In a 3-sided digital marketplace, we have these stakeholders-

  1. Consumers (driving engagement & capturing value)
  2. Creators (suppliers of business inventory)
  3. Advertisers (Monetization partners)

In a healthy ecosystem, these stakeholders engage in a seamless cycle: creators address consumer demand; consumers generate behavioral footprints; and advertisers bid on user attention to attract cross-platform sales.

The consumer floats at the centre — relaxed, capturing value. Everything else in the ecosystem exists to reach this person.

Acquisition & GrowthAcquiring new users and expanding product adoption
new sign-upsfree-to-paid conversionsD7/D30
EngagementMeasuring active user involvement and depth
DAU/MAUsessions/usertime spent
RevenueBusiness value generated via customer spend and ad performance
total revenueARPULTV:CAC
HappinessUser sentiment, satisfaction, and perceived ease of use
NPSapp rating
Task successFriction-less completion of high-intent user actions
time-to-discoverycrash rates%workflow completion

Different metric design types

Metric type
Example
Pros
Cons
Total counters
Total searches
Simplistic
Outliers can inflate total volume
Ratio
CTR = clicks ÷ impressions;
Frequently used to measure funnel progress
If both-numerator & denominator increase, ratio can still decrease, disagreeing with north star improvement. Also, Simpson's paradox across segments
Cumulative
Crash-free users since exposure; total revenue since first exposure
Can detect strong initial signals and regressions (like app crashes) very quickly
Stats Power can fall over time when feature is struggling with novelty; and only variance accumulates
Windowed
Returned between day 1 and day 7 post-exposure;
Stable metric; comparable across different experiments regardless of time
Late-exposed users are unusable diluting effect
Cumulative windowed
"D7 revenue so far" for a cohort exposed 3 days ago
Best for live dashboards during a running experiment
-
User-level aggregated
# active sessions per user
Best choice for primary OEC
Less sensitive than raw counts when the effect is intensity-based
Surrogate
12-month LTV predicted from first-30-day behaviour
Mapping long-horizon goals (LTV, 6-12m retention)
Prediction error inflates Type-I unless adjusted