you chose to represent value stops representing it the moment you start steering by it. It is adjacent to the vanity-metrics teardown and the north-star selection framework, but it is a distinct problem: here you picked a good metric, and the act of targeting it is what broke it.
A measure under pressure stops measuring
Goodhart's Law — "when a measure becomes a target, it ceases to be a good measure" — originates with the economist Charles Goodhart in 1975 and was sharpened into that phrasing by the anthropologist Marilyn Strathern. The underlying mechanism is simple and inescapable: every metric is a proxy for an outcome you cannot measure directly. Activation proxies "the user reached value." Weekly actives proxies "the product is habitually useful." NPS proxies "customers would recommend us." The proxy and the outcome move together right up until you attach a goal or a bonus to the proxy — at which point, as one explainer notes, people optimise for what they are measured on, and the cheapest way to move the proxy is rarely the thing that moves the outcome.
The corruption is usually not fraud. It is a sequence of small, defensible choices, each of which nudges the number toward the target: a slightly looser event definition here, a reminder email that lifts a login count there, a trial extension that improves conversion on paper. No single step looks like gaming, which is exactly why Goodhart damage accumulates unnoticed until the proxy and the outcome have visibly decoupled. The discipline that catches it is the same one behind a single source of truth for metric definitions: if the definition can drift silently, the target will drag it.
How Goodhart shows up on a SaaS dashboard
The most common SaaS instance is the activation metric that improves by redefinition rather than by faster time-to-value — the opening example. Because the boundary between "signed up," "activated," and "adopted" is a judgement call, an activation target creates constant downward pressure on where that line sits, which is why the activation-versus-adoption boundary needs to be instrumented and frozen, not negotiated each quarter. The same dynamic hits signups (a growth target met with low-intent or duplicate accounts), weekly actives (propped up by a notification that drives a hollow open), and NPS (lifted by question framing rather than service, a pattern the research literature repeatedly flags).
What makes these dangerous is that they are invisible in the metric itself. The activation rate that rose by redefinition looks identical to the one that rose because onboarding genuinely got faster; the dashboard shows the same green line. You can only distinguish them by looking at whether the outcome the metric proxies moved in step — did the extra activated users retain, did the extra signups convert — which is why a Goodharted metric is caught downstream, not on the chart that celebrates it. This is the same blind spot that makes a leading indicator worthless once it is being optimised directly rather than tracked.
Why proxies rot when you push on them
The reason a targeted metric degrades is that the map between proxy and outcome is many-to-one: there are many ways to move the number, and only some of them move the thing the number was standing in for. Before you attach a goal, behaviour is distributed across all those paths roughly in proportion to what actually creates value, so the proxy tracks the outcome. After you attach a goal, behaviour concentrates on whichever path moves the number cheapest — and the cheapest path is almost never the one that also moves the outcome, because if it were, you would not have needed the target to force it.
This is why "just pick a better metric" does not solve Goodhart: any single metric you elevate to a target becomes subject to the same pressure, however well chosen. It is a different failure from a bad proxy. A bad proxy never mapped to the outcome and you fix it by choosing a better one; a Goodharted proxy mapped fine and broke under load, and you fix it by changing how it is used. Treating the two as the same problem is how teams churn through north-star metrics every year, each new one decaying the moment it is targeted, when the issue was never the choice of metric.
Guardrail one: pair every target with a counter-metric
The highest-leverage defence is to refuse to ever set a single-number target. A lone metric is the most gameable object in analytics; a metric plus a guardrail it cannot break makes the cheap games visible. Activation is allowed to rise only if retention does not fall. Signups are allowed to rise only if trial-to-paid does not drop. Weekly actives are allowed to rise only if session depth or a value-event rate holds. The guardrail is the outcome-side check that a gamed proxy will fail, which converts an invisible corruption into a visible contradiction on the same dashboard.
This is the same construction as an experiment counter-metric, lifted from the A/B context to the KPI context: a small, permanent cost you pay so a confident-looking number cannot quietly be a loss. Balanced measurement — several linked metrics rather than one headline figure — is the most consistent recommendation across the Goodhart literature, which argues for multiple balanced measures tied to clear outcomes rather than single-number control. A driver tree is a natural home for this: the north-star at the top, the levers beneath it, and the guardrails that keep any one lever from being pulled at the others' expense.
Guardrail two: measure for learning, and design out the shortcut
The second defence changes the metric's job. A metric used purely for evaluation — hit the number or else — maximises the incentive to game it; a metric used for learning, where the question is "what is this telling us," largely removes that incentive because nobody is rewarded for the number alone. You cannot always separate the two, but you can choose which metrics carry stakes and which are diagnostic, and keep the most game-prone proxies on the diagnostic side.
The strongest version designs the metric so the easiest way to game it is the behaviour you wanted. The most-cited example is Microsoft's EngThrive programme, which measured time-to-first-pull-request for new hires; when managers reacted by assigning trivial first PRs to move the number, the designers treated that as success rather than cheating, because a small early PR inside the real workflow is precisely the onboarding behaviour the metric existed to encourage. The SaaS translation: define activation as an event that is only reachable by doing the valuable thing, so the shortcut to the target and the goal of the target are the same action.
A Goodhart audit you can run this sprint
Take your three most-watched metrics and, for each, ask four questions. One: what outcome is this a proxy for, and can I see whether that outcome moved in step over the last two quarters? Two: is there a goal or an incentive attached, and if so, what is the cheapest way to move the number without moving the outcome? Three: what guardrail would that cheap game break, and is it on the dashboard next to the metric? Four: is the metric definition frozen and owned, or can it drift under pressure? Any metric that fails question one or three is a Goodhart risk already in motion.
The fix is rarely a new metric; it is a guardrail, a frozen definition, and a decision about whether the number carries stakes or informs learning. The teams that get burned are not the ones with the wrong KPIs — they are the ones who elevated a reasonable proxy to a target and never asked what the cheapest way to move it was. Run the four-question audit before you set next quarter's goals, treat it with the same seriousness as a vanity-metric review, and the green dashboard stops drifting away from the business it is supposed to describe.
Frequently asked questions
The questions below cover what Goodhart's Law is, how it surfaces in SaaS product analytics, how to stop metrics from being gamed, how a Goodharted metric differs from a vanity metric, and whether you can design a metric that resists the effect.
FAQ
What is Goodhart's Law?
Goodhart's Law is the principle that when a measure becomes a target, it stops being a good measure. The economist Charles Goodhart made the point about monetary policy in 1975, and the anthropologist Marilyn Strathern later gave it the now-standard phrasing. The mechanism is that any metric is only a proxy for the outcome you actually care about, and once people are rewarded for moving the proxy, they optimise the proxy directly — including in ways that leave the real outcome unchanged or worse. In SaaS analytics it explains why a team can hit its activation or engagement target quarter after quarter while retention and revenue fail to follow: the number went up because the number was the goal, not because the underlying value did.
How does Goodhart's Law show up in SaaS product analytics?
It shows up whenever a proxy metric is attached to a goal or an incentive and the team optimises the proxy instead of the outcome. Classic patterns: an activation rate that rises because the activation event was redefined to something easier, not because users reach value faster; a signup number inflated by low-intent or duplicate accounts; an NPS that improves because of survey question framing rather than better service; a 'weekly active' count propped up by a notification that drives a hollow open. In each case the dashboard improves and the business does not, because the metric was gamed — sometimes deliberately, more often through small, reasonable-looking choices that each nudge the number toward the target.
How do you prevent metrics from being gamed?
Pair every target with a guardrail that gaming would break, measure for learning rather than pure evaluation where you can, and define the metric tightly so it cannot be quietly redefined. A single-number target is the most gameable object in analytics; a target plus a counter-metric (activation up without retention down, signups up without trial-to-paid down) makes the cheap games visible. Tie each metric to the outcome it proxies, review the definition on a fixed cadence so drift is caught, and involve the people doing the work in choosing metrics, since they know best how a given number could be gamed. The strongest version, used by some engineering-metrics programmes, designs the metric so that gaming it produces the behaviour you actually wanted.
What is the difference between a vanity metric and a Goodharted metric?
A vanity metric is a number that looks impressive but was never tied to a decision or an outcome — total registered users, cumulative pageviews. A Goodharted metric is different: it started as a genuine, useful proxy for an outcome and then degraded because it was turned into a target and optimised directly. The distinction matters for the fix. You solve a vanity-metric problem by choosing a better metric that maps to a decision. You solve a Goodhart problem by changing how an otherwise-good metric is used — adding guardrails, decoupling it from incentives, or redefining it — because the metric itself was fine until pressure corrupted it.
Can you design a metric that resists Goodhart's Law?
You cannot make a metric fully immune, but you can design one so that the easiest way to game it is also the behaviour you want. The most-cited example is Microsoft's EngThrive programme, which measured time-to-first-pull-request for new engineers; when managers started assigning trivial first PRs to hit the number, the designers treated that as a feature, because a small early PR inside the real workflow is exactly the onboarding behaviour the metric was meant to encourage. The general technique is to define the metric so that the shortcut to moving it requires doing the valuable thing, and to back it with a guardrail that catches the shortcuts that do not.