Since the introduction of Apple Mail Privacy Protection (AMPP), our teams have continuously been monitoring its effects on our systems, including the Send Time Optimization (STO) feature. Below are some of the findings.
How does STO perform since AMPP?
Since the release of AMPP we monitor the performance of STO algorithms very closely. We no longer measure by open uplift, but rather look at click uplifts. Here is an illustration of average STO click uplift since AMPP release for one of our clients.
As shown, there is no deterioration in STO performance.
What did we do so far to adapt STO to tackle AMPP?
We are experimenting with models that utilize clicks rather than opens, but there is a trade-off; even though clicks are more reliable, they are rare events compared to opens which has an impact on model performance. Our algorithms learn by analyzing data, there are simply far less click events and therefore less information on the preferences of contacts to use for optimization.
We are still planning to perform several experiments, combining click data with account-level non-AMPP opens to reach even better results.
We did, however, recently launched a new version of the algorithm that outperforms the previous one.
How is it possible that machine-generated opens do not influence the performance of our open based algorithm? What is our prediction regarding future uplifts?
The STO algorithm leverages two components.
- Account level opening patterns – the performance of each hour of the day from all users (some are non-Apple users!)
- Contact level behavior.
While contact-level opening patterns for AMPP contacts no longer provide useful information for STO, we are constantly monitoring the effect of STO and we do not observe significant decline in uplifts for those contacts.
This is possibly thanks to the information from account-level opening patterns. Account-level observations are noisier due to machine-generated opens, but it is spread randomly throughout the day making it irrelevant for the learning algorithm. Account level behavior still provide valuable information that is used to optimize send times. Moreover, for non-AMPP contacts, send time optimization functions properly, thus we expect positive global uplifts in the future.
Does STO ignore AMPP opens?
STO does not ignore AMPP opens and this is by design. The ‘noise’ from machine-generated opens does not affect the account-level performance as it is spread over random time slots.
What about new contacts?
For AMPP users (new or not new) with fake opens only, the algorithm utilizes account-level information to optimize sending times.
Why do we still show opens on the interface?
This aspect is being discussed and while we communicate to clients the status of open data, we are working on the best possible way to gradually adapt our platform accordingly, based on the data we observe.
For more information, see our collection of product-specific advice on handling lack of opens.