Effectiveness evaluation of generative intelligent search: a comparative study with traditional search engine

Purpose The emergence of generative artificial intelligence has transformed information retrieval, yet limited understanding of its differences from traditional search engines hinders clarity regarding changes in user search behavior and variations in search effectiveness, thereby impeding guidance for future search interface design. Design/methodology/approach We conducted a laboratory experiment comparing two versions of the Bing search engine: the traditional Bing search and the generative-AI integrated Bing (Bing Chat). We interpreted variations in user search behavior through querying, browsing, and clicking data, and established a multidimensional metric framework encompassing user, cost-benefit, and system dimensions to assess search effectiveness. 40 participants with good search skills completed the experiment, resulting in a final dataset comprising 160 observation records. Findings The Mann-Whitney U test revealed that users of generative AI search engines exhibited the following characteristics compared to those using traditional search engines: longer queries, more keyboard activity, more frequent query restructuring (including more narrowing queries, more following the system queries, and more changing the syntactic structure queries, with less broadening queries), more browsing during the first search session, and less scrolling and clicking. The results of the Linear Mixed Models show that, compared to traditional search engines, users achieve superior performance with generative AI search on users’ subjective perceptions and search efficiency metrics. Furthermore, we employed machine learning to validate which search behaviors can predict a user’s search effectiveness. The results indicate that the Gradient Boosting Machines are the optimal prediction models. The five most significant indicators for predicting search performance are Avelength. QueryConsTime, TimetofirSearch, ResViewTime and NumbofScroll. Research limitations/implications By revealing differences in user behavior and search efficiency across various search engines, this study deepens our understanding of generative AI search engines at a finer granularity and contributes to building adaptive search platforms capable of optimizing search recommendations. Originality/value By revealing differences in user behavior and efficiency between traditional and generative versions of the Bing search engine, this study deepens our understanding of generative AI search engines at a finer granularity and contributes to building adaptive search platforms capable of optimizing search recommendations.