Conversion optimization

Conversion optimization is efficient when all converting users are roughly equal, or when value is unknown at bid time. It breaks down when value varies widely: one-and-done buyers vs repeat customers, trialists who churn vs subscribers who renew, or purchasers who refund vs those who expand.

On the platform, conversion optimization learns fast because the event is simple: no value parameter, fewer eligibility gates, often lower volume requirements than value goals. That speed is also the risk. Teams can scale spend on cheap converters who look identical to high-LTV users in the ad account because value was never sent.

On the signal layer, the opposite failure mode is sending every purchase to the network. Platforms treat each event as positive reinforcement. A converter predicted to return, churn early, or abuse promos still teaches the algorithm to find more like them unless you optimize which conversions to send. Event incrementality here means withholding low predicted-value events, not only changing bid strategy.

Platform conversion optimization remains the right control arm in experiments. Churney pilots typically hold business as usual (BAU) conversion (send-all or platform conversion maximization) against selective, value-threshold sends to measure true incrementality.

Platform goals maximize conversion count. pLTV programs optimize which conversions the network sees and how many events cross the wire:

1. Model: Score user-level pLTV from first-party data in your data warehouse at the anchor event (purchase, trial start, install).
2. Threshold vs volume: Set a predicted-value floor and a minimum signal volume target. Too low a threshold reinforces returners and one-and-done buyers; too high starves learning phase.
3. Selective send: If predicted value is below threshold, Churney may not send the conversion event even though the transaction occurred. The network stops learning from customers you expect to destroy margin.
4. Activation: Above-threshold converters get the conversion event sent directly to ad networks (Meta Conversions API, Google Ads Conversion API, app paths). pLTV gates whether to send; campaigns still optimize on conversion count unless you add value-based bidding as a separate step.
5. Readout: Compare incremental conversions, incremental ROAS, and cohort quality vs BAU send-all or platform conversion maximization after cohort maturity.

Pair selective sends with platform conversion-maximization campaigns first. Graduate to value-based bidding or value optimization only when send quality and volume are stable.

1. Staying on conversion optimization when value variance is high. Scales volume of low-quality converters.
2. Using conversion optimization as the only readout metric after enabling pLTV elsewhere. Platform CPA can improve while profit does not.
3. Choosing weak proxy events. Short-window events that do not correlate with LTV teach the wrong lesson even before value bidding.
4. No holdout or BAU cell. Cannot isolate incrementality of a new value signal.
5. Assuming value bidding is always better. Low volume, uncalibrated pLTV, or poor match can underperform a stable conversion campaign.
6. Sending every conversion regardless of predicted value. Reinforces returners, promo abusers, and one-and-done buyers the model already flagged.

• Signal optimization
• Value optimization
• Value-based bidding
• Target ROAS (tROAS)
• Business as usual (BAU) conversion
• Signal volume
• Predicted lifetime value (pLTV)
• Holdout test
• Incremental ROAS