Why it matters
A pLTV model can be accurate in the data warehouse and still fail in production. Broken pipelines, stale scores, match loss, or erratic value magnitudes all show up as "the algorithm won't learn" or wild CPA swings. UA teams often blame creative or budget when the real issue is unhealthy inputs.
Signal health is an operating metric, not a one-time launch checklist. Privacy changes, catalog updates, promo calendars, and acquired mix shifts all stress the stack. Without monitoring, teams scale spend on platform ROAS while signal quality silently degrades.
Treating health as a single dashboard (volume + match + calibration + drift) prevents finger-pointing between engineering, data science, and performance marketing.
Signal health
Churney-style activation assumes ongoing health checks across the loop:
- Source: First-party data and revenue definitions in the data warehouse stay data-ready.
- Model: User-level pLTV scores refresh on schedule; watch model drift and feedback loop effects on features.
- Transform: Signal transformation rules stay bounded; calibration panels updated after mix shifts.
- Deliver: Meta CAPI, Google Ads Conversion API, and app paths monitored for signal freshness, errors, and conversion signal loss.
- Prove: Holdout tests and incremental ROAS vs BAU confirm health improvements move business outcomes.
Signal optimization is the remediation loop when health scores slip: adjust timing, volume, caps, or event definitions until learning stabilizes.
Category variants
| Vertical | Health watchpoint | Early warning |
|---|---|---|
| Ecommerce | Refund lag vs sent value | High-score deciles with rising returns |
| Subscription | Trial volume spikes | Calibration drift after promo trials |
| Mobile app | ATT / SKAN degradation | Match drop on iOS alongside flat values |
Common mistakes
- Monitoring only event count. , not value distribution or match rate.
- Declaring victory at launch. without 30 to 90 day health reviews.
- Ignoring feedback loops. that change who you acquire and retrain models.
- Scaling spend. while signal volume is below platform learning thresholds.
Advertiser lens
| Role | Cares about |
|---|---|
| UA / performance | Stable CPA/ROAS and exit from learning phase |
| Engineering | Pipeline SLAs, API errors, identifier coverage |
| Data science | Drift, calibration, score freshness |
| Growth analytics | Holdout readouts tying health to incrementality |
FAQ
What belongs in a signal health review?
Volume, match/delivery errors, value distribution, calibration vs realized LTV, freshness, and drift indicators vs last stable period.
How is signal health different from signal quality?
Health is the umbrella operational status; quality is a core ingredient (accuracy, completeness, timeliness).
How often should teams review health?
At least monthly for live pLTV programs; weekly during pilots or major promo/catalog changes.
Can platforms report signal health?
Partially (diagnostics, match scores like EMQ). Business-side calibration and cohort readouts are still required.
What is the first fix when health drops?
Diagnose delivery vs model vs transformation: lost events, stale scores, or miscalibrated values each have different owners.
Does Churney monitor signal health?
Product focus includes data warehouse-to-platform delivery, calibration discipline, and readouts tied to holdout tests and incremental ROAS.
Not the same as
| Term | Difference |
|---|---|
| Data readiness | Prerequisite setup, not live operational quality |
| Platform learning phase | Campaign state inside the ads UI, not end-to-end signal stack |
| Attribution health | Click/view matching for reporting, broader than value signal delivery |
