And what if brands could survey their customers without making them come back?

Listen Labs creates digital twins from real interviews to simulate consumer reactions to products, prices, and campaigns.

Listen Twins Creates Digital Replicas of Real Consumers to Test Brand Decisions

Before changing a price, launching a campaign, or introducing a new product, companies typically want to understand how their customers will respond. Traditional research may require several days of recruitment and interviews, making it difficult to investigate every decision marketing and product teams face. Listen Labs proposes a different approach with Listen Twins, a system that creates digital representations of people who have actually been interviewed. Each model corresponds to one respondent and draws on their answers, motivations, and the information collected during an in-depth conversation. Companies can then present proposals to a panel of more than 100 digital twins and receive individual responses within seconds, accompanied by explanations and references to the original interviews. One Digital Twin for Every Interview Participant

Market research has long relied on personas to represent consumer segments. Built from observations, interviews, and statistical information, these profiles summarize the characteristics of a target audience. They can also conceal important differences between people who share similar demographic backgrounds or purchasing habits.

Listen Twins uses individual representations instead. Rather than prompting a language model to impersonate an average customer, the platform builds a separate digital twin for each research participant. Two people in the same demographic category can therefore respond differently, reflecting the preferences and experiences expressed in their respective interviews.

The approach builds on Listen Labs' existing business in AI-moderated qualitative research. The company says its recruiting network includes more than 50 million verified participants. For each study, it selects respondents according to a client's criteria and conducts detailed interviews about their motivations, stated behaviors, and decision-making priorities.

Those conversations provide the foundation for each digital twin. The models can also incorporate customer information, previous research, and context specific to the company commissioning the study. According to Listen Labs, twins are created for particular purposes and are not shared across customers. Testing Prices, Campaigns, and Product Concepts

Companies can present the twins with product descriptions, pricing options, advertising messages, brand names, concepts, or complete campaigns. Each model produces an individual response, which the platform then organizes to reveal patterns and disagreements.

For example, a company could compare two positioning statements before launching a campaign. Some twins might favor a message emphasizing quality and reliability, while others respond more positively to price, convenience, or guarantees. The team can examine these differences across segments rather than relying solely on an average preference.

The same approach can be applied to product launches. A demonstration video can be shown to the panel to identify which elements attract interest, what objections emerge, and whether the message appears persuasive enough to encourage further action. Listen Labs presents these simulations as tools for campaign development, brand messaging, and product decision-making.

The responses remain predictions of stated reactions based on previous interviews. Someone who says they would book a demonstration or purchase a product may not actually do so. The system does not directly measure future sales or conversion outcomes. Every Answer Can Be Traced to an Original Interview

Traceability is one of Listen Twins' central features. Instead of displaying only a model-generated answer, the platform allows users to inspect the twin's explanation and locate quotations from the real person's interview that support its response.

This helps researchers understand whether a reaction comes from an explicitly stated preference. If a participant emphasized after-sales support during an interview, their twin might refer to that concern when evaluating a new product offer. The researcher can then examine the original quotation and determine whether the interpretation is relevant.

Each answer also includes a confidence score. When the available interview material does not provide enough evidence, the system can express uncertainty and respond that it does not know.

These mechanisms make the output easier to inspect than a generic AI persona's response. They do not guarantee that the model perfectly represents the individual's intentions. A quotation may be relevant in one context but much less informative when the question concerns a product or situation never discussed during the interview. Listen Labs Reports 93% Agreement With Human Answers

Listen Labs claims an average accuracy of 93% when comparing twin responses with the answers of the people they represent. The company describes a validation procedure in which part of the original interview is withheld during twin creation and later used to compare predictions with the respondent's actual answers.

The approach tests whether the model can reproduce information and preferences it did not directly receive during construction. It therefore goes beyond simply checking whether generated answers are consistent with the interview material used for training.

The published figure nevertheless requires qualification. The public materials reviewed do not provide enough information about panel composition, question categories, the exact accuracy calculation, or confidence intervals to independently reproduce the result. The founder's announcement also describes distribution accuracy, while the product materials discuss agreement with individual human responses.

More importantly, reproducing withheld interview answers is not equivalent to predicting future consumer behavior with the same accuracy. The 93% figure should not be interpreted as a probability of correctly forecasting