Value optimisation selects for what you can measure

A value-optimised campaign is a machine for finding people who resemble the labels you gave it. It is not a machine for finding valuable players. The two overlap, sometimes closely, but the gap between them is the single largest source of disappointment in ROAS bidding, and it widens exactly where measurement is weakest. On iOS under SKAN and AdAttributionKit, where the label is coarse and time-boxed, a campaign told to maximise value will maximise visible early value, and the players it selects will look excellent for a fortnight and ordinary by day ninety.

This is not a complaint about the platforms. It is a statement about how supervised learning works. The model predicts the label. If the label is a truncated, delayed, bucketed proxy for revenue, the model becomes very good at predicting that proxy and indifferent to everything the proxy misses.

What the label actually contains

Take the mechanism seriously. A value-optimised campaign trains on postbacks that say, for each attributed install, how much value the advertiser reported within some window. On Android with an MMP, that window can be long and the value fine-grained. On iOS, the conversion value is a small number of bits, the windows don't move, and postbacks arrive late and with privacy thresholds that null out low-volume cells entirely.

The label therefore contains: revenue that happened early, from players who converted within the window, in campaigns large enough to clear the threshold, encoded into a handful of buckets. It excludes: revenue from slow-burn players, subscription renewals past the window, ad revenue unless the advertiser has taken the trouble to send it, and everything from small campaigns whose cells the threshold suppressed.

The model does not know these are gaps. It sees the players who lit up the label and learns what they have in common. It then buys more of them.

The feedback loop nobody budgets for

The second-order effect is that the cohort you buy becomes next month's training data. If the label favours early payers, the campaign buys early payers, the postbacks confirm it found early payers, and the model's confidence in that pattern strengthens. Players whose value arrives late show up thin in the cohort, so they show up thin in the labels, so the model learns to avoid them further. Measurement gaps become selection gaps, and selection gaps become measurement gaps.

Sensor Tower's State of Mobile report earlier this month described the industry shifting from new-user volume towards lifetime-value expansion. Studios are running that shift through bidding systems whose definition of value is whatever fits in a conversion window. A studio whose actual economics rest on month-three retention or on ad revenue from non-payers is, in effect, asking a system trained on a different game to find its players.

There is a related trade-off in the choice between value and event optimisation. Event optimisation for a purchase asks the model to find anyone who buys; value optimisation asks it to find people who buy a lot, early, which is a more specific request and therefore far more sensitive to label quality. On a platform with rich labels it wins. On a platform with poor labels it can lose to the simpler objective, because the simpler objective does not amplify the truncation as hard.

A label audit before the bidding change

Before switching a campaign to value optimisation, the useful exercise is to measure how much of the revenue you care about is actually visible to the label. Call it label coverage: the share of day-ninety revenue per install that reaches the platform inside its window, in its encoding, for a representative cohort from your own data.

An illustrative case. A puzzle studio finds that its day-ninety revenue per install is 3.00 on a blended basis. Of that, 0.90 arrives within the first conversion window on iOS, and after bucketing into the available conversion-value ranges and applying the privacy threshold to its campaign sizes, roughly 0.60 is legible to the platform. Label coverage is twenty percent. A value-optimised campaign on that game is optimising for a fifth of the economics and is blind to the rest. On Android with a long MMP window and fine-grained revenue events, the same audit might show sixty or seventy percent coverage, and value optimisation is a reasonable bet.

A decision rule that follows:

• Coverage above half: value optimisation is likely to select for the players you want. Run it and check cohort quality at day thirty against the event-optimised baseline.
• Coverage between a quarter and a half: run value optimisation only alongside a predicted-value signal.
• Send a modelled LTV event at the point in the window where your prediction stabilises, so the label reflects what you expect the player to be worth rather than what they have paid so far.
• Coverage below a quarter: optimise for a well-chosen event instead, and do the value selection yourself through cohort review and budget allocation between campaigns. The model cannot see enough to help.

Making the label say what you mean

The lever most teams underuse is the label itself. Platforms and MMPs let the advertiser define what counts as value. Sending raw early revenue is the default and the worst option for most games. Sending a predicted lifetime value, computed from the first day or two of behaviour, gives the model a label that correlates with the outcome you actually care about, and it does so inside the window the platform imposes.

This is not free. A predicted-value label is only as good as the model behind it, and a bad prediction teaches the campaign to find the players your model overrates. It also requires that the studio can distinguish, after the fact, between players the prediction got right and players it got wrong, and feed that back. The discipline is the same as any measurement programme: the label is a hypothesis about value, and you need to test it against realised value on a fixed cadence.

The campaign will do exactly what the label tells it to do. Most teams spend their optimisation effort on bids and creative and leave the label at its default. The label is where the selection happens, and it is the one input the platform will let you change without charging you for it.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• Every SDK you install is a data-sharing agreement
• Acquiring high-value cohorts without buying whales