@AzureUser-9588 I think this guidance is based on the context that some customers use LUIS to train different apps or models where they intend to utilize the work done on one app to build different scenarios or similar responses in future iterations of the app. But, want to retain the original app so they have the option to export the app and import it with a different app id and train it before publishing it to a different prediction endpoint.
The balance here is basically a scenario where we want to use different prediction endpoints for different apps which have similar definitions and are used by the traffic manager to handle the quota limitations. This section or guidance from documentation should be relevant to understand this scenario. Hope this helps!!