Attribution is an allocation rule, not a truth

Every attribution debate in a UA team eventually collapses into the same argument: which model is closest to the truth. Last touch, first touch, data-driven, SKAN postbacks, probabilistic matching, modelled conversions. The framing is wrong. None of these measure what caused an install; they're rules for deciding who gets paid, and a team that treats them as measurements will make worse spending decisions than one that treats them as what they are.

The claim is stronger than it sounds. It means that arguing about whether last touch is "accurate" is a category error, in the same way that arguing about whether a tax code is accurate is a category error. A rule can be fair or unfair; it can be easy or hard to game; it can be useful or useless.

It cannot be true.

What an attribution rule actually does

An install has a causal history that nobody observes. A player saw a creator video on Tuesday, scrolled past a playable on Thursday, searched the store on Saturday and tapped a rewarded placement on Sunday. The install happened. Which touch caused it isn't a fact sitting in a log somewhere. It is a counterfactual, and the only way to estimate a counterfactual is with an experiment.

What the MMP does instead is apply a rule. Last touch within a window says the Sunday tap gets the install. SKAN says whichever network won the winning postback gets it, subject to Apple's privacy thresholds. A data-driven model shares the credit according to weights that the vendor trained on historical conversion paths, which an earlier rule had itself labelled.

The output of a rule is a ledger, not a finding. It says: under these conventions, network A has earned 4,000 installs and network B has earned 1,200. That is genuinely useful. Someone has to get paid, and a consistent rule means the payments are at least predictable. But the ledger is the input to a budget decision, not evidence about incremental installs.

The mechanism: rules shape the behaviour of the people they pay

Here is the part most attribution coverage skips. The networks under measurement know the rule, and their bidding systems optimise to win credit under it. This is not misconduct. It is the only rational response to a rule that decides who gets paid.

Under last touch, the winning move is to be the final impression before an install that was going to happen anyway. Retargeting, rewarded placements shown to already-engaged players and high-frequency remarketing all do well under last touch, and they do well partly because the rule rewards proximity to the install rather than distance from the organic baseline. Under SKAN, with its coarse conversion values and privacy thresholds, the winning move shifts toward volume and toward campaigns large enough to clear the thresholds, which changes what gets bid on.

So when you change your attribution rule, you're changing more than your reports. You are changing the incentives for every partner you buy from, and their algorithms will move to the new optimum within weeks. The reports that come back after the change reflect the new game, not a cleaner view of the old one.

This is the second-order effect that gets missed. A team switches from a seven-day to a one-day view-through window, sees a channel's attributed installs fall by half, and concludes the channel was weaker than believed. Both readings were artefacts of a rule. The channel's incremental contribution did not change on the day the window changed.

Judging a rule by its properties, not its accuracy

If a rule cannot be true, it can still be good. Three properties matter.

Stability comes first: the rule should produce similar allocations for similar behaviour week over week, so that budget decisions aren't chasing noise.

Resistance to gaming comes second. The rule should be hard for a partner to win without also producing incremental installs, and a rule that a partner can win by showing an impression to someone about to install is weak on this property.

The third is correlation with incrementality. When you run a holdout, the rule's allocation should move in the same direction as the experimental result, even if not by the same magnitude.

The third property is the one that turns a rule into something a finance team can live with. It does not require the rule to be right. It requires the rule to be wrong in a consistent direction, so that you can apply a correction factor and revisit it.

A practical decision rule follows. Run the same month's spend through two attribution conventions your MMP supports, for instance your current window and a materially shorter one. Rank channels under each. Where the ranking agrees, allocate with confidence. Where it disagrees, treat those channels as unresolved and put them in the queue for a holdout before any budget change. The disagreement set is where the rule is doing the deciding for you, and where an experiment earns its cost.

An illustrative example

Take a hypothetical mid-core title spending across four channels. Under a seven-day last-touch window, an illustrative allocation might show a rewarded video network at 35% of attributed installs and a social platform at 20%. Under a one-day window the same two channels flip: the social platform rises and the rewarded network drops sharply, because a large share of rewarded conversions were closing installs that had earlier social touches.

Neither table is the truth. What the pair tells you is that the rewarded network's attributed volume is sensitive to the rule and the social platform's is not. That is a reason to test the rewarded channel with a geo holdout before scaling it, and a reason to stop treating the seven-day number as a reason to scale. The comparison cost nothing beyond an afternoon in the MMP dashboard.

What changes in the review meeting

Once attribution is understood as a rule, several habits become obviously wrong. Reconciling MMP numbers to a network's self-reported dashboard, and treating the gap as an error to close, is one. The two are simply different rules, and closing the gap means adopting the network's rule. Arguing with a vendor about whose data-driven model is more accurate is another, since both models learned from labels that a prior rule produced.

The habit that replaces them is a standing question for every channel that carries meaningful budget: what is our rule paying this partner for, and does an experiment agree with the payment? Industry samples from the major MMPs now show the majority of gaming installs arriving through paid media rather than organically. A rule that somebody chose assigned every one of those paid installs to a partner. Choosing it deliberately, and knowing which partners it favours, is the measurement lead's actual job.

Related archive reading

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

• Modelled conversions: the numbers nobody audits
• Privacy made UA more honest, not harder