Recommendation systems are an important part of our daily lives, but they can struggle in cold-start scenarios, when there are no historical transaction data for items and users. This study addresses such cold-start scenarios. Our proposed two-phase solution involves first, identifying for each new item its most similar existing items using various item embeddings (image, textual-description, and attributes-based), and second, recommending frequently co-purchased items, alongside an augmented nearest neighbor collaborative filtering method. The proposed solution was evaluated using data from H&M's online store. Our preliminary outcomes highlight the effectiveness of textual-description embeddings for generating valuable recommendations under cold-start conditions. Notably, using textual embeddings, we accurately recommend at least one item among the top co-purchased items for 20% of new items, achieving substantial performance improvements over baseline strategies. This study contributes a versatile solution requiring no retraining of models and catering to scenarios with limited historical activity on websites.