Given the vast array of products on major online marketplaces, finding a specific item on the first few search attempts is unlikely. Users often see search results with prompts like “perhaps you meant this?”, which can lead them to reconsider their search criteria. These prompts are crucial in determining whether a user continues towards purchasing or abandons the search. Negative motivations are often caused by a search prompt that leads to zero search results. Zero search results, ranging from 6% to 8%, are a significant issue for online marketplaces. These occur when the search engine misinterprets the user’s intention and fails to locate relevant items from its extensive inventory. Search prompts must align with user intent, but narrowing down results may not exactly match the user’s needs. For instance, a slight variation in a model name for electronic device spare parts can lead to irrelevant results. Most methods for measuring text similarity are inaccurate and produce unsatisfactory results. We developed and tested machine learning models based on the discriminator-generator approach to accurately identify similar queries. These models use discrimination stages to generate and select appropriate search prompts. Our approach involves six discrimination stages: Narrowing, Rephrasing, Expansion, Modification of characteristics, Substitution, and Irrelevant. Experiments showed that our search prompts exceeded 90% on the F1 score. Our approach allows us to separate the inference model from business logic, which has a positive impact on the range of possible applications in marketplace search systems.