Attribution has always been one of the pillars of mobile marketing.
However, in the “new normal” environment, the way we do attribution has changed quite a bit. No, we’re not talking about the impact of the pandemic, but about the effects of Apple’s privacy updates.
In the privacy-first era, more and more marketers are starting to rely on probabilistic attribution.
Here’s everything there is to know about it.
Defining probabilistic attribution
Just like its name suggests, probabilistic attribution is a type of attribution based on – probabilities.
Probabilistic attribution is a process used to link users to certain mobile marketing campaigns.
In this process, users get assigned with campaign membership probabilities. Essentially, these are percentages that tell marketers how likely it is that a certain campaign triggered an install.
For example, with it, marketers can find out there is a 60% probability a user came from campaign A, 10% they came from campaign B, etc.
Why do marketers need probabilistic attribution in the privacy era?
Until recently, probabilistic attribution was overshadowed by deterministic attribution. This is the type of attribution in which there is 100% certainty a certain campaign drove an install.
Deterministic attribution is dependent on identifiers for advertisers. And ever since Apple devalued the IDFA with their ATT policy, marketers are looking for alternatives to attribute app installs.
Probabilistic attribution turned out to be one of the best ways to do so. It may not be 100% accurate, but it is 100% privacy-compliant.
With it, mobile marketers can get a head start over those who rely on deterministic attribution only. The data from probabilistic attribution is especially valuable for predicting lifetime values and allocating ad budgets.
Probabilistic attribution vs. fingerprinting
Probabilistic attribution is frequently mistaken for fingerprinting. While they have some common traits, they are not the same thing.
Fingerprinting falls into the group of probabilistic solutions. However, unlike probabilistic attribution, it does not comply with Apple’s privacy guidelines.
Fingerprinting is based on collecting IP addresses and other device-related data to distinguish users. Regarding this, Apple has explicitly stated that it is not allowed to collect data from a device with the intention of identifying it.
On the other hand, probabilistic attribution doesn’t use any kind of personally identifiable user information. Instead, it relies on the “allowed” data sets on iOS users.
A refresher: user-level data sets in iOS 14+
It’s pretty clear that probabilistic attribution is powered by machine learning.
But what do these data-hungry machines feed on exactly?
Probabilistic attribution solutions utilize different user-level data sets on iOS users:
- Anonymized user-level app data (e.g. revenue, in-app event data)
- Data from SKAdNetwork (e.g. conversion value)
- Available deterministic data (e.g. data from ATT opt-in users)
- Ad networks' campaign reporting data (e.g. impressions and ad spend)
The difference between probabilistic attribution and “winner takes all” attribution
Let’s dive deeper into how probabilistic attribution works and how it compares to other popular attribution approaches.
For this purpose, we can imagine a set of users.
These users can come from several different ad campaigns or from an organic source. For one of these users, membership probabilities can be distributed like this:
- 35% likely to come from campaign A
- 30% likely to come from campaign B
- 20% likely to come from campaign C
- 10% likely to come from campaign D
- 5% likely the install is organic
Based on these probabilities, it is possible to make revenue projections for each user. The revenue for every user can then be allocated to each campaign, with the membership probability factored in.
Next, these user-level projections are used to project campaign-level revenues. Finally, the projected revenues are compared against campaign spend, and used to predict ROAS (Return On Ad Spend).
Probabilistic attribution is commonly compared to another popular approach called “winner takes all”.
In it, the campaigns with the largest membership probabilities are given all the predicted revenue. Meanwhile, the campaigns with lower probabilities are considered “losers” and simply get ignored.
This approach is flawed, and here’s why.
When it comes to users who are 90% or 75% probable to come from a certain campaign, it makes sense. However, it puts all the other users in the same basket, and this can be misleading.
As a result, this “winner takes all” fails to recognize the long-tail probability of users coming from different campaigns. Just like that, big amounts of revenue can end up being mismanaged.
Probabilistic attribution and Conversion Value: The relation
Probabilistic attribution isn’t intended to become a replacement for SKAdNetwork. It is simply a valuable addition to it.
As mentioned before, one of the data sets used in probabilistic attribution is SKAdNetwork data. The most important piece of information received from it is Conversion Value (CV).
The CV tells marketers about the potential value of a single app user. It is mainly used to measure user engagement or in-app revenue early in the users’ in-app journeys.
For example, an event like “level 1 completion” can be traced back to certain campaigns. Thanks to CV, it is possible to find out how many “level 1 completion” events came from campaigns A, B, C, etc.
As such, conversion values give important early signals for optimizing ad campaigns.
However, CVs aren’t enough for precise ROAS estimates. For this purpose, advertisers need more than early signals. They need additional behavioral data to help guide their decisions.
This is where probabilistic attribution comes into play.
Probabilistic attribution utilizes diverse sets of behavioral data. It complements the CV data with the available data on matured user cohorts.
This creates a powerful data combination that helps advertisers make more accurate LTV projections. Based on that, they are able to predict asset level ROAS and distribute projections across different channels and campaigns.
How Tempr. uses probabilistic attribution to predict asset-level ROAS and CPE
The main point of probabilistic attribution is to use the power of data to make predictions.
Tempr.’s prediction solution is created to make the most out of data. It collects historical data, data from different UA channels, and tracking tools. Next, its powerful machine learning combines all this data and creates thousands of scenarios with varying parameters.
Granular predictions for KPIs such as return on ad spend (ROAS), Cost Per Event (CPE) & Life Time Value (LTV).
We can all agree that being aware of campaign predictions is great. But what’s even better is getting recommendations based on these predictions.
Tempr.’s solution has been tailored for this purpose.
Based on the probabilistic predictions, it gives advertisers tips on how to optimize their bids, budgets, and ad creatives. All with the goal of optimizing their campaigns towards the desired ROAS and other KPIs.
Let’s turn probabilities into results
Congrats – you now have a full understanding of how probabilistic attribution can help you out in the privacy era. Ready to reap its benefits? Reach out to Tempr.’s team and schedule a demo.