<p>In the digital age, digital watermarking is widely used for copyright protection. This work proposes an audio zero-watermarking scheme that operates in a fractional domain and explicitly examines the processing sequence between time–frequency transform and graph modeling. The framework first applies a fractional Fourier transform (FrFT) with an optimally selected order to map audio frames into a fractional time–frequency domain, then constructs a data-adaptive graph on the resulting coefficients, and finally uses singular value decomposition (SVD) to summarize the fractional-domain graph representation into a compact fingerprint that is bound to the watermark without modifying the host audio. Within a unified zero-watermarking protocol, we further compare transform–then–graph and graph–then–transform designs to examine how constructing the <i>k</i>NN graph in the time domain versus the transform domain affects synchronization robustness. Experiments show that, in the fractional domain, the transform–then–graph configuration achieves lower BER and higher NC than its counterpart in most cases, with especially clear gains under time-scale modification and cropping. These results indicate that building data-adaptive graphs in a suitably chosen fractional domain provides a robust yet transparent foundation for audio zero-watermarking.</p>