Social media has transformed into a primary communication and marketing platform where content engagement plays a crucial role in user influence and business strategy. Traditional analytics tools provide insights only after content is posted, limiting users' ability to optimize engagement beforehand. This study presents a novel AI-based framework for predicting social media engagement before posting, using BERT and DistilBERT for textual analysis. By leveraging transformer-based embeddings, the approach analyzes past social media data, extracts meaningful textual patterns, and forecasts potential user engagement (likes, shares, comments). The proposed model is trained using historical social media datasets, ensuring robust prediction across various post types. The dataset includes captions, engagement metrics, and user interactions, preprocessed through text normalization, tokenization, and embeddings generation. Experimental results demonstrate that the model outperforms traditional NLP models, offering high accuracy in engagement prediction. The proposed approach enables content creators, marketers, and businesses to strategically optimize their posts before sharing, leading to higher engagement and better audience reach. The findings highlight the effectiveness of transformer models in predicting user behavior, paving the way for AI-driven content optimization in social media analytics.