Why did CPMs increase following App Tracking Transparency?

A common question I field is: “If Apple’s App Tracking Transparency (ATT) privacy policy degraded the efficacy of ads targeting, making it harder for advertisers to reach relevant customers, why did observed CPMs increase on iOS following its rollout?”

It’s somewhat of an enigma. If targeting efficacy is diminished as a result of ATT, then, on a relative basis, an advertiser’s ads would be exposed to less economically valuable users, and so that advertiser would bid less for advertising inventory. It seems logical that CPMs should decrease as a result of inferior targeting, not increase. But CPMs did, indeed, increase: in its Q3 2021 earnings, on a year-over-year basis, Facebook noted that its average price per ad increased 23%. In its Q4 2021 earnings, reporting on the first full quarter for which ATT was operative at majority scale on iOS, the company indicated that its average price per ad increased by 6%. How could this happen if advertisers were reaching less relevant, or fewer relevant, users?

There are three potential explanations.

The first — rather mundane — explanation, cited by Facebook in its earnings, was advertiser demand increased from the same periods in 2020 as a result of the softening of COVID restrictions and increased consumer spending. Indeed, COVID caused a decline in CPMs across the digital advertising ecosystem. If observed CPMs were depressed in the early, global onset of COVID, then it’s certainly possible that CPMs would revert to pre-COVID levels as the economic impact of COVID waned, even against ATT headwinds.

The second is simply that some advertisers allowed their margins to compress in the face of ATT. If an advertiser was achieving a very favorable return on ad spend (ROAS) prior to ATT, it could likely afford to let its profit margin deteriorate while still earning a profit as ATT eroded targeting efficiency. Note that in both periods cited above, Facebook also saw year-over-year advertising revenue increases (35% and 20%, respectively).

And the third explanation relates to the way performance advertisers structure campaigns. An observed CPM metric, or cost-per-mille (thousand impressions), is simply an accounting of cost against ad exposures. Note that CPMs are not set by most ad platforms (except in the case where CPM floors are established) and are determined through advertising auction mechanics. Performance advertisers almost universally bid against outcomes, or conversions, in a CPA, or cost-per-action, model. An app advertiser might bid against app install outcomes or outcomes related to in-app purchases. An eCommerce advertiser might bid against outcomes related to product catalog engagement, such as when items are added to a shopping cart, or bid against purchases.

There is nuance to this: in a CPA model, an advertiser bids on an outcome (conversion), but that bid is adjusted on the basis of conversion probabilities for that action, and the expected value of that bid is used in the auction to determine the winner of the impression. For more background on digital advertising auctions, see this Mobile Dev Memo podcast episode; for background on how conversion probabilities are calculated, see this QuantMar thread.

Because the winner of the auction is determined on the basis of expected value — that is, the probability that serving a specific ad in the impression will generate the advertiser’s desired outcome, multiplied by the advertiser’s outcome-specific bid (eg. the advertiser bids $5 for an install) — an advertiser might pay for impressions that don’t result in conversion outcomes. The diagram above showcases how an auction for an impression might be resolved when each advertiser bids against different outcomes. The diagram below showcases how a single advertiser might pay for impressions that don’t result in outcomes even when they are bidding for those deliberate outcomes.

CPM is often thought of as an explicitly-determined price of advertising inventory, but it’s not: it’s simply a measure of the amount of money spent on advertising campaigns, which is derived from auction mechanics, divided by the impressions served for the campaign (divided by 1000). It’s important to reiterate that most ad platforms don’t set prices for ad inventory.

The auction mechanism takes as inputs an advertiser’s bid logic and conversion probabilities, which determine the cost an advertiser accumulates for its ad campaigns. And the bid an advertiser uses should be informed by the value ascribed by it to whatever action the campaign is optimized to generate. It is that CPA bid, informed by the advertiser’s own internal business logic and moderated by the efficiency at generating conversions of the advertiser’s targeting parameters and ad creative, that ultimately determines CPM.

It makes eminent sense that ATT would cause CPAs to increase given that ATT diminished ad campaign targeting relevancy. In the hub-and-spoke model of digital advertising, ad platforms are able to target individuals on the basis of their behavioral profiles, and so ads are paired to users on the basis of estimated propensities to engage with the products being advertised. Once those behavioral profiles cannot be aggregated, because of the disruption of data flow between advertisers (who can observe conversions in a first-party setting) and ad platforms (who can only observe ad views and clicks in a first-party setting), the ad platforms must resort to targeting users on the basis of group identities, such as with demographic features (age, geography, etc.) and interests. These groups are by definition more coarse and broader than individual-level targets, since an individual has a group size of one. And given the general rarity of conversion events on a cohort basis, it makes sense that ad relevancy decreases when targeting shifts from a specific individual with a known propensity for some product category to a group of people.

Another way of considering this is with a thought experiment that I have borrowed and adapted from this lecture by Nassim Taleb. Imagine you are in a busy shopping mall in a Midwestern city. Someone approaches you and informs you that the average net worth of a person in that mall at that moment is $50MM. Is it likelier that you’re surrounded by millionaires, or that most people in the mall have a net worth consistent with the median net worth in the United States ($121,700) and there is simply one billionaire present in the mall? Now imagine the same scenario, except that it happens at an exclusive dinner party at the World Economic Forum in Davos. What are the likely distributions of net worth in each of those settings?

Nassim Taleb calls the distributions in which extreme, rare events have a disproportionate impact on the parameters of those distributions “fat-tailed,” which in the vernacular of statistics translates into heavily skewed or highly kurtotic. In the Millionaires’ Mall, the distribution of net worths is more likely to be fat-tailed, with a high concentration of people with net worths between something like $0 and $250,000 and a single person with an extremely high net worth in the billions. At Davos, the distribution of net worths more closely resembles the Gaussian distribution, or at least it doesn’t skew towards the higher end of the distribution as meaningfully and its mode is closer to the center of the graph. Fat tailed distributions are commonly observed; below is the distribution of populations of US cities with a population of at least 10,000, which follows a power law.

This is relevant to the discussion of ads targeting because user-level targeting allows advertisers to reach audiences of relevant individuals based on behavioral data, and group-level targeting utilizes broad and somewhat arbitrary characteristics to classify users into audiences.

With group-level targeting, the advertiser assumes that the average economic value of the very diverse collection of people exposed to ads means something, when more often than not, the average is impacted by a small minority of high-value users. I unpack this idea in How does IDFA deprecation impact ad prices? and in this nearly decade-old post.

User-level, behavioral targeting produces audiences that look like the Davos dinner party; group-level targeting produces audiences that look like the Millionaires’ Mall, with just an incredibly small proportion of users dominating the descriptive statistics for past cohorts (it should be noted that for some power law distributions with extreme exponent parameters, a mean cannot be calculated at all). User-level targeting allows for audiences to be constructed in a “bottoms-up” approach: group users with similar behavioral histories and observable interests together, and expose ads to that audience. Group-level targeting segments users into audiences utilizing the only features available to advertisers, like demographic features and elected interests, which aren’t very helpful in determining relevance for many ad campaigns.

This dynamic resulted in a general increase in cost-per-action (CPA) metrics following the rollout of ATT. In order to reach relevant users, advertisers must use the demographic features from previously-acquired cohorts to establish priors regarding group-level average values.

But as in the case of the Millionaires’ Mall: the average value of cohorts might be meaningless, or at least not instructive for bid logic, given that it is skewed by a small number of high-value users. This targeting approach results in wasted impressions — which are paid for. When CPAs increase as a result of diminished ads targeting relevancy, then CPMs, a simple measure of ad spend divided by impressions purchased, also increase.