Modelled conversions: the numbers nobody audits

A growing share of the conversions in your ad platform dashboards did not happen, in the sense that no observed event links a specific install or purchase to a specific ad. The platform inferred them. It saw a gap between what it could observe and what it believed it delivered, and a model filled the gap. This is not fraud, and it is not a secret. What makes it a problem is that the party that gets paid on its output also builds and trains the model, then scores it, and almost no advertiser audits any of that.

The strong version of the claim: modelled conversions are now the single largest unreviewed input into mobile UA budget decisions, larger than any creative or targeting choice, because they flow directly into both the reports that justify spend and the bidding systems that place it.

Why modelling exists and why it is not going away

Since ATT, individual-level attribution on iOS covers only consenting users. Consent rates, as the MMPs track them, sit well under half of installs and barely move year to year. SKAN and AdAttributionKit postbacks cover the rest; they arrive late, though, carrying coarse conversion values, and below privacy thresholds they disappear entirely. An earlier UA Ledger piece (SKAN 4 and AdAttributionKit, a year in) covered how sparse the returned signal remains.

A platform optimising bids in real time cannot wait for a postback that may arrive in days and may not arrive at all. So it models. It trains on the consenting minority, on aggregate postbacks, on its own click and impression logs, and it estimates how many conversions the unobserved majority produced. The estimate is then reported to the advertiser as a conversion count and used internally as the feedback signal for automated bidding.

The mechanism is entirely reasonable as an engineering response. The incentive problem is that the platform's revenue rises with the advertiser's perceived ROAS, the perceived ROAS depends on the modelled count, and a system the platform controls, and the advertiser cannot inspect, produces the modelled count. Nobody needs to act in bad faith for the model's errors to drift in one direction. It is enough that optimistic errors are not punished and pessimistic errors are.

The second-order effect: models become spend

The consequence most coverage misses isn't inflated reports. It's that the inflation gets spent.

Automated bidding systems, including the AI-driven models that now dominate the large networks, treat the modelled conversion as the training label. If the model believes a segment converts well, the bidder buys more of that segment, generates more modelled conversions in it, and reinforces the belief. The advertiser's actual revenue does not participate in this loop unless the advertiser forces it in. Unity's Vector and AppLovin's Axon operate on the same principle, as do the large social platforms. The more spend that runs through self-optimising systems, the more of the budget a model's belief about its own performance ends up allocating.

There is a further effect on measurement itself. Compare two channels on platform-reported ROAS and you are increasingly comparing two models with different assumptions, not two outcomes. A channel with an aggressive model wins the comparison and the budget. The advertiser hasn't compared channels at all; it has compared modelling philosophies.

An audit framework that does not require the platform's cooperation

The platform won't open the model, and it doesn't need to. Four checks, run quarterly, are enough to know whether the modelled numbers are drifting from reality and by roughly how much.

• Modelled share. Ask each platform what proportion of reported conversions in your account are modelled rather than observed. Some will answer, some will not. Where they will not, the refusal itself is an input: the channel goes into a lower-confidence tier.
• Reconciliation to first-party revenue. Sum every platform's reported revenue for a quarter, plus your organic estimate, and compare it to the revenue your own backend booked. The excess is the aggregate over-claim. Track it as a single number over time. Rising is the signal that matters more than the level.
• Reconciliation to the MMP. Compare platform-reported conversions to MMP-attributed conversions per channel per month. The ratio will not be one, and it does not need to be. What it needs to be is stable. A channel whose ratio climbs quarter over quarter is a channel whose model is growing more confident than the evidence it is fed.
• Holdout calibration. Whenever a geo holdout runs on a channel, record the ratio of the platform's modelled conversions in the test geos to the incremental conversions the holdout measured. This is the only external calibration the model will ever receive, and it should be stored with a date and reused as a cap.

The cap is the decision rule. Budget models accept platform-reported uplift for a channel only up to the most recent holdout calibration ratio. Anything above it goes in the record and not in the budget. Until a channel has a calibration, its reported conversions are haircut to the portfolio's aggregate over-claim from the revenue reconciliation.

An illustrative example

A hypothetical studio sums platform-reported IAP revenue across four channels for a quarter and finds it exceeds the backend's total booked revenue, before organic is even considered, by an illustrative 30%. That is not a problem to reconcile away by adjusting windows. It is the size of the modelling optimism across the plan, and it becomes the default haircut.

Two of the four channels have holdout calibrations from the past six months, showing modelled-to-incremental ratios of roughly 1.2 and 1.8 in this illustration. Those ratios cap their reported conversions. The other two receive the 30% aggregate haircut and join the test queue, with the one whose MMP ratio has climbed fastest going first.

The resulting budget model flatters no channel, and it flatters the largest, most automated channels least. That is the expected outcome, because those are the channels where the model has the most room to be generous.

What to ask in the next platform review

The question that separates a vendor with a defensible model from one with a convenient one is simple to ask and hard to dodge: what would make your modelled conversions go down? A platform whose team can describe the conditions under which its model reports fewer conversions than the advertiser expects is running a model that is at least trying to be right. A platform whose answer is a restatement of how the model finds more value is describing a model nobody has ever asked to be wrong, and the advertiser is the only party with any reason to ask.

Related archive reading

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

• Attribution is an allocation rule, not a truth
• Retention is a UA metric