What Goodhart's law says
A measure earns its place because it tracks an outcome. Cost per lead is watched because cheaper leads usually mean more efficient acquisition; email open rate because opens usually mean attention; satisfaction scores because satisfied customers usually stay. The word doing the work is "usually". The link holds while nobody is pushing on the measure directly. Once the measure is a target, with bonuses, reviews or agency fees attached, effort flows to whatever moves it most cheaply. Some of that effort improves the outcome. Some of it moves the measure alone, and the link that made the measure useful weakens.
There are several ways this happens in practice. Definitions get stretched: a lead score threshold is lowered, so more people count as marketing qualified. Mix shiftsA change in a blended metric caused by a change in the mix of segments, such as devices or sources, rather than by any segment performing differently. towards the easy cases: budget moves to audiences that convert anyway. Work goes to what is measured and away from what is not: a team judged on clicks stops investing in the brand work that has no click. And sometimes the number is simply manipulated, though much of the drift comes from reasonable people doing what they were asked.
The guide to objectives and KPIs makes the same point in its section on incentives, and asks for constraints that stop a target being hit in ways that damage the goal. This page is about why that happens and how the design of a target changes how badly it happens.
How marketing metrics get gamed
| Target | How it is gamed | What it costs | Guardrail |
|---|---|---|---|
| Lead volume | Lower the qualification threshold, gate more content, count anyone who downloads anything | Sales time spent on leads that never buy; sales stops trusting marketing's leads | Share of leads sales accepts; pipeline created from marketing leads |
| Return on ad spend | Shift budget to brand search and retargeting of people already about to buy | Spend that adds little; prospecting starved; growth stalls | Separate targets for new-customer acquisition; periodic holdout tests |
| Cost per acquisition | Buy cheap, low-intent audiences; cut prospecting | Customers who churn or never repeat | Retention or repeat rate of the acquired cohort |
| Viewability rate | Buy placements built to stay on screen rather than to be read | Ads seen by nobody who matters; made-for-advertising inventory | Outcome measures such as lift or conversions; checks on where impressions ran |
| Click-through rate | Curiosity-gap headlines and misleading creative | High bounce, brand trust eroded | Conversion rate or engaged time after the click |
| Email open rate | Subject lines that overpromise; mailing only the most engaged | Unsubscribes and complaints; a shrinking list | Unsubscribe and complaint rates; revenue per recipient |
| Organic traffic | Thin pages for loosely related queries | Visitors with no intent; crawl and maintenance cost | Conversions or sign-ups from organic visits |
| Satisfaction score | Surveying only happy customers; prompting for a high score | A score that no longer warns of problems | Response rate; unprompted complaints; churn |
Several of these targets were reasonable choices when they were set. The problem is not the metric but the pressure on it combined with the absence of a counterweight. The branded versus non-branded search comparison shows the return on ad spend case in detail: a blended return borrows the flattering figure from people who were already looking for the brand. The guide to viewability and attention shows the inventory case: a viewability floor that becomes a target pulls a campaign towards sites built to score well rather than to be read. Customer service shows the same pattern with speed: the guide to customer service metrics pairs a fast first reply with repeat contacts on the same problem, so replies that answer nothing cannot pass as good service. An average time to first replyThe time from a customer's first message to the first reply from a person; read beside repeat contacts, as a quick reply that solves nothing still scores well. set as a target shows how: a quick holding message lowers it without bringing the answer any closer.
Designing targets that hold
Target the outcome, or as close to it as the team can influence. A KPIA metric tied directly to an objective. several steps away from revenue, such as impressions or clicks, gives the most room for drift. The closer the target sits to the outcome, the fewer shortcuts exist. Where the outcome is too slow or too shared to target directly, choose the measure at the handover point to the next team, because that team has a reason to reject poor work. The comparison of MQLs and SQLs shows why: an MQLA lead that marketing judges ready for sales follow-up against agreed criteria. is defined by marketing alone, while a sales-accepted lead has passed a second judgement.
Pair every target with at least one guardrail. A guardrail metricA metric an experiment must not worsen beyond an agreed bound, watched beside the primary metric that decides the test. is a measure that must not get worse while the target is pursued: lead acceptance beside lead volume, cohort retention beside acquisition cost, complaint rate beside open rate. The guardrail does not need its own target; it needs a limit and someone watching it.
Keep definitions out of the targeted team's hands. Whoever is judged on a number should not be the only person who can change how it is counted. Lead scoring rules, attribution settings and survey sampling should change through a recorded decision that someone else signs off.
Separate the few targets from the many diagnostics. The objectives and KPIs guide keeps one primary KPI per objective and puts everything else in a diagnostic layer that explains results without judging them. Metrics in that layer are under less pressure and stay more honest, which makes them useful for spotting drift in the target.
Check the link regularly. At each review, compare the target with the outcome it was chosen to predict. If the target rises and the outcome does not, the link has weakened. Reading a sample of the records behind the number, such as a batch of leads, a list of placements or a set of survey responses, catches drift that aggregate figures hide. When the number that has drifted is a company's north star rather than one team's target, the guide to changing a north star metric works through tightening its definition, refocusing an input or replacing it.
Be careful with external numbers. The data library's benchmarks section warns against turning an industry average into a target; a benchmark is a reference point, and a team pushed to match it will find the easiest route to the number, whether or not that route suits the business.
Worked example (illustrative)
An accounting-software firm moves its target from leads to pipeline
Grenmallow Software, an invented maker of accounting software for small firms, pays its marketing team a quarterly bonus on lead volume, with a target of 2,400 leads a quarter. Over three quarters, leads rise from 2,100 to 3,050. In the same period, the share of leads that sales accepts falls from 41% to 22%, and pipeline from marketing leads is flat at about £610,000 a quarter.
A review of 200 recent leads finds that most new volume comes from a gated template library and a webinar series aimed at students. The lead score threshold had also been lowered from 60 to 45 points after a dashboard change.
The firm changes the target to sales-accepted pipeline, set at £700,000 a quarter. Lead volume stays on the dashboard as a diagnostic. A guardrail is added: the acceptance rate must stay above 35%, and the scoring threshold can only change with the sales director's sign-off. In the next quarter, leads fall to 2,300, acceptance rises to 38% and sales-accepted pipeline reaches £680,000.
Limits and criticisms
The law is a warning, not a method. It says that pressure distorts measures, but it does not say which measures to use, how much pressure is too much or what to do instead. Teams still need a way to choose metrics, and Goodhart's law only tells them to expect trouble with whatever they choose.
Overcorrecting is its own failure. A team that takes the law too seriously can conclude that every target will be gamed and set none, or set so many balancing measures that nobody knows what success looks like. The result is no accountability and no direction, which is usually worse than a target that drifts a little. The law is an argument for designing targets carefully, not for avoiding them.
Some measures are hard to game and safe to target. Cash collected from paying customers, net of refunds, leaves little room for shortcuts that do not also help the business. Measures audited by another team, or verified by a controlled test, also resist drift. Treating every metric as equally fragile ignores these differences.
The law can also be used to dismiss inconvenient numbers. "That metric is being gamed" is sometimes true and sometimes a way to avoid a result someone dislikes. The claim should be checked against the records, like any other. And a metric that was never linked to an outcome, such as a vanity metricMetrics that look impressive but do not inform a decision or track a business outcome. like follower count, is not a Goodhart problem at all; it was never a good measure to begin with.
Related frameworks
The north star metric and OKRs framework builds a whole system around one metric, which makes Goodhart's law its main risk; that page's inputs and checks are one answer to it. The KPI tree places guardrails beside the branches of an outcome, so the cheapest route to one branch does not damage another. The measurement plan framework records who owns each metric's definition, which is one of the safeguards above. And control charts help tell a real improvement from a sudden jump that coincides with a change in how something is counted.
Common mistakes
- Targeting the easiest number to move. Impressions and clicks are easy to hit and easy to hit badly.
- Targets without guardrails. A target with nothing beside it rewards whichever shortcut is cheapest.
- Letting the targeted team own the definition. If the people judged can change how the number is counted, it will drift.
- Blending figures. A blended return or a combined lead count lets the easy part flatter the hard part.
- Never checking the link. A target that no longer tracks the outcome can be hit for quarters before anyone notices.
- Adopting a benchmark as a target. An outside average says nothing about what this business can or should achieve.
- Abandoning targets altogether. Fear of gaming is not a reason to have no accountability.
