The steering change comes first. Acquisition run on CAC alone optimises for cheapness — and cheapness systematically recruits deal-hunters whose cohort curves sag. Run the same channels on CLV:CAC by source and the map redraws: the influencer channel with ₹800 CAC and ₹1,200 predicted value is quietly worse than the content channel at ₹1,500 CAC and ₹6,000 value. The operational endgame is value-based bidding — feeding predicted customer value, rather than conversion counts, back into the auction platforms, so the machine hunts your best customers instead of your most numerous ones. The catch is that this only works if you trust your prediction enough to pay more per head, which is why prediction discipline is the spine of this page.
Prediction means acting at day 30 instead of regretting at year two. The Intermediate models (retention arithmetic, the probabilistic 'is she resting?' machinery) valued customers. The optimization layer forecasts them early, from behavioural signals with real predictive teeth: the interval to second purchase (in most retail data, the single loudest early signal — a customer who returns in three weeks and one who returns in three months are different assets wearing the same 'acquired' label), first-basket category and margin, channel of arrival, engagement depth. The forecast doesn't need decimal-point accuracy. It needs to sort — top-quintile pCLV treated differently from bottom-quintile is where the money is, and misclassifying the middle is affordable.
Intervention economics ranks the levers instead of loving them equally. Take the NPV sensitivity of each. A two-point lift in early repeat rate (nudging the second purchase forward) often outvalues a much larger win-back programme, because it compounds across every subsequent year. Margin mix (steering customers toward higher-contribution categories) raises the m in every future cell. Churn saves earn only when they're aimed: the save target is high value × high risk × high persuadability, because saving the low-value costs more than it returns, and 'saving' the never-leaving is just a discount to the loyal (the uplift logic — spend on those whose behaviour the spend actually changes). Reactivation has a half-life measured in months, so priced honestly it competes with acquisition, not with retention. The discipline: one table, levers down the side, ΔNPV per rupee in the column, budget follows the column.
Summed across the base, predicted CLV becomes customer equity — and the reporting altitude changes. Growth decomposes into equity added (new cohorts), equity defended (retention curves held), and equity expanded (frequency and margin lifted) — a decomposition that catches what revenue hides: a company can grow revenue while its equity shrinks, buying weak cohorts faster than strong ones decay. Two governance rules keep the machine honest. Backtest quarterly: last year's day-30 predictions against this year's realised curves. A model that flatters is a spend-justification engine wearing a lab coat (predicted CLV fed optimistic retention is how unit economics catch fire — Goodhart's law, invoiced). And keep the discount rate and margin definitions locked to the Intermediate page's standards, because optimization inherits every lie its inputs tell.