Instead of asking customers about a new advertisement, we can ask AI to think and answer like them. Can we trust those answers? This study shows that generated responses can resemble real people on average while missing important differences between individuals. Our marketing recommendation at DijitalPi is to use these responses to prepare ideas and interview questions, then test messages with actual customers. An AI response saying “I would buy this” is not evidence that a sale will happen.
ACADEMIC WORK REVIEWED
Synthetic Data in Marketing Research: How to Evaluate and When to Trust
ResearchersOded Netzer, Rajan Sambandam.
arXiv · 12 September 2026 · v1 · Preprint · Source language: English
English review: 21 September 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?
Which decisions can simulated answers support?
The authors examine when synthetic customer data can be trusted.
Explaining the context · DijitalPi commentary
DijitalPi commentary: Finding objections to a campaign idea and forecasting market demand are different tasks. A generated customer conversation may help create questions, but your decision record should clearly identify which statements came from actual people.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
A diagnostic for digital-twin answers is tested on 108 attitude questions from a 3,063-person survey.
Understanding the method · DijitalPi commentary
Write down the intended decision before evaluating a research output. Discovering possible misunderstandings in ad copy requires different validation from forecasting orders. One overall accuracy score cannot automatically approve both uses.
03 / RESEARCH FINDINGS
A matching average is not enough.
Aggregate similarity can hide poor individual differentiation; the paper proposes selective evaluation.
DijitalPi commentary: Label AI answers used during question preparation as assumptions to validate, rather than customer findings.
Preserve responses from real interviews that contradict the simulation. An unexpected counterexample may be more useful for a brand decision than agreement around an average. This is our proposed editorial application.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
The preprint does not demonstrate complete replacement of real customer research.
Our open question: which campaign assumptions change when checked with actual customers? We provide neither a sales forecast nor a completed DijitalPi experiment.
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
Ad copy: liking or understanding?
An AI persona praises three headlines, but it is unclear which communicates the product correctly.
Are positive comments enough to select a winner?
Open the recommendation for example 01+
Ask real customers what they understood.
Use AI to suggest possible misunderstandings. Show headlines to the intended audience under comparable conditions, then ask people to describe the offer in their own words. Record misunderstood promises, recall and intended next steps separately.
EXAMPLE 02
Product launch: simulated demand
Virtual customers say they would buy a new package. The team wants to turn this into a forecast.
Is generated purchase intent a sales forecast?
Open the recommendation for example 02+
Run a small real-world interest test first.
Use generated answers to prepare objections and interview questions. Present a clearly defined offer to actual people and track information requests and qualified conversations. Never enter synthetic answers into the CRM as genuine leads.
EXAMPLE 03
B2B segments: everyone agrees
Small-business and enterprise personas seem to care about the same features.
Are the segments similar, or are the questions weak?
Open the recommendation for example 03+
Test the distinguishing assumption with real interviews.
Ask separately about procurement approval, integration and support. Record counterexamples from actual customers. Preserve disagreements between AI output and customer feedback instead of reporting only a matching average.
TRY IT WITH YOUR TEAM
Compare a campaign assumption with actual customers.
Keep AI comments, team assumptions and real customer answers in separate columns. Decide in advance what evidence would change your decision. State the pilot sample limits; a few positive interviews do not represent the whole market.
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