Although lipid-dependent protein clustering in biomembranes mediates numerous functions, there is little consensus among membrane models on cluster organization or size. Here, we use influenza viral envelope protein hemagglutinin (HA(0)) to test the hypothesis that clustering results from proteins partitioning into preexisting, fluid-ordered "raft" domains, wherein they have a random distribution. Japan HA(0) expressed in fibroblasts was visualized by electron microscopy using immunogold labeling and probed by fluorescence resonance energy transfer (FRET). Labeled HA coincided with electron-dense, often noncircular membrane patches. Poisson and K-test (Ripley, B.D. 1977. J. R. Stat. Soc. Ser. B. 39:172-212) analyses reveal clustering on accessible length scales (20-900 nm). Membrane treatments with methyl-beta-cyclodextrin and glycosphingolipid synthesis inhibitors did not abolish clusters but did alter their pattern, especially at the shortest lengths, as was corroborated by changes in FRET efficiency. The magnitude and density dependence of the measured FRET efficiency also indicated a nonrandom distribution on molecular length scales (approximately 6-7 nm). This work rules out the tested hypothesis for HA over the accessible length scales, yet shows clearly how the spatial distribution of HA depends on lipid composition.