The role of segment similarity in early (i.e., partially learned) lexical entries was assessed using artificial lexicons in a referential context. During a learning phase participants heard 40 nonsense words, each accompanied by an unfamiliar picture. In testing, participants heard the direction ''Click on the [X]'', and chose which of four pictures was the target (X). Target lexical items (e.g., "pibo") appeared with foils that were similar: cohort items ("pibu"), rhymes ("dibo"), matched consonants ("pabu") or matched vowels ("diko"). Two initial experiments demonstrated cohort and rhyme confusions, similar to lexical activation findings. Four further experiments explored the role of segment similarity in word confusions. Consonant-matched CVCV stimuli were more strongly confused with each other than were vowel-matched CVCV stimuli. Placing consonants in syllable-final position (VC[f]VC) weakened consonant effects and strengthened vowel effects. These results suggest that syllable-initial segments play a strong role in word similarity and constrain the organization of new lexical items.