What a KPI tree is
A KPI tree is a hierarchy of metrics with one outcome at the root and a rule that connects each level to the one above it. The rule is arithmetic: the children of a node, combined by the stated formula, give the parent. That constraint is what separates a tree from a scorecard. A scorecard lists metrics a team cares about and leaves the reader to guess how they relate; a tree states the relationship, so that a movement in the parent must be explained by a movement in at least one child.
The tree form is widely used in management accounting and analytics and no single origin is claimed for it here. Its value for marketing comes from the three levels it insists on. The outcome is what the business wants and nobody can touch directly. The drivers are the metrics that compose it, still too abstract to act on. The inputs are the leaves: metrics a named team can change through its own work this week. A tree that stops at drivers explains what happened without saying who should do what; a tree that reaches inputs is a plan.
What the tree makes visible is attribution of change within the arithmetic. Revenue fell: the tree shows whether orders or order value fell, and beneath orders whether traffic or conversion fell, and beneath conversion which segment. What it hides is anything the formulas leave out: seasonality, a competitor's launch, a change in the traffic mix that pushes a rate down without any page getting worse. The tree says where; it does not say why. The metric definitions the leaves rely on should match the KPI framework reference, so that a tree built by one team reads the same to another.
Levels and arithmetic
| Level | What sits here | Relationship to the level above | Who owns it |
|---|---|---|---|
| Outcome | The one number the tree exists to explain: revenue, qualified leads, active customers | None; it is the root | The business, not a team |
| Driver | The metrics that compose the outcome by a stated formula: orders and order value; sessions and conversion rate | Multiplicative or additive, written beside the split | Usually shared between teams |
| Input | The leaves a named team can move: sessions from a channel, conversion rate on a page group, items per basket | The same arithmetic, one level further down | One team, named |
| Diagnostic | Measurements that explain an input without being levers: device share, form error rate, keyword position | Not in the arithmetic; read when the input above moves | Whoever reads the tree |
Two forms of arithmetic cover almost every split. A multiplicative split takes a count and a rate, or a count and a value per unit: orders are sessions multiplied by conversion rate (CVRShare of visits or clicks that convert.), revenue is orders multiplied by average order value (AOVAverage revenue per order.). An additive split takes a total and divides it into exhaustive, non-overlapping parts: sessions by channel, revenue by new and returning customers, accounts by plan. In a free-trial subscription tree, new paying customers split the multiplicative way: trial starts multiplied by the trial-to-paid conversion rateShare of free trials that became paying subscriptions..
Each form carries a rule that keeps the tree honest. In a multiplicative split, every child is measured on the same population and period as its siblings, or the formula does not balance. In an additive split, the parts must cover the whole and not overlap, or the tree quietly loses the ability to explain the parent; where part of the total is unmeasured, an explicit "unattributed" child is better than a gap. A tree whose arithmetic balances on last month's real figures is ready to use; one that does not is a diagram.
Guardrails
Guardrails are the metrics the tree would otherwise let a team break. They sit outside the arithmetic, so they do not explain the outcome, but they constrain how the drivers may be moved. An orders tree without a refund-rate guardrail will report success from a promotion that is later reversed by returns; a lead tree without a quality guardrail will report success from a form change that fills the pipeline with people sales cannot use; an acquisition tree without a cost ceiling (CACCost to acquire one new customer.) will report growth that the finance team later calls a loss.
A guardrail has three parts: the metric, a threshold it must not cross, and an owner who is told when it does. The threshold comes from history and from the business's tolerance rather than from a benchmark; it is the level at which the business has previously regretted continuing. Common marketing guardrails are refund and cancellation rates, unsubscribe and complaint rates, lead quality as judged by the people who receive the leads, acquisition cost against lifetime value, page speed, and brand-search volume as a check that performance activity is not eroding the brand. For a content programme the content goals and KPIs guide shows guardrails of a different kind: accuracy, refresh rate and editorial standard beside the traffic tree.
Guardrails are read in the same view as the tree, not on a separate dashboard, because an initiative that moved a driver and breached a guardrail has not succeeded, and the two facts need to be seen together.
Worked tree
A monthly recurring revenue tree for a small-business invoicing app
MRR = Paying accounts × Average revenue per account
Paying accounts (end of month) = Accounts at start + New paying − Churned
New paying = Trials started × Trial-to-paid rate
Trials started = Site visitors × Trial sign-up rate
In month one the app starts with 3,000 paying accounts at £24.00 each, giving £72,000 MRR. Visitors are 60,000 with a 2.5% sign-up rate, so 1,500 trials; 20% convert, so 300 new paying accounts; churn is 3%, so 90 leave. The month closes at 3,210 accounts and £77,040 MRR.
In month two visitors rise to 66,000, sign-up and trial-to-paid rates hold at 2.5% and 20%, so 330 new accounts; churn holds at 3%, so 96 leave; the month closes at 3,444 accounts. Yet MRR falls to £75,770. Reading the tree from the top: accounts rose, so the fall sits in the other driver, average revenue per account, which has dropped from £24.00 to £22.00. The diagnostic beneath it, plan mix, shows the share of accounts on the cheapest plan rose from 40% to 55%. The input that moved is the sign-up rate of a new comparison page ranking for sole-trader queries, which brought a wave of small customers onto the entry plan. The lever is not the page; it is plan packaging and the targeting of the page, and the guardrail check shows acquisition cost per paying account unchanged at £38.
The example shows the tree doing exactly what it is for: a counter-intuitive result, more customers but less revenue, resolved to a single branch in three readings, with the guardrail confirming that the acquisition itself was sound. It also shows the tree's blind spot. The arithmetic located the plan-mix shift; it took a person to connect that shift to a new page, because the page is not in the tree.
Finding the lever
A tree is read upwards for planning as well as downwards for diagnosis, and the upward reading is how it finds the lever. For each candidate leaf, four questions. How much headroom does it have: a conversion rate near the best the site has ever achieved has less room than one that has fallen behind its own history. How much weight does it carry: a small change in a large branch moves the outcome more than a large change in a small one, and the arithmetic says exactly how much. Who owns it: a leaf with no named team is not a lever however attractive its headroom. And what would moving it cost, in time, money and risk to a guardrail?
The comparison across leaves is the point. Without a tree, teams argue for the lever they happen to own; with one, the argument is settled by the arithmetic, which is why the north star metric and OKRs framework relies on a tree beneath the north star to decide which inputs each team commits to, and why the measurement plan framework uses the tree to choose KPIs that can be acted on rather than merely reported.
The lever also depends on the objective the tree serves. A revenue or lead tree finds its levers in the period's funnel; a retention tree finds them in the first weeks of a customer's life and has to be read by cohort age, which is why the KPI tree for retention changes the root, the arithmetic and the owners rather than only the label at the top.
One discipline matters more than the rest: distinguish the leaf that moved from the lever that will fix it. In the worked tree, revenue per account moved, but the lever sat two branches away in how the new traffic was targeted. A rate that falls because the mix of the population beneath it changed is a symptom, and pushing on the rate itself wastes a quarter.
Limits
The arithmetic is an identity, not a cause. The formulas guarantee that the children explain the parent numerically; they say nothing about why any child moved, and a team that treats a balanced tree as an explanation will stop asking questions one step too early.
External factors live outside the tree. Seasonality, a competitor's pricing, a platform policy change and a news event all move leaves without appearing anywhere in the structure, and a tree read in isolation attributes their effects to whichever leaf they landed on.
Branches interact. More traffic from a new channel lowers the blended conversion rate because the new visitors are colder, not because anything got worse; the tree shows two leaves moving in opposite directions and leaves the reader to work out that one caused the other.
A tree has one outcome, and businesses have several. Optimising the revenue tree can quietly damage the retention tree, which is why guardrails exist and why a second tree is sometimes the only honest answer.
Trees decay. Definitions drift, new channels appear, segments are renamed, and a tree that was accurate at the start of the year balances less well each month until nobody trusts it. It needs an owner and a re-balancing check at each planning cycle, and it needs its leaves kept aligned to the shared metric definitions.
Finally, the precision is seductive. A tree that balances to the pound invites decisions with the same confidence, when the inputs are estimates, the attribution is a stance, and the segments are approximate.
Related frameworks
| Framework | Reach for it when |
|---|---|
| North star metric and OKRs | The root of the tree needs to be the metric the whole organisation shares |
| Measurement plan | The leaves need event definitions, owners and a cadence |
| AARRR | A product business needs lifecycle stages before the multiplicative chain can be drawn |
| Marketing funnel | The tree's funnel-chain branch needs stage definitions and the constraint found |
When the tree serves a sales team, the KPI tree for lead generation roots it in qualified pipeline, carries the lag between lead and opportunity and hands sales-side leaves to their owners. For apps, the KPI tree for app installs puts activated users at the root, gives the store listing its own branch and labels the leaves that privacy rules turn into estimates. Where visits are the goal, the KPI tree for website traffic roots the tree in engaged visits, adds channel branches together and separates search demand from the share the site wins.
Common mistakes
- Stopping at drivers. A tree whose leaves no team owns explains the past and plans nothing.
- Arithmetic that does not balance. Children on different populations or periods, additive parts that overlap, gaps left unlabelled.
- No guardrails in the same view. A driver moved and a guardrail breached is reported as a win.
- Pushing on a symptom. A rate that fell because the mix changed is fixed one branch over.
- Two outcomes in one tree. Shared drivers are noted on both trees, not merged into one.
- Never re-balancing. A tree built once and trusted for a year is quietly wrong by spring.
- Reading a balanced tree as an explanation. The arithmetic says which leaf moved; research and experiments say why.
