The yeast GCN4 transcriptional activator protein binds as a dimer to a dyad-symmetric sequence, indicative of a protein-DNA complex in which two protein monomers interact with adjacent half-sites. However, the optimal GCN4 recognition site, ATGA(C/G)TCAT, is inherently asymmetric because it contains an odd number of base pairs and because mutation of the central C.G base pair strongly reduces specific DNA binding. From this asymmetry, we suggested previously that GCN4 interacts with nonequivalent and possibly overlapping half-sites (ATGAC and ATGAG) that have different affinities. Here, we examine the nature of GCN4 half-sites by creating symmetrical derivatives of the optimal GCN4 binding sequence that delete or insert a single base pair at the center of the site. In vitro, GCN4 bound efficiently to the sequence ATGACGTCAT, whereas it failed to bind to ATGAGCTCAT or ATGATCAT. These observations strongly suggest that (i) GCN4 specifically recognizes the central base pair, (ii) the optimal half-site for GCN4 binding is ATGAC, not ATGAG, and (iii) GCN4 is a surprisingly flexible protein that can accommodate the insertion of a single base pair in the center of its compact binding site. The ATGACGTCAT sequence strongly resembles sites bound by the yeast and mammalian ATF/CREB family of proteins, suggesting that GCN4 and the ATF/CREB proteins recognize similar half-sites but have different spacing requirements. Unexpectedly, in the context of the his3 promoter, the ATGACGTCAT derivative reduced transcription below the basal level in a GCN4-independent manner, presumably reflecting DNA binding by a distinct ATF/CREB-like repressor protein. In other promoter contexts, however, the same site acted as a weak upstream activating sequence.