At ZyG, Lifetime Value (LTV) is the unit of truth that changes how you think about everything from creative spend to retention. While last quarter's numbers tell you what happened yesterday, LTV fuels the agentic decision-making process and tells you what each customer, cohort, and brand is likely to be worth tomorrow.

The asymmetry that makes LTV so important in the first place: the gap between the knowability of expenses vs. revenues. Future spend on ads, inventory, shipping, and platform fees are mostly under our control and can be modeled. Expected revenue is the opposite. A single order today is just the first sample of a distribution that unfolds over months: will this customer come back, how often, at what cart size, and when are they likely to churn? Modeling LTV is how we turn the hard side of the equation into something we can actually plan against.

We model LTV at three levels of granularity, each answering a different question:

• Brand level: Is this brand likely to be profitable over the long run, net of acquisition cost? LTV here is a strategic input - it  feeds the ZyG Score, the result of our agentic PMF test, and drives decisions on which brands to scale and how to deploy capital.

• Cohort level: How does one acquisition cohort behave versus another over a comparable time horizon? Our LTV analysis quantifies the effect of a creative shift, a pricing change, or a new channel, and lets us measure, with statistical significance, whether the change actually moved retention and repeat-rate curves, not just first-order conversion.

• Customer level: What is each customer worth, given their journey so far? This powers personalization, win-back targeting, and bid optimization in near real-time, impacting UA decisions. Combined with creative attribution, it lets us judge a creative by the LTV of the customers it acquired, not just the conversions it produced.

On the data science side, predicting LTV is a layered problem. Repeat-purchase frequency and churn dynamics come from BG/NBD or shifted-beta-geometric (sBG); basket value from Gamma-Gamma; and gradient boosting on top of a rich feature store that captures the per-customer signal that probabilistic models can't. 

Three principles keep the LTV model usable across ZyG:

• Calibration: Predicted LTV must match realized LTV at the cohort level

  within a known confidence interval. Without calibration, statistical

  significance claims collapse, allowing you to detect a shift but not to size it.
  
  

• Explainability: Each prediction surfaces the customer behaviors that

  drove it. At the customer level this powers UA bid decisions; at the

  brand level the aggregated drivers feed the ZyG Score's diagnosis of

  why a brand is expected to be profitable, going beyond a binary go/no go signal.
  
  

• Freshness: Models retrain as new orders, sessions, and ad events land,

  so feature and population distributions don't drift and stay aligned with current data.

The outcome: a mechanism that produces a single metric - the core signal in ZyG's agentic Operating System, powering every agent from UA to retention. Predicted well, it's the golden key to scale.