Send Time Optimization (STO) is an AI-driven add-on for your email campaigns, which analyzes the email responses of every recipient in relation to the campaign launch time, and then sends the next email at the optimal time for them.
What is Send Time Optimization?
Send Time Optimization staggers email campaign launches over 24 hours and sends each email at a time after which the recipient is most likely to open it.
How does it work?
It uses machine learning to analyze your contacts’ behavior and identify the times when they are most responsive. This is independent of time zone, language or region. The contacts are then divided among 12 child campaigns, sent on the hour at two-hour intervals.
For new contacts with no behavior history, their first emails will be sent at a time calculated on the basis of your account’s aggregate historical data on open times. But the algorithm analyzes new responses for every contact on a daily basis and updates its model accordingly. This means that we will have more accurate data over time. Our research shows that, after sending 10 send time optimized campaigns, you can expect around 2-12% increase not only in opens, but also in clicks and purchases.
So that you can measure the benefits of STO, a control group of 10% of the launch list is also created for each campaign. The engagement results of this control group are used as benchmarks against which STO results are compared. This is done completely automatically, so you do not need to devote time and effort to performing A/B testing on your own.
Are there any limitations?
- Send Time Optimization can currently only be enabled for ad hoc email campaigns and for batch email nodes in Automation Center programs that do not start with a transactional entry node.
- Currently, the option Do not send to duplicate email addresses, which you can select in the Email Settings step when creating your campaign, does not apply to STO campaigns.
- The campaign launch cannot be scheduled for the current day, the earliest available option offered in the date picker is the following day.
- Send Time Optimization can handle user lists that contain maximum 10 million contacts: the target segment or the Automation Center list of participants cannot contain more users. If the user list is bigger, sending will be cancelled and the account owner will be notified via Emarsys Support.
As Send Time Optimization needs past contact behavior data to determine the optimal sending times, we do not recommend using STO until you have at least 1 month of email campaign history. Turning STO on too early can actually result in a decline in your engagement results.
Why is this good for my customers and for me?
You want to maximize the chances that your emails will engage your customers. Send Time Optimization is designed to help you achieve just that. For your customers, this is one more example of how you are treating them as an individual, strengthening their loyalty to your brand and improving their experience.
In most cases, no. For time-sensitive emails, such as limited special offers or discounts, you should consider the duration of the offer before using this feature. If an offer is only available for a 24-hour period, for example, you will not be giving all your customers the same advantage if STO is used.
Currently, no. Our algorithm analyzes contact events in order to determine the best possible send times. Excluding hours manually would decrease the efficiency and the performance of the algorithm.
No. Scheduled campaigns are launched from 00:00 through to 22:00 on the calendar day of the launch (in the timezone of the account), and manually launched campaigns are staggered over 24 hours from the time the calculations are finished. No other control over the timescale is possible.
The technology behind STO
Both. We cannot know for sure which is the best hour for any contact, so it is important to experiment with send times we have minimal or no information about. On the other hand, we obviously want to exploit the knowledge we already have and send at times which perform well with high probability. The algorithm finds the best trade-off between exploration and exploitation. The more historical data you have the better results you can expect as STO needs less exploration and can do more exploitation.
If you are an existing customer, we use your account’s aggregated historical data on open times to assign the most likely send time to your new contacts. If you are a new customer with no historical data, new contacts are randomly distributed among the launches if STO is turned on. That is why it is highly recommended to wait at least a month before activating STO.
The algorithm automatically adapts to consumer preference changes: for example, if the client moves to a country in a different timezone or their daily routine changes, the algorithm will gradually adjust the sending time as we receive this feedback.
If there are two, or more, similarly performing send times the algorithm will send to these with similar frequency; it is not a goal to pick one of them. This ensures that we have the possibility to find out if one of them is actually better in the long run, even if both produce very similar results for a period.
Partially, yes. It still collects response data on an account level, so the underlying statistical model will be better when you next switch it on, benefiting future contacts.
We count any open as a success, no matter how long after the launch it comes, so a contact will continue to receive emails at that time, as long as they continue to respond. No open response for a campaign will cause the algorithm to reevaluate the time for that contact, based on its current statistical model.
It usually needs at least 10-20 launches to see the first stable improvements, until then, results may fluctuate, even deteriorate, so be patient. You will see the results in time.
The calculations begin as soon as Launch now is clicked, and control group email will be sent as soon as they are finished, that is, within a maximum of two hours. The first real STO launch will happen on the following even hour.
Therefore a delay of up to three hours after clicking the Launch now button is, in theory, possible.
Each contact’s profile is updated every 2 days with the latest engagement information in order that we have the most recent data at our disposal.
STO applies a method called a multi-armed Bayesian bandit algorithm, if you must know.
The effects of STO
For any given week, we compute the aggregated open rate for both STO and CONTROL launches, then calculate the difference as a percentage of the CONTROL open rate: (STO open rate - CONTROL open rate) / CONTROL open rate as %.
- If the two groups performed identically, e.g. both have a 15% open rate, then the added value of STO will be 0%.
- If the control group launches have a 15% open rate and the STO launches 16.5%, then the added value of STO will be +10%, since the +1.5% difference is 10% of control group's 15% open rate.
No. We have found no significant difference during the pilot phase in unsubscribe rates between STO and non-STO campaigns.
If the Frequency Cap is active, it will apply to each launch at the moment of launch. For example, if an STO campaign is scheduled for a day when contacts whose Frequency Cap is set to 1 have already received a message, the STO child campaigns launched on that calendar day will not be sent to them. But they will receive the launches sent on the next calendar day.
Yes. In such cases the given launch will not be created at all, which means that there will be fewer than 12 launches in the campaign. But if you have a large contact list containing thousands of email addresses, and especially if these are geographically dispersed, this is highly unlikely to happen. You can always check the statistics breakdown for each STO campaign by going to Analysis > Emails and clicking the plus sign to see the results of the individual launches.
About the control group
Completely randomly. For each campaign, the members of the control group are picked at random so when we compare the results the only deliberate difference between the two groups is STO.
We need a large enough control group in order to reliably measure results. On the other hand, we do not want to adversely impact your engagement results. We think that 10% is a good trade-off.
No. The control group is not optional. It is the basis of all monitoring and reporting, and it also saves you the time and effort you would need to invest in A/B testing.