Two-sided → two North Stars
canonTwo-sided marketplaces need TWO North Stars — supply + demand liquidity. A single number hides the constrained side.
North Star · AARRR · HEART · AI stack
Instrument one metric, not the table — and every metric is a bet that it still tracks real value.
A North Star is one value-delivery metric plus 3–5 movable inputs, each with a guardrail. Pick a framework and instrument it; don't dump the whole menu and call it measurement.
One value-delivery metric
North Star01
Movable input
02
Movable input
03
Movable input
04
Movable input
Spotify = time spent listening
Airbnb = nights booked
Slack = messages sent
3–5 movable inputs — survivorship caveat: those look obvious only in hindsight.
Above the fold
The three moves that carry the measurement.
01
Instrument ONE metric, not the table.
Pick a single framework and wire it up — a North Star is one value-delivery metric plus 3–5 movable inputs. Reproduce the canon from the name; don't dump the whole table.
02
Every metric is a proxy — pair a guardrail.
Name how it could rise while value falls, then always ship a paired guardrail (what must NOT degrade). Goodhart: a North Star that becomes the target gets gamed.
03
Once the product acts, count jobs not events.
For agentic/AI features, measure jobs completed, not events fired. "Messages sent" rewards a chatbot that loops — track resolved-without-handoff and task success.
North Star · Amplitude
One outcome the team aligns on — plus what protects it.
One value-delivery metric with 3–5 movable inputs aligns a team on a single outcome. It is a proxy, not the goal: re-validate quarterly that it still correlates with retention and revenue.
Proxy, re-validated
canonThe North Star is a bet on value. Re-validate quarterly that it still correlates with retention/revenue — a metric that drifts from value silently misleads.
Survivorship caveat
houseSpotify's "time listening", Airbnb's "nights booked", Slack's "messages sent" look obvious only in hindsight. Derive yours from value delivered, not by analogy.
Guardrail taxonomy · pair at least one
Quality / health
Does the core experience hold?
Customer sentiment
Are users still glad they came?
Trust & safety
Is growth clean, not gamed?
Unit economics
Does each unit still pay?
Retention / churn
Do they stay, or leak away?
AARRR / Pirate Metrics · McClure
The funnel — and the stage most teams forget.
Trace a retention problem upstream — it's often an Activation / time-to-value miss, not Acquisition. Pair any retention fix with a guardrail + North-Star input so you don't fix one stage while degrading another.
AAcquisitionSignups / downloads / MAU — vanity without what follows.
AActivationThe "aha" within ~7 days — the real gate. Value, not vanity.
RRetentionWeek-4 cohort retention — trace a leak here back upstream.
RRevenueWhere value gets captured — still value, not a headline.
RReferralUsers bringing users — compounding, if the loop is real.
RResurrectionWin-back — the often-missed stage. Churned isn't gone forever.
Growth accounting
Quick Ratio = (new + resurrected) ÷ churned MAU
Below 1 you're leaking faster than you fill — regardless of how strong acquisition looks.
Vanity vs value
Acquisition — signups, downloads, MAU — is vanity; Activation, Retention and Revenue are value. Contrast vanity signups against an activation "aha" within ~7 days and week-4 cohort retention before you celebrate a growth chart.
HEART · Google — and the bands
Feature-level UX, and the numbers that anchor realism.
HEART — Happiness / Engagement / Adoption / Retention / Task-success — runs through Goals → Signals → Metrics at the feature/UX level, not the company level. Pick the 1–2 rows the feature actually moves.
Goals → Signals → Metrics
GoogleName the goal, find an observable signal, then define the metric. Alt lenses: HEART reads feature-UX regression; Hook / trigger-decay reads habit loss.
Counting-method drift
warningN-day vs rolling vs bracketed retention give different curves. Hold the definition fixed across benchmark comparisons and tool migrations, or you'll chase artifacts.
Retention benchmark bands · directional, point-in-time — verify current
~20% / 50%+
DAU/MAU stickiness — ~20% is decent; 50%+ is daily-habit territory.
Directional · verify current
≥40%
Sean Ellis PMF probe — ≥40% of users "very disappointed" without the product signals product-market fit.
Directional · verify currentJudgment that overrides the tables
Two rules that outrank any metric menu.
Proxy-not-goalGoodhart's law
Tell
Engagement is up — because users are lost and clicking around. The number rose while value fell.
Fix
Name how it could rise while value falls, and always ship a paired guardrail (what must NOT degrade).
Outcome-not-eventagentic / AI
Tell
"Messages sent" rewards a chatbot that loops. Event volume becomes vanity the moment the product acts for the user.
Fix
Count jobs completed — resolved-without-handoff, task success, time-to-resolution.
Aumayr · L.E.K. Key Value Drivers
The finance metrics most PMs skip — and the tree that ranks them.
Exec and finance weigh these even when PMs don't. Then, when a stakeholder asks "does this metric actually move the business?", decompose enterprise value into the operating drivers a team can move.
Share by volume AND value
AumayrA value-share gap vs volume share signals a discounting / mix problem.
Multi-level contribution margin
AumayrThe real "is this worth shipping?" number. Plus ROS / ROI / break-even point.
PLC age-structure
AumayrA portfolio skewed to mature / decline is hidden risk. Track satisfaction & relationship-quality as metrics, not survey afterthoughts.
L.E.K. value-driver tree · three steps
1MapDecompose enterprise value (NPV / operating profit) into drivers a team can move — price, volume, mix, cost, capital, retention.
2SensitivityPerturb each by the same small %; rank by sensitivity, not salience — keep the steep ones.
3ControllabilityDrop high-impact-but-uncontrollable factors (regulated price, commodity cost) — a risk to hedge, not a KPI to chase.
Two tests for a KPI
Significant value impact AND controllable. The trap it catches: teams reward managers for metrics that barely move value. This is the finance-rooted complement to the North Star's product input tree.
AI products · three-tier metric stack — house framing
Never report a single accuracy number.
Track three tiers, each with an owner. Quality is the leading indicator; experience is the point; unit economics is the constraint.
1 · Qualityleading indicator
Eval pass-rate on the golden set + failure-mode counts. Moves on every prompt / model swap.
2 · Experiencedid the job get done?
Task success, no-retry rate, escalation-to-human rate, trust / adoption.
3 · Unit economicsthe envelope
Cost per successful task = (tokens + infra) ÷ resolved jobs, plus latency and margin. Delight that loses money per call doesn't ship.
Quality → Experience → Unit economics · each with an owner
The vanity rule
A quality-tier win that doesn't move the experience tier is vanity; an experience win that blows the cost envelope isn't a win. Read all three together or you'll optimize a number nobody feels.
Try it
Run your own growth accounting.
Two metrics instruments, live — the Quick Ratio gauge with its retention bands, and the Sean Ellis PMF meter against the 40% bar.
The Quick Ratio gauge needs JavaScript to run. Run it on the Tools page.
The PMF meter needs JavaScript to run. Run it on the Tools page.