An LLM agent simulation can be used to pre-test marketing strategies, with early studies reproducing effects already known from real cases
Strategy review & optimization
- Reader Note:
- Eleven LLM agents in a ten-location virtual town over one simulated week, run on DeepSeek-V3. There are no real consumers anywhere in this study. The authors establish validity by noting that the simulation reproduced a substitution pattern already documented in the discount literature, not by comparing it against a fresh real-world dataset, so the simulated figures — revenue up 51% from Day 2 to Day 3, share moving from 30% to 41% while a competitor fell from 62% to 48% — describe the model and not a market. The paper also records that agents sometimes chose locations that did not exist in the town, and that the personas were least realistic for children and older adults.
- Evidence:
- Modelled / projected
- Vendor interest:
- None
- Sample:
- 11 LLM-based generative agents with assigned personas, budgets, needs, and social ties operating in a virtual town with 10 locations; not real human participants.
Good for
Planning with a rough expectation, and sanity-checking your own assumptions.
Not for
- Reporting as what happened. This is what a model expects, not what anyone counted.
- Nobody counted this directly — it's a calculation. The publisher worked it out itself, or took the ad platform's word for what its ads achieved.




