The main difference between the new, improved and the formal (old) version of STO algorithm is how they behave for new contacts and contacts with relatively limited behavior history. Note that limited behavior history includes those cases when there is a solid amount of engagement data for a contact, but email sends are limited to only a few send times. As described in the Send Time Optimization - Overview, the algorithm relies on account level opening patterns to decide on optimal send times for these contacts and explores their personal preferences with time. While in general this is true for both algorithms, there is a difference in how account level information is utilized in the process. The new algorithm has a more sophisticated way of using account level opening patterns that differentiates between contacts based on their overall activity level.

As a general rule, the improved algorithm tends to have a longer learning process than the old one had. However, as a result, it has a better overall performance in the long run. As described in the  Send Time Optimization - Overview, any STO algorithm has to balance the cost of exploring send times we have minimal or no information about and the benefit we gain by exploiting the information we already have. It turns out that the formal STO algorithm did too little exploration for some contacts that resulted in a sub-optimal performance. The improved algorithm tends to explore more potential sending times and consequently it is more likely to learn the actual preferences of the contacts.

All the data collected by the formal algorithm is utilized by the improved algorithm. However, as described above, the reason why the improved algorithm improves overall performance is that it better balances collecting new information and exploiting the existing data. At the beginning of its introduction, the improved algorithm tends to explore potential new sending times at a cost of a temporary performance drop. On the flip side, this exploration phase results in a better long-run performance.

The unusual send times are the result of the learning period of the improved algorithm. The formal algorithm had the tendency to rely too heavily on the information collected on popular send times at the cost of missing the opportunity to explore the individual preferences of contacts. With thorough measurements we proved that exploring these ‘unusual’ send times results in a better long-run performance in most cases.

Yes it will be able to detect it. The improved algorithm is even more likely to learn quickly the new patterns in contact behavior than the formal algorithm was able to learn it.

Currently, the algorithm treats AMPP opens as real engagement (no change compared to the previous algorithm in this sense). There is ongoing research on how to mitigate the effect of these fake opens. At the same time, we are monitoring the performance of the STO algorithm using click rate uplifts which we believe to be unaffected by the AMPP feature.