Can AI support market research with fewer survey responses?
When researchers reanalysed existing pricing survey data, could AI support estimates based on fewer human responses? Using AI to help analyse human answers produced 12.4% estimation error with 100 responses, compared with 12.5% using 300 human responses alone. This suggests that pricing research may be able to reach similar accuracy with fewer responses.
ACADEMIC WORK REVIEWED
Generative Augmented Inference of LLM-generated Data for Market Research: Theory and Empirical Evidence
ResearchersCheng Lu, Mengxin Wang, Dennis J. Zhang, Heng Zhang
arXiv · 22 July 2026 · v3, not peer-reviewed · Preprint · Source language: English
English review: 1 November 2026 · The source date appears in the citation above.
Editorial review draft
The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.
LET'S READ THE RESEARCH TOGETHER
We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.
01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?
Can AI reliably improve market estimates when only a small number of real human responses is available?
The researchers investigate whether AI outputs can reduce estimation error when human data are limited. Their aim includes assessing uncertainty correctly, alongside producing more accurate estimates. The proposed GAI framework uses AI outputs as information that helps predict human responses.
Explaining the context · DijitalPi commentary
DijitalPi commentary: A large response count in a pricing or product study does not establish that customers are represented accurately. How AI-generated answers enter the analysis matters. This paper clarifies the practical distinction between replacing real responses with synthetic answers and using AI information to support the analysis of human data.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
In the v3 manuscript dated 22 July 2026, Cheng Lu, Mengxin Wang, Dennis J. Zhang and Heng Zhang present a theoretical method and experiments that repeatedly select different samples from three existing datasets. They collect no new data. The first dataset comes from a 2020 study of US adults’ vaccine preferences: 1,971 respondents each complete 5 tasks choosing between two hypothetical vaccines with 11 attributes. In the second dataset, 2,058 participants answer more than 500 questions and indicate whether they would buy each of 40 products offered at randomised prices. The main text reports a household cleaning product. The country for this dataset is unspecified. The third application uses observational census data on 318,215 people in California to examine the relationship between private health insurance coverage and income. The supplied source file does not specify data collection periods.
Understanding the method · DijitalPi commentary
The researchers retain a small subset of human responses for analysis and use AI outputs in a larger auxiliary sample. Human sample sizes are 50-200 for vaccines, 100-500 for pricing and 100-1,000 for insurance. Auxiliary samples are 1,000, 1,000 and 2,000 respectively. Each configuration is repeated in 50 independent trials. The comparison methods include Primary, which uses human data alone, and Naive, which pools human and AI responses without correction. They also include PPI and another method that account for errors in AI predictions. The researchers compare results with a coefficient calculated from all the human responses. A coefficient is a numerical value describing the relationship being studied. This comparison value is not an independently known market fact.
The vaccine application uses GPT-4o choice predictions and explanations, with OpenAI’s text-embedding-3-large model converting the explanation text into a numerical representation. In pricing, AI representations built from participants’ answers, called digital twins, generate yes/no predictions. Price is the demand model’s only explanatory variable. The insurance application uses a conventional machine learning classifier rather than a language model. The main measure compares coefficients estimated from a small sample with those calculated from all the human responses. It is the average percentage difference between them. Lower error is better. The researchers also examine ranges that express uncertainty at the 95% level. They check how often these ranges contain the comparison value and how wide they are. They also count errors in the estimated direction of a relationship and cases where an existing relationship is incorrectly judged statistically insignificant. The estimation error denominator includes an added 1 for vaccines and pricing but not insurance, so error percentages across applications should not be read on the same scale.
03 / RESEARCH FINDINGS
Using AI as supporting information produced low estimation error with fewer human responses
In vaccine preferences, the researchers found 16.5% estimation error for GAI with 50 human responses, compared with 19.0% for the human-only method with 200 responses. They interpreted this as a reduction of more than 75% in required human responses. Directly pooling AI and human responses produced 40-49% estimation error. For GAI using numerical representations of explanations, uncertainty ranges at the 95% level contained the reference value in 94.6-99.5% of trials. For GAI using choice predictions alone, this fell to 90.9% with 200 human responses. In pricing, digital twins predicted a positive purchase response in 30.07% of cases, compared with 44.10% positive human survey responses. These are not completed sales rates. Given the same auxiliary predictions, GAI achieved 12.4% estimation error with 100 human responses, while the human-only method achieved 12.5% with 300. The authors reported a 67% reduction in human responses. However, another comparison method produced slightly narrower uncertainty ranges in this application, so GAI did not lead on every measure. In insurance, GAI’s uncertainty ranges contained the reference value in 99-100% of trials and no decision errors were observed across the tested configurations. These results come from resampling existing data, rather than realised gains in a commercial deployment.
GAI with 50 human responses for vaccine preferences
16.5% estimation errorCompared with 19.0% estimation error for the human-only method using 200 responses. The outcome is estimation error in model coefficients.
GAI with 100 human responses for pricing
12.4% estimation errorThe human-only method achieves 12.5% estimation error with 300 responses. This comparison does not measure sales or revenue growth.
DijitalPi commentary: Ask your research provider whether AI-generated answers count as real customer responses or serve as supporting information when analysing human data. Retain a human-only result as the comparator in a small evaluation. When deciding, check whether responses were randomly selected, how large the estimation error is and how uncertain the result is. Do not focus only on reducing the number of survey responses.
The authors’ theoretical result that the method cannot perform worse than human data alone depends on random selection of human responses. Deliberately selecting easy cases can invalidate it. Increasing auxiliary AI data without limit can also be counterproductive. Different GAI inputs do not reflect uncertainty equally accurately, and integration into real-time systems may create computational challenges. The method therefore needs specialist assessment and checks in the target market. It does not supply a ready-made survey reduction rate.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
This study is a preprint, a research manuscript that has not yet undergone peer review. It did not measure sales, revenue or realised research budget savings in a new business deployment. The study shows that retaining human data while drawing on AI outputs can reduce estimation error in existing datasets. The findings do not come from a new business field deployment: vaccine data concern US adults, insurance data come from California, the pricing dataset’s country is unspecified, and no validation is provided for Türkiye or Turkish. The reviewed v3 is a preprint; acceptance of an earlier, narrower version at ICML does not make this expanded manuscript peer-reviewed. All authors are affiliated with US university business schools. No funding or competing-interest statement was found in the source, and no company affiliation was specified.
Our open question at DijitalPi: could you retain randomly selected human responses in your own pricing or product research and measure how AI assistance changes estimation error and uncertainty?
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
E-commerce: researching a cleaning product’s price
An illustrative e-commerce business wants to understand customers’ willingness to buy a cleaning product at different prices.
Could AI predictions let us ask fewer customers?
Open the recommendation for example 01+
Retain real responses and evaluate AI predictions as separate supporting information
Randomly select customers for human responses and record the displayed price alongside each answer. With a statistician, compare estimation error and uncertainty between a human-only method and GAI. Do not count stated willingness to buy as completed orders or transfer the paper’s sample reductions into expected budget savings.
EXAMPLE 02
B2B software: comparing package features
An illustrative B2B software business is researching preferences between packages combining different features and prices.
Could AI-generated explanations help estimate package preferences?
Open the recommendation for example 02+
Test explanations as supporting information for predicting human choices without treating them as customer opinions
Randomly select records for collecting human responses rather than choosing only accessible customers or predictable decisions. With specialist support, evaluate AI explanations for the same package options as inputs for estimating human preferences. The paper’s attribute-choice application concerns vaccine preferences. Its findings have not been validated for B2B software buyers.
EXAMPLE 03
Tourism: choosing how much supporting data to use
An illustrative tourism business plans to supplement limited human responses about holiday packages with a large number of AI predictions.
Do more AI responses always produce a more reliable estimate?
Open the recommendation for example 03+
Compare different quantities of supporting data instead of increasing them without limit
Plan analyses using different quantities of AI outputs while retaining randomly selected human responses. Record estimation error and uncertainty for package preferences in each analysis. The authors’ warning about excessive supporting data motivates this comparison. The study does not establish a suitable quantity for tourism.
TRY IT WITH YOUR TEAM
Define the role of human responses
Choose one pricing or product preference question and plan a random human sample alongside a human-only comparison. Evaluate GAI with a statistics specialist. Measure estimation error and how uncertain the estimate is. This is a proposed pilot that DijitalPi has not conducted.
Cheng Lu, Mengxin Wang, Dennis J. Zhang, Heng Zhang Olin Business School, Washington University in St. Louis, USA (Cheng Lu, Dennis J. Zhang); Naveen Jindal School of Management, University of Texas at Dallas, USA (Mengxin Wang); W. P. Carey School of Business, Arizona State University, USA (Heng Zhang).
The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.
Cookie preferences