pLTV (predicted lifetime value) is a class of machine-learning models that estimate how much a user will eventually spend — not from observed multi-year revenue, but from early-window signal. The model scores a user within hours or days of install based on what they did in onboarding, how deep their first session went, whether they completed the tutorial, whether they made a first purchase, what event sequences they generated. The output is a forecasted LTV per user, used for bid optimization.
Why pLTV matters now: a decade ago, UA optimization was install-count or cost-per-install. The mature version was ROAS — target return on ad spend at D7 or D30. pLTV is the next evolution: instead of waiting for D30 observed revenue, you can bid into estimated D365 revenue starting from D1. Meta and TikTok's app campaigns now consume pLTV signals as their optimization event directly — when you send their network a "this user scored 85th-percentile pLTV" event, the network will bid up similar lookalike users.
What signals drive pLTV models in practice
- Onboarding depth: did the user complete onboarding? How long did it take? Did they personalize their experience?
- Session-1 engagement: time-in-app, screens visited, content interactions, feature toggles activated.
- First-purchase signal: any IAP / subscription start within the first session is the single strongest predictor. Even a low-price IAP changes the user's pLTV cohort dramatically.
- Event sequences: not just "did the user do X", but "did the user do X then Y then Z in this order".
- Device + acquisition channel + country + creative cluster: macro-features baked in alongside per-user signal.
Build threshold: pLTV is not a quick win. You need (a) at least 3-6 months of historical cohort data to fit a model with predictive power, (b) a feature pipeline that can score users in near-real-time (a few hours of latency is usually acceptable; days is not), (c) an event-export pipeline to fire pLTV scores back into Meta / TikTok / AppLovin as conversion events, (d) a feedback loop that retrains the model regularly as user behaviour drifts and the product changes.
Vendors abstract this for mid-market. Apps that don't want to build pLTV in-house can use vendors like AppsFlyer Predict, Adjust Predictive Analytics, Tenjin pLTV, Liftoff Vungle's audience tools. The trade-off is the vendor's model is less customized to your app's specific monetization shape — but they handle the data pipeline, model training, and network event-export plumbing for you.
When pLTV doesn't work: apps with monetization that ramps slowly (LTV materializes mostly in months 6-12), apps with very low daily activity (insufficient early signal to score on), and apps with monetization shaped by external events (seasonal F2P games where in-game events drive revenue). pLTV is most powerful when first-week behaviour strongly predicts year-1 revenue — common in subscription, less common in deep F2P meta loops.
pLTV vendors vs in-house build
| Approach | Setup time | Customization | Best for |
|---|---|---|---|
| AppsFlyer Predict | Days to integrate | Limited — vendor model | Mid-market apps already on AppsFlyer |
| Adjust Predictive Analytics | Days to integrate | Limited | Adjust customers |
| Tenjin pLTV | Days | Limited | Indie + smaller publishers |
| Liftoff audience tools | Days | Limited | Apps spending on Liftoff DSP |
| In-house build | 3-6 months | Full — tailored to your monetization | Large publishers ($500K+/mo UA) with ML capacity |
Vendor pLTV works for most apps that don't have unusual monetization shapes. In-house pLTV is justified when (a) you have 3-6 months of cohort data and ML engineering capacity, (b) your monetization shape is atypical, (c) UA spend is high enough that 5-10% improvement justifies the engineering investment.