Large online trading platforms introduce thousands of new products every day. To purchase a new item, the user has to find it through a search system. Modern search systems primarily rely on user behavior signals like purchases, clicks, and views to index products, which poses a challenge for newly listed items with no existing data, known as the cold start problem. However, generative language models can be trained on historical user interactions to produce simulated search queries for these new items. This generates artificial behavior data, enabling search engines to learn and rank previously unavailable products effectively. The main objective of this article is to test how well the search engine’s autonomous metrics trained on this synthetic dataset for new items perform. The Prod2Query model employs an Encoder-Decoder architecture built on BERT transformer models. By applying Prod2Query to new products, we achieved an mAP@12 score of 77.2%, surpassing the industry average of around 75%. This outcome suggests that the cold-start sales issue can be successfully mitigated by leveraging seller signals to create synthetic search queries and train search models.