Much research examining the development of semantic knowledge has focused on patterns of co-occurrence in language. An often-made distinction concerns ‘local’ co-occurrences (word pairs that appear close together) and ‘global’ co-occurrences (word pairs that appear in similar linguistic contexts), with some recent accounts arguing that only local co-occurrences are exploitable by young children. We tested the alternative view that this dichotomy may be misaligned with natural language input. First, we performed a descriptive analysis of the co-occurrence structure present in the input. This analysis suggests that global co-occurrences are frequent in the input to children. Second, we performed a descriptive analysis of the semantic information encoded by local and global statistics in the CHILDES corpus; this analysis suggests that both local and global co-occurrences encode patterns of similarity that could support the acquisition of structured semantic representations.