a builder's codex
codex · operators · Archie Abrams · ins_long-term-holdouts-30-40-evaporate

30–40% of growth experiments with short-term lift show no incremental value at one year

By Archie Abrams · VP Product & Growth, Shopify · 2026-04-28 · podcast · Churn optimization & long-term holdout experiments

Tier A · TL;DR
30–40% of growth experiments with short-term lift show no incremental value at one year

Claim

Run a permanent 5% holdout on every shipped change (or a 50/50 split on new-user cohorts) and revisit one to three years later. Roughly 30–40% of changes that produced a clear short-term lift will show no incremental long-term value. Most growth measurement misses this.

Mechanism

Short-term lifts are often pull-forward effects — accelerating signups that would have happened anyway — or low-quality user attraction that does not durable retain. Long-term holdouts catch this by separating "moved the metric" from "created incremental value." Without the holdout, the team takes credit for evaporated wins and re-invests behind theatre.

Conditions

Holds when:

Fails when:

Evidence

"We want to lower barriers to get started and help folks grow, and those winners make the whole thing work."

Finding: 30–40% of experiments showing short-term lift (paying signup, first sale) show NO incremental GMV lift after 1 year.

Archie has never seen short-term positive flip to long-term negative. So safe to ship — but don't take credit.

Operating discipline at Shopify: 5% always-on holdout, automated email reports at 3 / 6 / 9 / 12 months. "You can't hide from reality."

— Archie Abrams on Lenny's Podcast, 2026-04-28

Signals

Counter-evidence

For very early-stage products, you don't have the volume or runway to run multi-year holdouts; short-term metrics are all you have. And the 30–40% number is Shopify's, not a universal constant — sample-size and category dependent. Some operators argue holdouts at scale leak revenue Shopify can afford to leak; smaller orgs cannot.

Cross-references

Open the interactive view → View original source → Markdown source →