You have the ability to enable auto-optimization for traffic allocation among experiments for trigger(s) for segment triggered, event triggered and recurring campaigns. For a trigger, you can either chose to have current behavior of allocating percentages to each template for a trigger or you can let Blueshift auto-optimize traffic allocation based on the creatives performance in that trigger.
Each time you create a trigger, auto-optimization strategy with "unique click rate" would be marked a default under "Optimize allocation to creatives" and you won't see custom percentages against creatives. If you wish to change it to custom percentages or to change the auto-optimization metric, you can click on the hyperlinked link and change it in the pop-up. Please note, once you select either strategy, you won't be able to switch it, once the campaign launches. Also, if you add a new trigger you won't be able to change strategy once you hit "Create Trigger".
How it works
Once you select auto-optimization for a trigger and launch the corresponding campaign, equal traffic (percentages) will be allocated to all the creatives. e.g. if there are 3 creatives in trigger, percentages set would be 34%, 33% and 33% respectively. At any point in time, you can disable traffic being allocated to a creative. There should at least be 1 enabled creative in that trigger. Disabling all experiments for this trigger won't let you save that trigger.
Once you disable a creative, percentages would automatically be refreshed and updated.
For auto-optimization to kick in, trigger should satisfy 2 parameters.
1. Minimum sends from this trigger should be > 10000
2. 14 days (learning window) must have passed since latest creative was added to that trigger
The learning window for auto-optimization algorithm is 14 days from the latest creative added to that trigger. e.g. if you add 2 creatives on day 1 and then add 1 creative on day 5 and then don't add any creative afterwards, auto-optimization will kick in from day 19 (5 + 14). The trigger will allocate traffic uniformly to enabled creatives in learning window.
Once, the above 2 criteria's are satisfied, auto-optimization will kick in and update percentages as determined by multi arm bandit algorithm based on selected optimization metric and sends per creative, among all creatives, for that trigger.
Once auto-optimization for a trigger starts, it will update the percentages for all the enabled creatives in that trigger. We define "priors(a,b)" for all the new experiment added to the trigger and initialize with the optimization metric stats and send stats for enabled winner creative respectively, so as not to starve the newly added creative and give it a head start as equal to winner creative. The value of priors for creatives pre-optimization period are 0,0 respectively.
At any point of time, you could enable/disable/add a creative with following scenarios -
1. If a creative is disabled, its stats won't be considered in calculating for any further newly added creatives
2. If a creative is enabled, its priors will be set (if not already set earlier).
3. If a creative is added, its priors will be set to be equal to enabled winner creative's statistics.
Winner will be selected from only those creatives which have completed their duration for learning window time.