IEEE ICEBE 2025 — LLM Agent Simulation of Consumer Response to a Discount Promotion

Summary:

This paper is about using large language model agents to act like shoppers in a fake town, so marketers can test promotions before trying them in the real world.

The authors built a small virtual world where agents work, eat, talk, remember things, and decide where to spend money.

The main test was simple: one fried chicken shop ran a 20% discount in the middle of the week, and the researchers watched what happened.

The work was done by researchers from Clark University and Quinnipiac University.

The setup had 11 agents and 10 locations in a virtual town.

Agents were given personas, budgets, energy levels, grocery needs, memories, and social ties. A DeepSeek-V3 model generated their plans, conversations, and buying decisions in JSON. The system then ran those choices in parallel and tracked where agents went, who they met, what they bought, and how sales changed over a week. The paper also compares the simulation patterns with older marketing research, but it does not validate the model against a fresh real-world dataset.

The main results were that the discounted fried chicken shop's revenue rose 51% from Day 2 to Day 3, even after the price cut, and its market share rose from 30% on Day 1 to 41% on Day 3.

At the same time, the local diner lost share and saw revenue fall 7%, while the coffee shop stayed fairly steady, suggesting the discount mostly pulled people away from one competitor rather than growing the whole market.

The strongest effect peaked with a delay, which the authors read as word spreading through the agent network. Total market spending moved around but did not clearly rise during the promotion, which again points to substitution, not extra demand.

The simulation also showed repeat visits and rough loyalty patterns: some agents returned to the fried chicken shop during the promotion, while the diner kept broad appeal without discounts.

On top of that, agents made group plans, invited friends, and spread interest through conversations, creating social coordination that was not hard-coded step by step.

There are important catches. This is a model world built by the researchers, not an experiment on real consumers, so the findings are best seen as simulated outputs, not evidence that real shoppers would react the same way.

The sample is tiny at 11 agents, and their behavior depends heavily on prompts, the chosen LLM, and design choices like the town layout and rules.

The paper itself says the model can hallucinate nonexistent places or items, and that results may change across LLMs. It also notes weak realism for children and older adults.