It depends on your beliefs about the nature of stock returns. Whether or not those beliefs reflect reality is a different matter. I think it's helpful to consider what is being modeled in GARCH by breaking down the term:
Generalized - i.e., can take a number of parameters in order to fit generic data types.
Autoregressive - i.e., terms tend to revert to their means, a-la a Ornstein-Uhlenbeck process (also, Brownian Motion under friction).
Conditional - i.e., future terms are dependent on past terms and/or best estimates of future terms are based on Bayesian inference; reflects stylized beliefs on absolute return/variance clustering.
Heteroskedasticity - i.e., literally, "differing variance"; parameters change over time.
While it is technically possible to use GARCH to model the conditional expectations of stock returns, GARCH models were not intended to model returns.
Implicit in the name is its intent as skedasticity (i.e., volatility) metric in which terms are squared residuals of periodic returns.
Relaxing the squared error condition, however, results in a particular form of Autoregressive Moving Averages (ARMA). ARMAs are a more generic class of econometric model which allow for negative terms and which can conform to many stylized facts of security returns.