The use of predictive models for evaluating enteric methane emissions from ruminants could be used to assess mitigation strategies on farms. Predictive models of these emissions often use animal and diet characteristics as predictors. As microbes are key players in methanogenesis, we hypothesized that inclusion of microbial data improves existing models’ performances. In this work, we developed linear mixed-effect models with milk composition, animal and diet characteristics data for predicting methane production (g/day), methane yield (g/kg dry matter intake) and methane intensity (g/kg milk) from dairy cattle. We used data from 4 lactating dairy cows fed diets supplemented with concentrates rich in fiber or starch and without or with bicarbonate in a 4x4 Latin-square design. Enteric methane emissions were measured. Microbial taxa from rumen and faeces samples were grouped on the family level to have more consistent variables; row counts were transformed with the centered Log ratio transformation. Root-mean-square prediction error expressed as a percentage of the observation mean (RMSPE%) and concordance correlation coefficient (CCC) were used to compare the models. Including the bacterial (Campylobacteraceae and Prevotellaceae) or archaeal (Group9_sp, Methanocorpusculum_sp, ucl_Euryarchaeota and ucl_Methanomicrobia) families from faeces to methane production model and including the archaeal (Group11_sp, Methanobrevibacter_ruminantium_clade and ucl_Euryarchaeota) families from rumen to methane intensity model strongly improved their respective performances with a considerable decrease in RMSPE% (-56%, -52% and -33% , respectively) and a considerable increase in CCC (+32%, +41% and +54%, respectively). However, microbial taxa had no effect on methane yield. This study highlighted the interest in considering faecal microbial composition in the model most commonly used under field conditions. This unexpected result could be due to the dietary treatments tested. Moreover, rumen archaeal improved methane intensity prediction. These results suggest that microbial data might significantly improve the predictive ability of current models.