The term probabilistic attribution has been co-opted, somewhat cynically and disingenuously, as a synonym for mobile device fingerprinting. This is unfortunate because, conceptually, probabilistic attribution describes a broad basket of model-based activities that attempt to credit ad interactions with user-level engagements with products. The models that power this probabilistic assignment process can take a range of inputs:
- Device-level parameters for the purposes of pairing a user with some ad engagement. The process of fingerprinting (or, as I’ve termed it, “probabilistic install attribution using device parameters,” or PIAUDP) invokes such logic: this user saw that ad, and this user appeared later in the advertised app, therefore that ad is credited with the acquisition of this user;
- On-site behavioral signals that can be matched to similar behaviors from users where the source channel is known. This form of probabilistic attribution invokes different logic: this user behaves similarly to that user, and that user is deterministically known to have been sourced from some channel, therefore we should credit some channel with the acquisition of this user.
Probabilistic attribution, either via fingerprinting or the use of on-site behavioral signals, supports the promise of user-level campaign attribution in a setting where user-level identifiers are unavailable for deterministic attribution. And this is the setting in which the entire mobile ecosystem operates, at least on iOS. This new environment was mostly catalyzed by Apple through initiatives like Intelligent Tracking Prevention (ITP) and App Tracking Transparency (ATT), but the same structures of user identity are likewise being demolished by Google generally with the deprecation of third-party cookies in Chrome and the GAID on Android. Advertisers must embrace a future in which conversions cannot be deterministically attributed to ad engagements with user identifiers.
I believe that fingerprinting on mobile will, in the very near-term future, be policed by Apple with potentially the approach that I hypothesized back in February in How Apple might break fingerprinting in iOS 16. That leaves probabilistic attribution using in-app (on-site) behaviors for accomplishing user-level campaign associations. When ATT was first announced, many advertisers pursued classification algorithms that would use in-app signals to probabilistically attribute users to source channels: these classification systems attempted to use in-app behaviors to determine the advertising channel responsible for a given user’s provenance.
This always seemed unrealistic to me, for a number of reasons:
- Most advertisers operate across multiple channels: while the “quality” (observed, cohort-based average value of users) delivered by these channels often differs meaningfully, those differences are not necessarily perceptible early in cohort lifetimes in a way that makes probabilistic, channel-level advertising attribution helpful or useful (that is: an advertiser would want to attribute ad engagements within a very short amount of time — 24-48 hours — in order to do anything useful with that information);
- Almost definitionally, the channels that absorb the most ad spend tend to provide the advertiser with the most users. This dynamic will engender a hazardous bias in any model: the users that appear to be the most valuable will be classified as having converted from the channels that historically produce the most valuable users. Because spend is not evenly distributed but often tends to be captured in the majority by just one or two channels, a channel-level classifier is likely to privilege the channels that historically deliver the highest-quality cohorts. I speak to this effect in The “Quality vs. Volume” fallacy in mobile user acquisition;
- Channel as a feature of cohorts is not operationally helpful. Attempting to break cohorts apart by source channel doesn’t produce any operational advantages for a marketing team, since every “channel” is effectively a bucket of campaigns, but the added complexity in attempting to accommodate channel in a classifier risks inviting Simpson’s Paradox into an analysis. Classifying users by source campaign would dimensionalize the data too granularly to be viable, but classifying users into source channel doesn’t necessarily aid with campaign optimization, although it can be helpful for cross-channel optimization when utilized by a Media Mix Model (MMM);
- Only a few interactions matter in evaluating the economic quality of cohorts. Retention and monetization are the only real behavioral categories to consider in assessing the value of cohorts, and users within those cohorts either do those things — retain, purchase — or they don’t. This idea gets muddied when thinking about “average” values for properties of cohorts (eg. ARPU, LTV) that are determined by fat-tailed distributions. If Cohort Y is “worse” than Cohort Z because it has a lower measured ARPU, is it more likely that (a) every user in Cohort Y monetized to a smaller magnitude than in Cohort Z, or that (b) fewer users in Cohort Y monetized than did in Cohort Z, but all monetizing users in both cohorts monetized to roughly the same magnitude? Often, the latter. So if “good” users look the same across all channels because any app only presents a few opportunities to monetize, but some channels produce more monetizing users, then it becomes difficult to distinguish good users by channel, and the challenge raised in the second bullet is activated.
To my mind, probabilistic attribution at the level of the channel, much less at the level of the campaign, is impractical to impossible using in-app behavioral signals. And if fingerprinting is policed, that activity will not be possible using device parameters, either. That leaves advertisers with one tactic for improving funnel conversion and thus the efficiency of advertising spend: in-app personalization.
As I note in How to scale and optimize marketing spend with SKAdNetwork, prior to the deprecation of device identifiers, all app personalization was outsourced to advertising platforms. These platforms reached the most relevant audiences based on aggregated behavioral profiles sourced from other properties. Now, absent those profiles, ad platforms are less able to locate relevant users for exposure with ads.
Given this limitation, it is incumbent on app developers to find ways to parse apart the broader, more heterogeneous accumulations of users that are presented to them as cohorts by ad platforms into meaningfully-defined groups. These groups can then be exposed to appropriately differentiated in-app product treatments, designed to optimize the user experience at the smaller group level.
Many developers were pursuing this approach prior to ATT; some colleagues and I wrote a case study on how we were able to increase game revenue by 10% using very early in-game signals as inputs to a personalization engine for an in-game special offer. These kinds of efforts become profoundly more valuable in the post-ATT world because ad platforms cannot facilitate the conversion-optimizing feedback loop between advertisers’ properties and the platforms’ own data environments any longer, as I describe in this piece.
Instead of receiving users that were targeted on the basis of past relevant behaviors, advertisers in the post-ATT environment receive a “blob” of heterogeneous traffic, targeted via broad demographic features like geography, gender, and, potentially, expressed interests. As I propose in Why did CPMs increase following App Tracking Transparency?, targeting at the level of a group is axiomatically less efficient than targeting individuals. And because ad platforms can no longer fulfill this service on behalf of advertisers, then advertisers must ingest this functionality into the product environment, optimizing the product experience — as opposed to the ad experience — to the tastes and preferences of individual users, utilizing first-party data. This isn’t an advertising exercise; it’s an in-app personalization exercise that exclusively relies on first-party data in a way that is fully privacy compliant. And it will be increasingly necessary as identifiers are jettisoned in the advertising ecosystem.