Why it matters
Finance and leadership think in cohorts: "What did Q1 paid social buyers yield by month six?" Cohort models answer that question well. They smooth noise across users and align with cohort LTV dashboards.
Ad platforms, however, bid per impression and per user. Sending one average value for everyone in a campaign removes the gradient that separates future VIPs from one-and-done buyers. Two-model thinking is common: cohort models for planning, user-level models for activation.
Cohort-based LTV model
| Layer | Grain | Job |
|-------|-------|-----|
| Cohort-based LTV model | Segment / cohort | Reporting, targets, sanity checks |
| User-level pLTV | Per user | Value on each conversion event for bidding |
Churney activates the second layer. Cohort models can inform priors and calibration, but aggregation error grows when segments hide heterogeneity inside the group.
Category variants
| Vertical | Typical cohort slice | Limitation for bidding |
|---|---|---|
| Ecommerce | Monthly acquisition cohort | Hides bracketing vs loyalists in same month |
| Subscription | Plan or channel cohort | Trial quality varies within cohort |
| Mobile app | Geo or network cohort | ATT bias by platform |
Common mistakes
- Using cohort LTV as the bid value for every user. Destroys within-campaign ranking.
- Comparing cohort model output to platform ROAS directly. Different windows and attribution.
- Ignoring maturity. Early cohort curves are unstable.
- Assuming one model serves finance and UA. Different grain, different decisions.
- Confusing with customer LTV reporting. Reporting is backward; models may be forward-looking.
Advertiser lens
| Role | Cares about |
|---|---|
| Finance | Cohort curves, payback, forecast accuracy |
| UA / performance | Needs per-user scores, not only cohort averages |
| Analytics | When cohort and user models disagree |
| Data science | Feature sets and horizon per model type |
FAQ
What is a cohort-based LTV model?
A model that predicts or summarizes LTV for defined customer groups rather than individuals.
Why are two models often required?
Cohorts support planning; user-level scores support real-time bidding differentiation.
Can cohort models feed user-level models?
Yes, as priors or calibration benchmarks, not as a replacement for per-user output.
What is aggregation inaccuracy?
Error introduced when heterogeneous users are averaged into one segment value.
Which horizon should cohort models use?
Match the business decision (D90, D180, full LTV) and cohort maturity available.
Is cohort LTV the same as a cohort-based model?
Cohort LTV is often the reported metric; the model is the method that produces or forecasts it.
Not the same as
| Term | Difference |
|---|---|
| User-level pLTV | Per-user activation score |
| LTV reporting | Historical dashboards, not necessarily a predictive model |
| Media mix modeling (MMM) | Channel-level macro model, not customer grain |
