What information do autoencoders extract

[Pang-0093]

Hi everyone, I have a question about autoencoders. I've noticed that a well-trained autoencoder can perform well on an unseen dataset. Does this mean autoencoders don't rely on the data distribution features? In other words, are the features extracted by autoencoders not the characteristics of the data distribution, but rather just some structural information used for compression-reconstruction? I am so curious about it. Can someone help me out?