A new study maps 20,000 English words by their visual mouth shapes (visemes), demonstrating that lip-reading errors are non-random and structurally driven by dense, compressed clusters of visual look-alike words within the linguistic network landscape.