Personal Delivery Time helps to improve the results of your campaigns by optimizing the "when" of your message strategy. Personal Delivery Time combines machine learning algorithms and hand-crafted rule-based strategies to determine an optimal time to deliver a push message. The optimal time is identified as the time when a particular user has the maximum likelihood of starting a session within your mobile app or website.
When scheduling an App or Web Push message, you can Personalize Delivery Time to deliver your message at the optimal time of day for each recipient, which is calculated daily.
For your active users (users with at least 3 sessions in the last 30 days), PDT uses historical session data computing number of sessions, session start time, session recency, and session duration. The predicted send time is based on the hour with the most activity.
For less active, non-active, or new users, PDT leverages aggregate data across your entire app to predict the best time to send. First, the algorithm defines a window of time when your users are least likely to use your application. This “low usage time” is our way of applying Quiet Hours algorithmically and usually falls between 1am and 6am in a user's local timezone. These values vary for a particular app and the day of week. This window of time is calculated daily. Messages will never be sent within this window.
Next, the algorithm distributes the delivery time for each message across the day (with the exception of "low usage time"). Based on the data available from previous PDT campaigns, messages will be delivered at times that have historically showed higher open rates. For example, if we know that messages delivered between 9am to 10am receive more engagement than delivered between 10pm to 11pm, more messages will be delivered during the first interval than the second.
Personalizing the Delivery Time happens at the Scheduling stage of sending a push message. You can Personalize Delivery Time for one-time and recurring campaigns.
Simply check the box for Personalize Delivery Time. You can schedule the message to be delivered on a specific day and Localytics will decide the specific time. You can also set a fallback time—more on that in the next section.
How does the algorithm work for my first PDT campaign?
When you use PDT for the first time, all active users will receive your message based on their activity. For less active users, we do not have any statistically reliable information so messages are distributed at random, excluding "low usage time" hours.
How fast does the algorithm learn?
The algorithm starts learning right after your first PDT campaign is complete in all timezones. Our observations show that you should see a stable improvement after 5-20 campaigns. This figure strongly depends on your campaign size. The algorithm learns slower on small campaigns.
How do you define the efficiency of PDT and how it is reported?
We look at the session open metric within 1 hour after a user receives a push message to calculate the open rate (OR). Then we compare the open rate for those who received the message at the predicted time against those in a control group. The control group is a group of users who receive your message at a random time (excluding "low activity time").
The uplift for PDT is calculated as (open rate (PDT) / open rate (Control) - 1)*100%
What is the PDT control group and can I turn it off?
A part of your audience will be assigned as a control group. The users in the control group will receive a message at a time randomly generated for each user. The time which was identified as the time with low usage will not be used. Currently, it is not possible to disable the control group, because it allows the machine learning algorithms to dynamically adapt for changing user patterns.
What is the fallback time?
The fallback time is used as the delivery time in cases where our machine learning algorithms fail to execute. This is very rare.
Why is PDT better than a user timezone campaign?
PDT automatically adjusts the time of message delivery to maximize engagement (session open). While it is possible to manually explore which time works better using many time zone campaigns, machine learning does exactly the same job much more efficiently and eliminates the guesswork.
What is a typical use-case for PDT?
Our research shows that PDT is beneficial for a variety of campaign goals and audiences. There are, however, two important limitations you should consider before using PDT in your campaigns:
- The uplift can be reliably calculated only for campaigns with audience for at least thousands of users.
- You should not use PDT for time sensitive communication, for example, informing a user that a promotion is about to end soon, because some users will potentially have too little time to react.
Is throttling supported for PDT campaigns?
No, PDT and throttling are mutually exclusive. However, as PDT distributes the delivery time across the day you should see the load on your servers also distributed in time. If you are using throttling, we advise you to try PDT. However we have no guarantees for the time distribution used in PDT.
Is the data on my users shared with other Upland Localytics customers?
No. All machine learning models are trained on a per-customer basis, there is no data sharing between customer accounts.
What happens if a user changes their timezone?
The optimal delivery times are calculated in the user’s timezone. If a user moves to a different time zone their information will be updated and they will receive a message at the optimal time in their current time zone. There may be a delay between that moment when the user moves to a new time zone and the information is updated in our servers.
Does the algorithm learn while I do not use PDT in my campaigns?
Yes, but not as efficiently when you use PDT.
Why do I not see PDT uplift in my true impact campaign report?
There are a few reasons for PDT uplift is not shown on the true impact page:
- You have not completed enough PDT campaigns to generate statistically reliable data. You should complete at least 5 campaigns before you will start getting results
- Your audience size is too small to calculate the lift in a statistically reliable way. Try to use PDT on your larger campaigns.
What is the machine learning technology behind PDT?
Usage of PDT does not require any background knowledge in machine learning. But If you are curious, we are using contextual multi-armed Bayesian bandits with a combination of Thompson and epsilon-greedy sampling.