A new goodness-of-fit test for spectrum sensing in cognitive radios under heavy-tailed noise is proposed, based on the geometric power (also called the zero-order statistics) of the received observations. The noise statistics is assumed to follow a symmetric-alpha-stable distribution, motivated by statistics observed in realistic scenarios. The expressions are provided for the test statistic and the asymptotic detection threshold, in terms of the number of observations under the null hypothesis. Through extensive Monte Carlo simulations, the superior performance of the proposed technique over existing non-linear detection techniques is demonstrated, such as the fractional lower-order statistics, zero-memory non-linear and myriad filtering. In addition, the advantages of the proposed technique on experiment-captured data are demonstrated.