What is a north star metric?

A north star metric is a company-level answer to "what does progress look like". Nights booked for a marketplace, weekly active teams for a collaboration product, hours watched for a streaming service. Its purpose is coordination: when six teams are choosing between roadmaps, a shared definition of progress stops them optimising six different things that pull against each other.

The distinction that matters most is between this and a primary metric. A primary metric decides one experiment — it must be sensitive enough to move detectably in two or three weeks, and it is nominated per test. A north star operates over quarters at company scale and is deliberately slow, broad and stable. It is almost always far too insensitive to serve as an experiment's decision rule: a checkout improvement that adds real value will not visibly shift weekly active teams, and a test powered to detect that would need more traffic than the company has.

The connection between them should be explicit rather than assumed. A well-constructed programme can draw a line from each experiment's primary metric up through intermediate outcomes to the north star, and can say roughly what a 1% move in the primary is worth at the top. Most organisations cannot draw that line, which is when the north star becomes decorative — quoted in all-hands presentations and connected to nothing anyone does on a Tuesday.

A good one has three properties. It reflects value delivered to customers rather than value extracted from them, so that improving it is genuinely good news. It is hard to move without doing real work, which rules out most vanity counts. And it is resistant to gaming, because any metric given organisational authority will be optimised whether or not anyone intends to cheat. "Hours watched" is the standard cautionary example: it can be improved by making content genuinely better and equally by making it harder to stop watching, and the metric cannot tell those apart.

The usual accompaniment is a small set of counter-metrics that must not degrade — the organisational equivalent of guardrail metrics. A north star for engagement paired with a watch on churn and support contacts is much harder to game than one on its own, because the cheap ways to move the headline number tend to show up immediately in the counterweights.

A streaming service uses hours watched per subscriber per month as its north star, currently 34.2. A team wants to know whether their autoplay-next-episode experiment, which lifted session length by 4%, can be evaluated against it — and what the change is worth at company level.

This is the ordinary case rather than an extreme one, and it is why the two kinds of metric must not be conflated. The experiment is judged on session length, which moves detectably in two weeks; the north star is the thing session length is believed to feed, and the belief is what needs periodic validation rather than per-experiment testing. The second half of the question — what is it worth — is where the chain earns its keep: 4% on session length maps to roughly 1.2% on monthly hours, which is a defensible forecast only if that mapping has been estimated from past shipped changes rather than assumed. There is also a gaming caution sitting in plain sight here. Autoplay increases hours watched and can do so by making it harder to stop rather than by delivering more value, and hours watched alone cannot distinguish those. Whether this ships should depend on the churn and satisfaction counter-metrics as much as on the headline.