Your bids are buying your cheapest customers

Sit in your buyer's chair; this is happening in her company right now. The growth lead of a subscription app is presenting a quarter of channel performance. Cost per trial signup, down 18 percent. Trials, up. The ad platforms' dashboards glow with optimization wins. Then finance shows the other chart: cohort revenue, sliding three straight quarters. Two dashboards, opposite stories, and the meeting ends with everyone agreeing to "look into attribution."

Attribution is not the problem. The bidding objective is, and the repair has a name: predicted lifetime value, fed to the platforms as the signal to optimize. Every major ad platform's algorithm chases the conversion event you feed it, and this company, like most, feeds it the cheapest one to measure: the trial signup. The platform's machine learning then does exactly its job, finding the users most likely to sign up for a trial at the lowest cost. But "likely to sign up" and "likely to be worth anything" are different populations. The cheap-to-acquire cluster skews toward deal-hunters, tire-kickers, and users whose trials die at the paywall. Quarter by quarter, the algorithm gets better at buying them, which is why efficiency improved while cohort value fell. The system was aimed at the wrong target and hit it.

An ad platform optimizes toward whatever you call a conversion. Call a 50-cent user and a 500-dollar user by the same name, and it will buy you the cheap one every time.

The platforms themselves quietly agree; that is why value-based bidding products exist across Google and Meta. Feed the algorithm revenue instead of signups and it optimizes for revenue. The catch, and the reason most teams stall here, is that subscription revenue arrives over months and years, while bidding signals must arrive within days to be usable. The platform needs to know what a user is worth long before the user has decided.

Sending trial-to-paid conversions helps and is still late. The signal that actually changes the economics is a prediction: this specific new user, based on their first hours of behavior, will be worth roughly this much over their lifetime. Which turns the bidding problem into a modeling problem.

The method is predicted lifetime value, pLTV, engineered as a bidding input. A model trained on your own historical cohorts learns the mapping from early signals, acquisition context, onboarding progression, first-session depth, early feature usage, plan viewed, to eventual revenue. The statistics underneath are unforgiving in specific ways: revenue outcomes are censored, because most of every recent cohort's lifetime has not happened yet, so the model family comes from survival analysis and censoring-aware regression, not from a naive revenue fit that trains only on old cohorts and drifts with every product change. Calibration matters more than ranking, because the dollar values feed an auction; a model that is directionally right but 2x optimistic buys you the wrong bids at scale.

The predictions flow back through the conversions API within the platforms' signal windows, and the auction begins discriminating: 80 dollars for the user whose first day looks like your best cohorts, 2 for the lookalike of the paywall bouncers. The budget stops being spent on volume and starts being spent on the integral under the value curve. We built the censoring-aware scaffolding for our own usage-revenue models, and the recurring lesson was unglamorous: the feature pipeline's freshness and the calibration checks decide whether the platform trusts your signal, long before model sophistication matters.

(There is an honest tension to hold: pLTV bidding narrows targeting toward users who resemble past winners, so it needs an exploration allowance or it slowly forgets everyone it never tried. That is a budget line, not a footnote.)

The version of that quarterly meeting we like better runs on one chart instead of two: cost per predicted dollar of LTV, by channel. In published case studies of value-based bidding done with properly calibrated models, the pattern repeats: signup costs rise, sometimes 20 percent, and revenue per acquisition cohort rises far more, because the algorithm was always this good and is finally pointed at the right number. The growth lead and the CFO are, for the first time in three quarters, reading the same dashboard.

TensorLabs builds pLTV pipelines for subscription businesses whose acquisition efficiency keeps improving while cohort revenue disagrees. If that divergence is on your charts, get in touch and tell us which event your bids currently optimize toward. The answer usually is the diagnosis.