The paper shows that large language models can closely replicate human consumer purchase intent when asked to act as synthetic respondents. Traditional numeric rating methods produce unrealistic results, but a new approach called Semantic Similarity Rating (SSR)fixes this by converting model-written text into Likert-scale ratings using embedding similarity. Tested on 57 real product surveys with over 9,000 participants, SSR achieved about 90% of human reliability and realistic response distributions. It also provided richer qualitative feedback than typical human surveys. The method worked best when models were prompted with demographic details like age and income. SSR allows accurate, scalable, and low-cost simulations of consumer behavior, enabling companies to test product ideas synthetically before real surveys, cutting time and cost while maintaining interpretability
