Consumer Digital Twin

Why this matters for your decision

Digital twins make existing knowledge about a target group queryable. Instead of searching through last year's study, a team can ask the twin new questions: how does the segment react to this wording? Which objections are to be expected? For exploration and pre-selection, this is fast and inexpensive, especially for categories and brands for which a lot of data is available.

A digital twin reflects what is already known about a target group. For offers that do not yet exist, it lacks this basis. That is where Horizon measures the purchase intent of real people, and nothing is sold. For incremental developments in a familiar setting, the data basis is denser; for a new tariff, a new pricing model or a new category, it is thinner.

How digital twins are built and checked

The data basis usually consists of surveys and interviews. A public research dataset for digital twins covers 2,058 US respondents who answered around 500 questions over four waves (Toubia et al. 2025). Agents based on in-depth interviews with 1,052 people reproduce answers in the General Social Survey 85% as accurately as the people themselves repeat their answers two weeks later (Park et al. 2024).

Both studies show how far the technology has come with survey answers. Both measure against surveys and lab tasks. They do not answer how well a twin reflects the choice between real offers with price and alternatives.

Example

A home appliance manufacturer has a digital twin of its core target group built from customer surveys of the last three years. For a new cordless kitchen machine, the team has the twin rate three positionings; all three receive similarly positive scores. In the subsequent behavioural test with real people, at the same price, the measured purchase intent of the strongest positioning is clearly above that of the weakest. The twin had no data on a product type that did not yet exist in the range.

How it differs

Synthetic respondents are usually created only from demographic profiles, a digital twin additionally from real data. AI personas are often conversational characters without their own data basis. A model trained on observed behaviour rather than on surveys is a different class of method. A Painted Door Test does not simulate a target group, but observes which variant real people choose.

Limitations

A twin is only as good as its data basis and inherits its biases. If that basis comes from surveys, it also inherits the gap between statement and behaviour. It ages with the data, and it knows no reaction to something that did not exist before. The behavioural test, in turn, only checks the variants it is given and explains motives only in part; for the breadth beforehand, a twin is useful.