Most pricing and insights teams have been asked some version of this: "Instead of paying for a panel and waiting for fieldwork, why not ask an AI model to play 1,000 customers and answer the conjoint survey?"
"Synthetic respondents" are an appealing idea. If a large language model (LLM) has read most of the internet, perhaps it already knows how people trade off brand, features and price.
Since 2023, marketing researchers have tested this directly. Their findings point the same way: LLMs are useful for exploring ideas and preparing a study, but they are not a reliable substitute for real respondents when the output is a price.
Right direction, wrong numbers
Brand, Israeli and Ngwe's Harvard Business School working paper, Using LLMs for Market Research (latest version April 2026), tested whether GPT models can stand in for consumers in conjoint-style studies. The models often reproduce familiar patterns, such as demand falling as price rises. But the willingness to pay they produce is "often inaccurate and in some cases wrong-signed", it does not reflect real differences between customer segments, and fine-tuning helped only within product categories where human data already existed.
Changing the price changes more than the price
Gui and Toubia (2023) looked at a basic problem with simulated experiments. In a real experiment, you change one thing (the price) and hold everything else constant. When you change the price in a prompt, the LLM may also change what it assumes about the product, the shopper or the context. That confounds the demand curve: you no longer know what caused the change in the answer.
Plausible is not the same as correct
Oetzel and Maiberger (2026), How price sensitive is AI?, compared LLM willingness to pay with human data. LLM answers "can appear economically plausible", yet led to materially different optimal prices; the authors conclude synthetic respondents "should not be treated as direct substitutes".
Wang, Zhang and Zhang (2024) found that substituting LLM data for human responses in conjoint "can exacerbate bias". Combining a small human sample with LLM data, on the other hand, saved 24.9% to 79.8% of the data and cost needed.
Arora, Chakraborty and Nishimura (2025), AI–Human Hybrids for Marketing Research, found that LLMs get the direction and tone of responses right, and that a human-plus-LLM combination beats either on its own.
Why price is where LLMs struggle most
Three reasons stand out:
No budget. An LLM has no income and no bills. When it "chooses" between a £39 and a £79 plan, it is predicting likely text, not making a trade-off against money it has to spend.
New products aren't in the training data. A new plan, a new pack format or this year's price change has no history for the model to draw on.
Everyone sounds the same. Real markets contain bargain hunters, loyalists and sceptics. LLM answers drift towards a stereotypical average, which flattens segments and narrows the spread of preferences that simulator intervals depend on.
How Ordeen uses AI
Ordeen's rule is verified data first, AI last. We draw a hard line between preparing a study, where AI saves hours, and respondent data, where only real people count.
Where we use AI: recommending a method from your brief; drafting attributes, levels and a starting competitor market for you to edit; checking open-ended answers for fraud; and suggesting names and descriptions for segments that have already been calculated.
Where we never use AI: every choice task, Van Westendorp price and Gabor-Granger answer comes from real respondents, through the integrated panel or your own customer list. Fieldwork actually uses AI the other way round: to catch and reject respondents who use an LLM to write their open-ended answers.
Related reading: Methodology · What is willingness to pay? · What is conjoint analysis?
Sources
Arora, N., Chakraborty, I., & Nishimura, Y. (2025). AI–Human Hybrids for Marketing Research: Leveraging Large Language Models (LLMs) as Collaborators. Journal of Marketing. DOI 10.1177/00222429241276529.
Brand, J., Israeli, A., & Ngwe, D. Using LLMs for Market Research. Harvard Business School Working Paper 23-062 (version of April 2026). SSRN DOI 10.2139/ssrn.4395751.
Gui, G., & Toubia, O. (2023). The Challenge of Using LLMs to Simulate Human Behavior: A Causal Inference Perspective. arXiv:2312.15524.
Oetzel & Maiberger (2026). How price sensitive is AI? Journal of Marketing Analytics. DOI 10.1057/s41270-026-00542-7.
Wang, Zhang & Zhang (2024). arXiv:2412.19363.
