GPT for Research - James Brand
Using GPT for Market Research
Summary
Brand, Israeli, and Ngwe (2023) have made a pioneer study on utilizing GPT 3.5 to the market research on people’s willingness-to-pay preferences. Instead of directly asking GPT, the authors used the “distributional nature” of the model’s responses. The temperature of GPT’s text completion was set as 1. Each prompt was repeated hundreds of times so that GPT can generate a “distributional” reply instead of a single reply.
Methodology
The authors have used Julia Code to process the questionnaires. The conversation with GPT was performed using OpenAI API. Prompts are in Appendix B, and respond examples are in Appendix C. There are 2 groups of studies:
- Testing predictions from theory - to find if GPT reacts same as economical theories.
- WTP comparison with real humans - to find if GPT has same react as customer behavior.
Set-up of Customer Identities and Goods
Before the purchasing prompts, there should be a set-up process of customer identities and goods. The authors generated functions for a prompt of purchasing decision, with its arguments being the customer annual income and product price.
Study 1 - Downward-sloping demand curve
This study was performed with 1 option and 2 options. Regardless of the option amount, the price ranges from low to high. GPT has generated a reasonable response: although both cases have downward-sloping demand curve, the one with 2 options is more obvious, especially at the price where the reference good. This follows human nature: when there is comparison, people’s decisions tend to be anchored.
Study 2 - Impact of Income and Demand
Economic theory suggests that higher-income customers are less sensitive to price changes. This is also proved by the simulation with GPT.
Study 3 - Previous Purchase Dependence
Usually, there is an inertia for a customer to stay with original selection to prevent unnecessary changes. This is proved by the GPT simulation as well.
Study 4 - Diminishing marginal utility
Diminishing marginal utility is a natural phenomenon and is also reproduced by GPT.
Results
- OpenAI API can help generate automatic workflow in customer preference generation and data collection.
- GPT’s simulation is proved to be realistic in demand curve, annual income, past purchase, and diminishing marginal utility.
Gaps
- This study targeted GPT 3.5 with reference of GPT 3.0. Since GPT 4.0 is already in wide use, this paper is not the most emerging study.
- The examples are all simple cases. They are simple in two ways: firstly, the customers are simple, only with an anchored preference, or with a pre-set annual income; secondly, the dilemmas are simple, either one product with different prices, or to products with different prices and configurations.