Alphabetic writing systems use combinations of individual characters (e.g., letters) to convey information. These combinations exhibit complex statistical regularities whose role in reading is hotly debated in the psycholinguistic community. This theoretical problem could be approached using the framework of Statistical Learning, but most work done so far in visual statistical learning use sequential presentation of stimuli, leaving reading-like material rather unexplored. Therefore we have developed a statistical learning paradigm in which stimuli are novel words written in pseudofont for which all characters are presented simultaneously. This experimental design allowed us to orthogonally manipulate word frequency and bigram frequency. Our results show that, across different string lengths, when facing a word learning task, participants implicitly learned bigram frequencies and had difficulties rejecting incorrect words that shared bigram frequencies with correct words. This implies that participants used the learned statistical structure of the stimuli to make lexical decisions. The novel experimental design we present here can easily be extended to other fields beyond reading in which the relevant statistics for learning depend on co-occurrence of features instead of sequential presentation.