AI "Synthetic Consumers" Can Predict Real Purchase Intent with 90% Accuracy. No Training Required.
How Generative AI Is Revolutionising Purchase Intent Prediction: What Kiwi CMOs Need to Know
A new study has shown that large language models (LLMs) like GPT-4o and Gemini-2.0 can replicate human purchase intent with 90% accuracy, using an out-of-the-box approach that’s fast, scalable, and doesn’t require any custom training or fine-tuning.
What’s the big idea?
Instead of classic, expensive consumer panels, researchers asked LLMs to impersonate consumers with specific demographic profiles, presented them with product images, and prompted them for impressions.
These impressions, rich, free-text responses, were then mapped onto five-point Likert scales via semantic similarity.
How? By comparing the AI’s answers to anchor statements (like “definitely would buy” to “definitely wouldn’t buy”) using advanced text embeddings.
Key findings for marketers:
Realistic Human Simulation: AI-generated responses mirrored real survey results, both in distribution and concept ranking, achieving 90% of human test–retest reliability, massively outperforming classic machine learning methods, which only reached 65%.
Plug-and-play & Cost-effective: No retraining is required. This means you can run wide-scale synthetic consumer panels at a fraction of the traditional cost, testing everything from products to messaging, fast.
Scalable, Rich Insights: The method preserves quantitative results (Likert scores) and offers qualitative feedback, AI consumers explain their ratings, highlight strengths, and call out concerns, offering more actionable commentary than most panel participants.
Demographic Sensitivity: The AI’s responses track with key demographic markers, such as age and income, offering realistic segmentation and nuanced market insights.
Limits & Opportunities: Some demographic details (like gender or region) aren't yet perfectly captured, and synthetic panels are only as good as the LLM’s domain knowledge. For products outside the AI’s training data, human validation remains essential.
Why this matters to New Zealand brands:
For CMOs, this means rethinking expensive market research. LLM-powered synthetic panels can:
Accelerate product validation—from ideation to launch
Reduce costs—screen more concepts, more often
Empower SMEs—allowing smaller players to access critical consumer insights
Complement human research—quickly filter promising ideas, saving time for deeper qualitative work.
My optimistic take:
Generative AI lets marketing leaders focus resources on genuine innovation and deeper engagement, not just routine survey logistics.
Brands can move faster, test bolder ideas, and fuel growth even in challenging markets.
Put this into action:
Here’s how a marketer could use LLMs like ChatGPT to predict purchase intent, based on this research:
Create customer profiles: The marketer defines typical buyer personas—age, income, interests, etc.
Show product images: Product photos or descriptions are presented to the LLM, along with each persona prompt.
Ask for impressions: The LLM (ChatGPT or similar) is asked to impersonate the customer, sharing their thoughts and likelihood of purchase, just like a human panelist would.
Rate the intent: The AI’s response—such as “I’m very likely to buy this”—is mapped to a Likert scale by comparing its language to anchor statements using text similarity.
Quantify purchase intent: This semantic comparison turns qualitative AI responses into quantitative intent scores, directly comparable to human survey data.
Aggregate results: Marketers analyze these AI-generated ratings and comments for trends, demographics, and insights, just as they would with real consumer panels, but faster and at scale.
