Microeconomic data show two important facts about new products. First, some products are more important than others. And second, it takes them years to penetrate the market significantly. Our calibrated model with these features overpredicts the autocovariance of U.S. GNP at long lags, but underpredicts it at short lags. The latter is not surprising, since the model leaves out other obvious high-frequency shocks. The puzzle is why the U.S. GNP data do not show stronger autocorrelation at higher lags. A surprising finding is that while the speed of diffusion has huge level effects, it plays a minor role in shaping the business cycle.