An efficient way of modelling web users is needed in the optimisation of an advertising campaign: a lot of different methods have been used, including recurrent neural networks. Using those networks, it is possible to take into account the order of the events in a sequence of browsing. Nevertheless, in order for those to have good enough performance, a consequent amount of data is needed. Because of that, it seems impossible to use them at the beginning of an ad campaign, when not enough data about browsing behaviour have been gathered: and the ad budget is spent randomly in order to gather unbiased data: this is the cold-start problem. To avoid that problem, we suggest leveraging data from previous campaigns, from other announcers, reusing an already trained neural network.