Accurate sales forecasting is essential for the mobile phone industry to make informed decisions, such as government policy-making and corporate operational decisions. This study aims to construct a mobile phone sales forecasting model based on brand exposure and explore the predictive role of brand exposure. Firstly, a quantification method for brand exposure, which integrates word-of-mouth reviews and online search data, is proposed. Word2Vec is used to train word vectors in the word-of-mouth corpus to extract initial keywords. Core keywords are then selected through time-lag correlation analysis, and brand exposure is synthesized using principal component analysis(PCA). Secondly, a mobile phone sales forecasting model based on Attention-LSTM and brand exposure is proposed. In this model, LSTM leverages the time-series related features of the data to perform deep learning, which improves the long-term dependency issue inherent in the RNN model and incorporates an Attention mechanism to eliminate data redundancy, thereby enhancing the prediction accuracy. The experiments show that the Attention-LSTM model, incorporating brand exposure, reduces the RMSE and MAPE indicators by 2.18% and 0.89%, respectively. Furthermore, compared to the ARIMA, SVR, BP neural network, and LSTM models, the Attention-LSTM model reduces the average percentage errors by 6.49%, 3.37%, 2.43%, and 0.87%, respectively. The proposed Attention-LSTM model incorporating brand exposure can effectively predict the dynamic changes in mobile phone sales trends.