AI synthetic shoppers predict real customer reactions with 90 percent accuracy, study finds
New research shows that large language models can act as synthetic consumers and predict how real people will respond to products with about 90 percent of the accuracy of surveying actual humans. The old way of testing a product concept meant surveying hundreds of people, waiting days or weeks, paying thousands of dollars and still risking biased answers. The new way takes minutes: tell the AI to adopt a detailed persona, such as a 35-year-old mother from Texas earning 50,000 dollars a year, show it the product, and ask whether that person would buy.
The technique behind the accuracy is called semantic similarity rating. Instead of asking the AI to rate a product from one to five, which produces erratic results, researchers ask it to explain its reaction in plain language, then match that explanation against five pre-written reference statements, one per rating level. The method plays to what AI does best: reasoning in words rather than picking numbers.
Validation came from 57 real product surveys covering 9,300 people rating personal care products. The AI matched the distribution of human star ratings, ranked the products in the same order, reproduced demographic patterns by age and income, and often wrote more detailed feedback than the human respondents. Sample reactions read like genuine shoppers: one persona called a body wash too fancy and said it would stick with what it knows, while another weighed ease of use and safety against wanting more proof of effectiveness. Traditional machine-learning models trained on past customer data managed only about 65 percent accuracy by comparison.
The applications for sellers are immediate. Product concepts can be tested before manufacturing by having AI personas from different demographics review each candidate and only producing the winner, whether the target buyers are college students, business professionals, outdoor enthusiasts or parents. Listings can be improved by showing pages to AI customers and rewriting around the objections and concerns they raise. Prices can be tested across income levels to find where purchase intent falls off. And images, packaging, messaging and bundles can all be A/B tested in simulation before a dollar goes to real traffic.
The deeper shift, the researchers argue, is from analyzing what customers did in the past to simulating what they will do in the future, with psychological insight rather than bare metrics. Their conclusion is measured but striking: paired with appropriate methods, large language models act as credible synthetic shoppers for testing concepts, meaning brands can stop guessing at what buyers want before spending on inventory or ads.
In platform news, Walmart and Sam's Club unveiled a strategic tie-up with OpenAI that lets customers complete purchases directly inside ChatGPT conversations, with no visit to Walmart's site or app. Because the AI learns preferences and anticipates needs, retailers using it can intercept purchase intent before shoppers ever reach a marketplace, and the deal extends Walmart's broader AI push into listings, customer service and staff training.
Sellers are also digesting detailed guidance on optimizing for Rufus, Amazon's AI shopping assistant. The core shift is from keywords to storylines: mapping the missions a product serves, such as remote work setups or back-pain relief for an office chair, and making sure the listing works for every pathway a customer might arrive by, whether fitness goals, joint health, meditation or travel workouts for a yoga mat.
The guidance continues: help shoppers complete their mission by spelling out what pairs with the product; answer the questions Rufus needs answered, like whether something suits beginners or fits small spaces; and replace bare specs with context-rich copy, turning a six-millimeter non-slip yoga mat into an eco-friendly, physical-therapist-recommended mat for home practice that pairs with blocks, straps and bolsters. Because Rufus remembers past purchases, first-order quality is critical, and because it consults outside reviews and expert opinion, brands are advised to pursue buying-guide features, expert reviews and awards. Traditional keyword work should continue in parallel, since keywords still tell the assistant what the product is.
Amazon reportedly expects Rufus to generate 700 million dollars in profit in 2025, a sign of how central the assistant has become to its commerce strategy.
