- Validate a “winning segment” for breadth before building plans on it – reach estimates plus reconciliation against backend LTV, not just early CPA; scale in increments with a pixel-overfit tripwire.
- The reset test: a fresh concept that opens at a normal CPM with frequency building normally means fatigue; no reset means the pool the pixel will serve is exhausted, most often pixel overfit. A failed reset is stop-and-diagnose, confirmed on a clean pixel before the creative is written off.
- Pixel overfit – campaigns optimizing to a shallow or narrow pixel event that correlates with short-term conversion but not with the true addressable, revenue-generating audience, so reported performance looks strong precisely because the model has fit itself to a pocket too small to scale.
- Don't let one early pocket train the pixel: spread early learning budget across concepts, never set unit-economics benchmarks from a single segment, and check event–LTV correlation before choosing an optimization event.
- Creative freshness is checked by hand, weekly: ad names encode the concept, ad-level spend is aggregated by concept, and two numbers are read – each concept's share of budget and its CPA. Spend concentrating on one concept is the early sign of overfitting.
The part that generalizes beyond this one case
- Validate a "winning segment" for breadth before building plans on it – reach estimates plus reconciliation against backend LTV, not just early CPA.
- Scale in increments with a pixel-overfit tripwire. The reset test: a fresh concept that opens at a normal CPM with frequency building normally = fatigue; no reset = the pool the pixel will serve is exhausted, most often pixel overfit. Treat a failed reset as stop-and-diagnose – and confirm on a clean pixel before writing off the creative.
- Don't let one early pocket train the pixel. Concentrated early spend on a single high-intent audience teaches the optimizer to keep returning to it, starving new creative of impressions outside it. Spread early learning budget across concepts, and periodically validate breadth on a clean pixel or via reach estimates.
- Never set unit-economics benchmarks from a single segment's early performance.
- Run a pixel-vs-backend reconciliation cadence, and check event–LTV correlation before choosing an optimization event – so the model optimizes toward revenue, not toward a shallow proxy.
Current Processes & Policies
These are the areas we maintain to keep CPMs in check.
Technical
On the technical side, we control the following – each item is checked by Ops pre-launch or on a weekly cadence (details intentionally omitted):
- Conversions API
- Advanced matching quality
- Required funnel events before purchase
- Domain verification
- Event deduplication
- Aggregated Event Measurement – standard events only
- Landing page speed
Creative freshness – weekly manual review
Meta doesn't give us a creative-freshness tool, so we run the check ourselves, by hand, once a week. Two things make it possible:
1. Ad naming convention. Every ad name and creative file name encodes the creative concept and its subtype – not just the name or number of the individual creative. Without this, spend can't be aggregated at the concept level at all.
2. Weekly spend-by-concept review. Once a week we export ad-level spend, aggregate it by concept, and look at two numbers: the share of budget each concept takes, and the CPA it delivers. If spend starts concentrating on a single creative concept, that's an early sign of overfitting – the signal to rotate fresh concepts in before CPMs react.