Ensuring similarity of user interfaces (UI) is often desirable, e.g. in software migration and redesign projects, to minimize experience disruption for regular users and increase subjective satisfaction with new versions. In our paper we explore applicability of artificial neural networks (ANNs) to support test-driven development by predicting similarity assessments without employing the actual users. Having reviewed requirements engineering (RE) standards and practices for HCI-related requirements, we identified two dimensions for similarity of web UIs: 1) objective, the data for which we collected with a dedicated web intelligence miner and 2) user-subjective, operationalized with the renowned Kansei Engineering method. Then we constructed the respective ANN models predicting perceived similarity between websites of a same domain and trained the models with the data we collected in experimental sessions with 209 participants of different nationalities and 21 operational university websites. The results of our pilot study suggest that subjective "emotional" factors are considerably more important in predicting similarity evaluations provided by users. Thus, employment of trained ANNs as test oracles may be feasible in automated measurement and control of UI similarity.