Digital twins in marketing and market research

What is a digital twin in marketing?

The term comes from manufacturing and aerospace. IBM describes a digital twin as “a virtual representation of a physical object or system that uses real-time data to accurately reflect its real-world counterpart’s behavior, performance and conditions.” The Fraunhofer ASE wiki distinguishes it from simulation models: a digital twin represents a real product that is in operation.

Gartner applies the idea to customers and calls it the “digital twin of a customer” (DToC). According to a 2023 Gartner press release, as reported by DC Velocity, Gartner defines it as “a dynamic, virtual representation of a customer that simulates and learns to emulate and anticipate behavior.” In its Hype Cycle for Emerging Technologies 2024, as quoted by the vendor edgeTI, Gartner adds that customers can be individuals, enterprise customers, personas, groups of people or machines. Gartner’s own pages could not be read when we checked.

In marketing today, the term almost always means a language model that is given data about a person or an audience and then answers questions or reacts to copy.

Digital twin, persona, synthetic persona: the difference

The terms are often mixed up. This is how we tell them apart:

Gartner’s term also covers personas and groups. So when a vendor offers a “digital twin of your audience”, it often means a synthetic persona. Ask what it is based on. More: Synthetic users.

How marketing and market research use digital twins

• Synthetic surveys. A language model answers a questionnaire on behalf of many respondents. Argyle and colleagues (2023) created “silicon samples” this way from the backstories of real survey participants.
• Product and concept tests. Maier and colleagues (Colgate-Palmolive, PyMC Labs) tested 57 personal care product surveys with 9,300 human responses against synthetic ones.
• Pricing and willingness to pay. Brand, Israeli and Ngwe (Harvard Business School) estimated willingness to pay from language model responses and compared them with human studies.
• Predictions for individual people. Park and colleagues (Stanford) and Toubia and colleagues (Marketing Science) built twins of over 1,000 and 2,058 real people respectively.
• Ad, message and customer experience tests. Vendors such as iMotions, Genesys and Dovetail name these uses on their pages. We found no study that measures twins on ad copy against market results.

How accurate are digital twins?

They are measured against a strict benchmark: people answer the same questions again after a while, and they do not always answer the same way. This test-retest accuracy is the ceiling. Studies report what share of it the twins reach.

An example from Twin-2K-500: people repeated their own answers 81.72% of the time, the twins got 71.72% right. The authors give the ratio across 17 tasks as 87.67%.

Be careful with reports of “85% accuracy”: that figure comes from the first version of the Stanford study from 2024 and also means a share of test-retest accuracy, not 85% correct answers. The Future of Marketing Institute, for example, writes that the twins replicate how participants think “with 85% accuracy”.

Maier and colleagues also found that asking language models directly for numerical ratings produces unrealistic response distributions. Their result applies to their own method, not to any prompt.

Where digital twins get it wrong

• Too normative. In Twin-2K-500, the twins always picked the normative option on one task, while their human counterparts did so about 30% of the time. About 45% of the people refused a vaccine, but only 4.0% of the twins.
• Willingness to pay with the wrong sign. According to Brand and colleagues, estimates from language models are sometimes comparable to human studies, but often inaccurate and in some cases wrong-signed. Fluoride in toothpaste was valued roughly three times too high, a larger laptop screen negatively.
• No reliable differences between segments. Same study: the models reflect the average and do not provide meaningful differences between customer groups. Fine-tuning helped within a product category, but not for new categories or between segments.
• Too little variance, unstable results. Bisbee and colleagues (2024) found less variation than in real surveys. 48% of estimated coefficients differed significantly, and in 32% of those cases the sign flipped. The same prompt gave significantly different results over three months.
• Too favorable. Nielsen Norman Group observes shallow or overly favorable feedback from synthetic users. They seem to care about everything.

Brand and colleagues conclude that language models in market research are a supplement to human studies, not a substitute. Without a human benchmark, a researcher would have had no way to tell which estimates were unreliable.

How to evaluate a digital twin vendor

• Measured against real answers? Ask for a comparison with real people’s answers the twin had not seen, and for the benchmark: share of test-retest accuracy or raw hit rate.
• What is the twin based on? Interviews or surveys with real people did clearly better than demographics alone in Park and colleagues’ study (83% vs 74%).
• Does it show differences? Ask to see whether different segments answer differently. This is exactly where the models failed in Brand and colleagues’ study.
• Does it push back? A twin that likes every concept tells you nothing about objections.
• How current is the data? A twin is a snapshot, not a live data stream as in industry.
• Whose data is in it? A twin of real customers is built on personal data. Clarify the legal basis for processing it first.

What this means for ad and page copy

None of the studies tested ad or page copy against market results. The patterns are still the ones to watch for in copy tests: twins judge more conventionally and more uniformly than people, so they tend to underestimate objections.

99takes does not call its personas digital twins, because they are not copies of individual customers. They are personas from your audience, researched from competitor pages and public customer voices, each with its source; without evidence, an audience is labeled a hypothesis. An audience typically has 10 to 14 personas covering a wide range, and versions are checked blind. How often the jury was right is shown by the hit rate per jury, with the number of results and a range, published once there are ten decidable results.

99takes runs on Claude and has no language model of its own. It is a simulation before launch, not a market measurement.

Frequently asked questions