The electron temperature of the plasma is one important aspect of the environment. Electrons created by photoionization or impact ionization of atmospheric gas have energies $\sim$10 eV. In an active comet coma, the gas density is high enough for rapid cooling of the electron gas to the neutral gas temperature (a few hundred kelvin). How cooling evolves in less active comets has not been studied before. Aims. We aim to investigate how electron cooling varied as comet 67P/Churyumov-Gerasimenko changed its activity by three orders of magnitude during the Rosetta mission. We used in situ data from the Rosetta plasma and neutral gas sensors. By combining Langmuir probe bias voltage sweeps and mutual impedance probe measurements, we determined at which time cold electrons formed at least 25\% of the total electron density. We compared the results to what is expected from simple models of electron cooling, using the observed neutral gas density as input. We demonstrate that the slope of the Langmuir probe sweep can be used as a proxy for the presence of cold electrons. We show statistics of cold electron observations over the two-year mission period. We find cold electrons at lower activity than expected by a simple model based on free radial expansion and continuous loss of electron energy. Cold electrons are seen mainly when the gas density indicates that an exobase may have formed. Collisional cooling of electrons following a radial outward path is not sufficient to explain the observations. We suggest that the ambipolar electric field keeps electrons in the inner coma for a much longer time, giving them time to dissipate energy by collisions with the neutrals. We conclude that better models are required to describe the plasma environment of comets. They need to include at least two populations of electrons and the ambipolar field.