Satellite observations have revolutionized Earth science, yet uneven coverage and frequent gaps limit the capture of surface dynamics. Processing multi‑modal data demands heavy computation and sophisticated alignment pipelines that many remote‑sensing practitioners cannot afford. Earth embeddings, learned from heterogeneous observations, serve as general‑purpose representations that support downstream tasks without task‑specific feature engineering or preprocessing. This paper first outlines the theoretical basis of Earth embeddings, covering self‑supervised feature extraction and cross‑modal alignment. It then surveys representative products, including Transformer‑based global embeddings, convolutional auto‑encoder local embeddings, and hybrid spatio‑temporal graph networks for sequential embeddings. The review proceeds to six major remote‑sensing domains—climate monitoring, land‑cover change, vegetation health, urban expansion, disaster response, and ocean observation—highlighting how embeddings are employed. Four case studies—hyperspectral image classification, individual tree‑crown segmentation, crop yield prediction, and above‑ground biomass mapping—compare the performance of Earth embeddings, raw imagery, and handcrafted features. Results show embeddings excel in computational efficiency and spatio‑temporal consistency, yet they do not always surpass traditional imagery or specialized features on fine‑grained tasks. Consequently, embeddings should be viewed as complementary representations rather than a direct replacement for conventional EO data. Practitioners are advised to first assess embedding suitability for their specific task and then decide whether to adopt them alone or fuse with other features.\n\nReview: Earth embeddings offer a unified feature framework that lowers preprocessing barriers for remote sensing, but careful task‑specific evaluation remains essential.