1. LandOffer
  2. Blog
  3. All articles
  4. Build a Metric Tree That Connects Team Work to a Product Outcome

Data & Decisions

Build a Metric Tree That Connects Team Work to a Product Outcome

Build a defined outcome-and-input tree, then diagnose a flat result despite more assignments. Exact cohort counts, guardrails and a completed decision keep causal claims limited.

9 min readLandOffer team

Start a metric tree with one defined product outcome, then connect it to a small set of measurable inputs the team can influence. Label which connections are exact arithmetic and which are hypotheses about user behavior. Give each input an owner, an action and a quality constraint. A tree that makes every arrow look causal can turn a useful planning model into an unsupported explanation of success.

This LandOffer.ai guide builds a complete tree for a fictional feedback product, then diagnoses a case where an input improves while the outcome stays flat. Its records are synthetic. The counts and formulas were checked offline; no real product instrumentation, user experiment or business outcome was observed.

Two gardeners examine an espalier fruit tree against a cream wall, with separate harvest baskets below.

Define the outcome before drawing branches

Name the product boundary and the value a user receives before naming a feature release. Write down who receives the value, the event that represents it and the counting window. “More engagement” leaves too much room for disagreement. “Weekly completed requests” is more specific, but may still count work that recipients do not find useful.

The public North Star Framework guide separates a central customer-value metric from measurable, influenceable product drivers and the work beneath them. Amplitude's North Star Playbook similarly connects a central metric to inputs, while describing the tree as containing assumptions that need testing. These sources support a planning structure; they do not validate the arrows in your own product.

Mixpanel's focus-versus-input explanation distinguishes an outcome from more actionable inputs. Its language about relationships should not be read as experimental proof for a newly drawn tree. An owner and a plausible story identify what to investigate; they do not establish causation.

Use “focus outcome” if that is clearer than “North Star” in your organization. A task count, a quality constraint and a financial result have different jobs. Giving all of them the same grand label will not help a team make a trade-off.

A completed metric tree for CritiqueCircle

CritiqueCircle is an invented service where adult writers request feedback on short drafts. A suitable reviewer is assigned, submits feedback and the recipient can confirm whether it was useful. Product lead Esme wants the service to deliver useful feedback reliably, rather than merely assign more reviewers.

The team's chosen outcome is eligible requests with feedback that the recipient confirms useful within seven days of submission. It counts requests, not writers, comments or sessions. “Confirmed useful” is an observable response; it does not mean every unconfirmed response was useless.

An eligible request is a genuine in-scope request from a participating writer. The case excludes staff tests and duplicate follow-up requests before forming the cohort. A cohort contains submissions from Monday through Sunday, using UTC. Each request gets its own seven-day window, so the final weekly result is not read until the last request has matured.

The following completed tree uses request cohort A, submitted November 2–8, 2026. Its final window ends November 15. The recorded analysis date is November 23, after both example cohorts have matured.

48 requests with timely, recipient-confirmed useful feedback
  = 100 eligible requests
  × 80/100 assigned within seven days
  × 60/80 assigned requests receiving feedback within seven days
  × 48/60 timely-feedback requests confirmed useful within seven days

Candidate work below those measured inputs:
  assignment route -> assignment rate                 [hypothesis]
  reviewer capacity and reminders -> timely feedback [hypotheses]
  matching relevance -> useful confirmation          [hypothesis]

Side constraints:
  unwanted-feedback reports; moderator time

The multiplication is an exact decomposition of nested sets in this defined cohort. The work-to-rate connections are hypotheses. Shipping a matching change does not mathematically guarantee a higher assignment rate, and a higher assignment rate does not guarantee a higher final count.

Node Filled definition and cohort A value Owner and possible work Evidence boundary
Focus outcome 48 distinct eligible requests with timely useful confirmation Esme owns the decision and overall review Measures confirmation, not all recipient value or long-term retention
Eligible demand 100 requests after five exclusions from 105 raw records Analyst Beck owns eligibility and duplicate rules Higher demand is not automatically better matching
Assignment rate 80/100 = 80%, assigned within each request's seven-day window Engineer Dune owns routing instrumentation Routing-to-rate effect untested
Timely feedback among assigned 60/80 = 75%, feedback received by day seven from submission Reviewer coordinator Nessa owns capacity review Must not restart the deadline at assignment
Useful confirmation among timely feedback 48/60 = 80%, confirmation also received by day seven Researcher Bram owns response interpretation Nonresponse is not a negative rating
Unwanted-feedback constraint Two reports per 100 eligible requests Moderator Hollis reviews reports and severity Separate from the usefulness count, not subtracted automatically
Moderator workload 20 hours for the cohort's defined follow-up work Hollis owns the workload record Authored case amount, not measured staff productivity

An input owner owns its definition and review, not every cause of movement. Nessa cannot create unlimited reviewer time, and Bram cannot force a recipient to confirm usefulness. Write dependencies next to ownership so accountability does not become a promise to control user behavior.

Make the arithmetic survive implementation

The nested-set identity only works when every term refers to the same requests, observation window and eligibility rule. Using all assigned requests from one week with completions from another can produce a persuasive number that no longer describes either cohort.

For CritiqueCircle, a confirmation belongs in the outcome only if its eligible request also received feedback within that request's seven-day window. A later positive response may be valuable, but it belongs in a separate later-response view. Do not move it backward to make the original cohort look better.

Specify how corrected or duplicate records are handled. Here, each eligible request has one unique identifier. Repeated events do not add additional requests. Five raw records in each cohort are excluded: two staff tests and three duplicate follow-ups. Those exclusions stay visible rather than disappearing silently from the denominator.

The offline fixture creates 210 authored records across two cohorts, applies the same eligibility and time-window rules, and uses exact fractions to reconcile the decomposition. That verifies these synthetic counts and arithmetic. It does not validate an event pipeline, prove real-world deduplication or establish that recipient confirmation is the best value measure for a live service.

Diagnose an input improvement with a flat outcome

Suppose the fictional team introduces a simpler routing screen before cohort B, submitted November 9–15. The last request's window ends November 22. By November 23, the team has the following complete comparison.

Measure Cohort A Cohort B What the comparison establishes
Raw records / exclusions / eligible 105 / 5 / 100 105 / 5 / 100 Same authored denominator size and rules
Assigned requests 80 90 Ten more assignments; rate rises from 80% to 90%
Timely feedback 60 60 No increase in the timely-feedback count
Timely feedback among assigned 75% 66⅔% The conditional rate falls as assigned volume grows
Timely useful confirmations 48 48 Defined focus outcome stays flat
Confirmation among timely feedback 80% 80% Same observed fraction in the synthetic cohorts
Unwanted-feedback reports / eligible 2/100 5/100 More reports under the same denominator
Moderator hours 20 25 More authored workload; no causal allocation established

For cohort B, the exact identity is 100 × 90/100 × 60/90 × 48/60 = 48. Rounding the middle rate to 67% would produce 48.24 if multiplied literally. Keep exact counts for reconciliation and round percentages only for display.

The immediate diagnosis is narrower than “routing failed.” More requests were assigned, but that did not produce more timely feedback in this comparison. A reviewer-capacity constraint is one hypothesis. Different draft difficulty, reviewer mix, recipient response behavior or another contemporaneous change could also matter. The two weekly cohorts were constructed for this example, not produced by a randomized experiment.

The team cannot claim the routing screen caused either the higher assignment rate or the flat outcome from this record alone. The tree directs the team to the gap between assignment and timely feedback. It does not identify the cause of that pattern.

Hands hold a pear above a basket, with blemished fruit set aside and a brass balance scale nearby.

Attach a decision, not just a new target

Before the fictional comparison, Esme and Hollis set a review trigger of no more than three unwanted-feedback reports per 100 eligible requests. This is a local authored decision threshold, not a universal product-safety standard. The team also records that any serious report requires individual review regardless of the total.

Cohort B reaches five reports per 100 and crosses that trigger. Counting 48 useful confirmations does not cancel the reports, and averaging the two cohorts to make the number look less alarming would hide the newer result. The defined constraint has a job precisely when another metric improves.

The completed November 23 decision record says:

Esme will hold wider routing-screen rollout while Hollis reviews all five cohort B reports through the normal moderation process. Dune will reconcile assignment and feedback timestamps for the 30 assigned requests without timely feedback, using authorized records. Nessa will compare the available reviewer slots with the 90 assignments and propose an assignment limit for the next review. Bram will examine confirmation nonresponse separately from negative responses. We have not established a causal effect of the routing change. The next review is November 30; no new assignment-rate target or automatic expansion is approved today.

The 30 requests come from 90 assigned minus 60 timely completions. They are cases for investigation, not 30 proven capacity failures. Their state after day seven and the reason for delay need separate evidence. Some may have been completed later; the original outcome definition deliberately does not count those as timely.

For a possible next test, the team proposes a bounded assignment-capacity comparison with clear eligibility, outcome and report monitoring. The design and required approvals are still pending. No experiment has run, no sample size has been justified and no expected lift is asserted.

Keep the tree useful when evidence changes

Review definitions before adding more branches. If an input is a broad activity nobody can act on, identify a more specific candidate or remove the implied ownership. If several inputs count the same requests, label overlap instead of adding their counts as though they were independent contributions.

Keep a distinction between a leading signal and a mathematical factor. Assignment rate may provide an early signal, but the exact identity also depends on later conditional rates. A request assigned on day six has little time remaining under the seven-day outcome window. Changing that window would create a different metric, not simply improve the old one.

Version the tree with definitions, owners, unresolved relationships and the date of the last review. Retain the previous version when the strategy or evidence changes. Do not rewrite past denominators or remove an inconvenient constraint without explaining what changed and why.

For your own product, start with one customer-value event, one mature cohort and the actual numerator behind every rate. Reconcile any arithmetic connections, mark the work-to-outcome hypotheses and choose the decision each owner will make when the pattern disappoints. Save the definitions, exact cohort counts and resulting decision together so the next review can reproduce the comparison. For related current-work analysis guides, visit LandOffer.ai.

Sources and Further Reading

  • North Star Framework: A Practical Guide — Central value metric, product drivers and work connections. Undated; checked October 28, 2026. A framework description, not proof of this case's relationships.
  • Amplitude: The North Star Playbook — Chapter 1, printed pages 6–8, distinguishes central metric and inputs and identifies assumptions in the tree. Publication date not established from the inspected pages; checked October 28, 2026.
  • Mixpanel: Focus metrics vs. input metrics — Outcome/input terminology; published January 29, 2026, last edited June 9, 2026; checked October 28, 2026. Its general relationship language is not a causal test of CritiqueCircle.