This field record asks: Which simple rule built from early spend share and early CPA maximizes holdout lift on later spend retention and conversion-cost performance? It is part of a seven-study Darwin IQ Labs series on early signals for newly introduced Meta creatives (Nos. 113–119).

On holdout Meta accounts from clients sampled of various industries (2046 weekly new ads; ~30% holdout), a train-chosen rule—early share ≥ half the duplicate baseline and CPA ≤1.25× the ad set when available—lifts later spend retention 1.85× (95% CI 1.28–2.76) but spends-and-performs only 1.14× (CI 0.83–1.48), short of a 1.5× CPA claim.

Sample

Dataset EMR-2026-10G combines the weekly and daily new-ad panels underlying Nos. 115–117. Holdout is ~30% of accounts by hash. Clients sampled of various industries.

Method

Candidate screens use thresholds fixed from the duplicate-baseline study (No. 117) and early CPA from the CPC/CPA study (No. 116). Meta rankings and CPC/CTR stand-ins are omitted because prior records did not support them. The winning rule is chosen on training lift for spends-and-performs; only holdout lift is reported as the result. Observational late-add backtest — not a live container RCT.

Findings

Spend screen holds; CPA claim does not. Holdout supports the screen for spend, not yet for CPA: "Pass kill AND (CPA ≤ 1.25× or too few conversions yet)" → 1.85x on keeps-spending (CI 1.28x - 2.76x), but 'spends and performs' is only 1.14x (CI 0.83x - 1.48x).

Holdout: the screen lifts spend retention, not spends-and-performs

Lift vs random new ads on holdout accounts

1.50x barKeeps spending1.85xBeats ad-set CPA0.92xSpends and performs1.14x

Source: Darwin IQ Labs, dataset EMR-2026-10G · Success bar 1.5× on spends-and-performs

Pitch-safe claim from holdout: challengers that passed were ~1.8× as likely to keep spending; they were not reliably more likely to hit target CPA.

What this record does not claim

  • No account list is released with EMR-2026-10G.
  • This is a historical late-add backtest inside BAU ad sets, not a prospective container pilot.
  • Many ads never reach five conversions; performs analyses favor ads that already got delivery.