Two opinions currently exist on the role of zero-filling in data processing for in vivo MRS: that it results in a purely cosmetic interpolation; or that it confers a benefit. Most commonly, in vivo MRS data are acquired as complex time-domain half-echoes, that are Fourier transformed to the give a real spectrum that is modeled for quantification. In this manuscript, we highlight that performing zero-filling draws the independent information from the imaginary part of the spectrum into the real spectrum, improving modeling accuracy. In order to demonstrate this, 10,000 time-domain datasets were simulated as decaying exponentials and noise was added. Data were then Fourier transformed with no-zero filling, 2x, 4x and 8x zero-filling. All spectra were then modeled using a simple single-Lorentzian frequency-domain model. It was demonstrated that 2x zero-filling results in a ~√2 benefit in modeling accuracy, compared to no zero-filling. There was no additional advantage for further zero-filling.