This paper addresses the issue of coalescence of frequent collocations and its consequences for their realization and mental representation. The items examined are ‘semi-modal’ instantiations of the type V- to -V inf , namely have to, used to, trying to and need to , in American English. We explore and compare their realization variants in speech, considering the effects of speech-internal and extra-linguistic factors (speech rate, stress accent, phonological context, speech situation, age of the speaker), as well as possible effects of analogy with established contractions like gonna, wanna . Our findings show a high degree of coalescence in the items under study, but no clear pattern of contraction. The propensity for contraction in analogy to gonna/wanna is strongly affected by phonological properties – it is inhibited by the presence of a fricative in have/used to . Moreover, the most frequent reduced realizations are conservative in terms of transparency and still allow morphological parsing of the structure. More radical contractions are restricted to rapid and informal speech, and less entrenched as variants. This shows the limitations of reduction as a frequency effect in light of the balance between articulatory ease and explicitness in speaker–hearer interaction. Even in highly frequent and strongly coalesced items, reduction (articulatory ease) is restricted by a tendency to retain cues to morphological structure (explicitness). Finally, we propose a network of pronunciation variants that includes representation strengths as well as analogy relations across constructional types.