I also asked the question to Copilot, providing the slack question, growthbook html docs and SQL queries as context. The answer seems reasonable: Looking at your SQL queries, I can now explain the key difference between these two metrics: ## The Core Difference Ratio Metric (1): Only includes users who have at least one order in the denominator (5k users) Mean Metric (2): Includes all users exposed to the experiment in the denominator (83k users), even if they never placed an order ## How This Happens in the SQL ### Ratio Metric Query

sql
-- In __userMetricAgg:
SELECT
  umj.variation,
  umj.client_id,
  SUM(COALESCE(umj.m0_value, 0)) AS m0_value,
  COALESCE(MAX(umj.m0_denominator), 0) AS m0_denominator  -- Key line
FROM __userMetricJoin umj
GROUP BY umj.variation, umj.client_id

Notice

m0_denominator

in the fact table is set to

1

for each order row. When aggregating: • Users with orders get

MAX(m0_denominator) = 1

• Users without orders get

MAX(NULL) = NULL

, then

COALESCE(NULL, 0) = 0

In the final aggregation:

sql
SUM(COALESCE(m.m0_denominator, 0)) AS m0_denominator_sum

Only users with

m0_denominator = 1

contribute to the denominator. Users with zero orders are excluded. ### Mean Metric Query

sql
SELECT
  m.variation AS variation,
  COUNT(*) AS users,  -- This counts ALL exposed users
  ...

The denominator is simply

COUNT(*)

of all users in

__userMetricAgg

, which includes every user from

__distinctUsers

(all experiment exposures), regardless of whether they have orders. ## Why This Matters for Your A/A Test With 83k total exposed users but only 5k with orders: Ratio Metric: 1.7M / 5k ≈ $340 per purchasing user • Higher variance (smaller sample size) • More likely to show false positives in A/A test Mean Metric: 1.7M / 83k ≈ $20 per exposed user • Lower variance (larger sample size) • More stable, less likely to show false significance ## Which Should You Use? Use Ratio Metric when you want to understand: • “Among users who purchased, how much did they spend?” (Average Order Value style metrics) Use Mean Metric when you want to understand: • “What’s the average value per user exposed to the experiment?” (True per-user impact) For most A/B tests, Mean metrics are more appropriate because they measure the true causal impact per user exposed, including those who didn’t convert (which is often the effect you’re trying to change).