A metric is a measure of how well intent is being met. It is what closes the loop — without one, every other node in this model is an assertion.
A metric attaches to what it measures: a strategy, an initiative, or an expectation. Its definition is replaced wholesale on each revision and the history is kept, so "we changed how we counted this in March" is recoverable rather than an argument.
Observations are the readings taken against a metric over time. The metric is the instrument; the observations are what it recorded.
What a metric is for here
Not a dashboard. This tier is not trying to be your analytics stack, and a metric here is not a replacement for the number in whatever system already computes it.
It answers one question:
Can this promise ever be judged — and by what?
A metric is the link between something you said you would achieve and evidence about whether you did. Its value is in the link, not in the number.
The question worth asking before you build
An expectation nothing measures cannot be validated, however well the work goes. That is checkable up front:
What measures this expectation?
If the answer is nothing, you have three choices and all of them are better than proceeding:
- add a metric, if the expectation can be measured;
- accept it deliberately as unmeasured, and record that;
- reconsider whether it is an expectation at all — an unmeasurable one is usually a goal in disguise.
Asking before the work costs a moment. Discovering afterwards that nothing was ever going to tell you whether it worked costs the whole cycle.
Measuring the right side
The most common mistake is measuring your own activity instead of the outcome:
| Measures activity | Measures the promise |
|---|---|
| onboarding flows shipped | share of sign-ups that never ask for a document the user lacks |
| support tickets closed | share of questions answered the same day |
| features delivered | share of users who say they know where the money went |
The left column always improves when you are busy. It improves when you are busy doing the wrong thing, which is exactly when you need a measurement that can tell you.
Attach metrics to expectations wherever you can, because those are the promises. Metrics on strategies and initiatives tell you whether the response is progressing; metrics on expectations tell you whether it is working.
Goals are not measured
Worth restating, because the model enforces it and people try anyway.
You cannot measure to feel in control of my finances. You measure the expectations that proxy it, and you accept that they are proxies — which means keeping an eye on whether they still stand for the goal, or whether the number has started improving on its own while the customer feels no different.
That drift is the single most useful thing this structure can show you, and it is only visible because the goal and its proxy are separate nodes with a link between them.
Working practices
- Define the metric before the work, not after. A metric chosen after delivery is chosen, consciously or not, from among the numbers that moved.
- Few, and load-bearing. A metric nobody has read in six months is not measuring anything; it is decorating something.
- Record the definition, not just the name. "Activation rate" means four different things in most companies. The wholesale-revision history is there so that when the definition changes, older observations are not silently reinterpreted under the new one.
- Let a metric disconfirm you. The point of attaching one to a bet is to find out the bet is wrong while it is still cheap. A metric that could only ever look good is a metric that will tell you nothing.
Read next: Products, journeys and user stories.