Over the last decade, nova shells have been discovered around a small number of cataclysmic variables that had not been known to be post-novae, while other searches around much larger samples have been mostly unsuccessful. This raises the question about how long such shells are detectable after the eruption and whether this time limit depends on the characteristics of the nova. So far, there has been only one comprehensive study of the luminosity evolution of nova shells, undertaken almost two decades ago. Here, we present a re-analysis of the Hα and [O III] flux data from that study, determining the luminosities while also taking into account newly available distances and extinction values, and including additional luminosity data of “ancient” nova shells. We compare the long-term behaviour with respect to nova speed class and light curve type. We find that, in general, the luminosity as a function of time can be described as consisting of three phases: an initial shallow logarithmic decline or constant behaviour, followed by a logarithmic main decline phase, with a possible return to a shallow decline or constancy at very late stages. The luminosity evolution in the first two phases is likely to be dominated by the expansion of the shell and the corresponding changes in volume and density, while for the older nova shells, the interaction with the interstellar medium comes into play. The slope of the main decline is very similar for almost all groups for a given emission line, but it is significantly steeper for [O III], compared to Hα, which we attribute to the more efficient cooling provided by the forbidden lines. The recurrent novae are among the notable exceptions, along with the plateau light curve type novae and the nova V838 Her. We speculate that this is due to the presence of denser material, possibly in the form of remnants from previous nova eruptions, or of planetary nebulae, which might also explain some of the brighter ancient nova shells. While there is no significant difference in the formal quality of the fits to the decline when grouped according to light curve type or to speed class, the former presents less systematic scatter. It is also found to be advantageous in identifying points that would otherwise distort the general behaviour. As a by-product of our study, we revised the identification of all novae included in our investigation with sources in the Gaia Data Release 2 catalogue.