Scientifically validated: AI personas deliver reliable results

Discover why researchers and companies rely on AI-based audience research

Stanford study confirms validity

Generative AI agents simulate human behavior with high accuracy

A recent study led by researchers at Stanford University examines the ability of generative AI agents to simulate human behavior and attitudes. The research findings provide empirical evidence for the validity of AI-based approaches in audience research.

Key findings of the study:

The researchers created generative agents based on interviews with 1,052 representative US participants and tested their predictive ability using several standardized measurement instruments:

  • High consistency rate: The generative agents achieved 85% of the consistency with which humans reproduced their own answers after two weeks.
  • Methodological superiority: The interview-based agents showed significantly better results than those based solely on demographic data (14-15 percentage points higher normalized accuracy).
  • Reduced demographic disparities: The interview-based agents consistently showed smaller performance differences between demographic groups.
View full study
"The correlation between the effects predicted by agents and actually observed treatment effects reached r = 0.98, indicating precise simulation of group behavior."

— From the Stanford study

"AI personas discovered the same UX problems as real users, but provided more detailed feedback explanations."

— From our website test, April 7, 2025

Website test with AI personas

Case study: Persona-based vs. generic AI test of a zoo website

On April 7, 2025, we conducted an insightful proof-of-concept comparing two different AI approaches in website testing: A generic AI without personality and an AI with a simulated persona. Both were given the same task: Planning a zoo visit on the website of a major German zoo.

  • The AI with persona ("Lara Schneider", a 28-year-old project manager) provided more nuanced insights into the actual user journey and identified specific UX problems.
  • While the generic AI discovered many basic problems, the persona AI could show emotional friction points and context-related hurdles that would otherwise be overlooked.
  • The persona AI suggested more innovative and targeted features, such as "A simple 'Plan your visit' function" or "Interactive map with optimal routes between animal feedings".

"As the product manager responsible for the zoo website, Lara Schneider's report is significantly more valuable to me. It identifies concrete UI problems like the non-intuitive date selection and the overwhelming ticket page that directly affect our conversion rate. The suggested features like the 'Plan your visit' assistant and the interactive map with optimal routes exactly match the expectations of modern users. These insights allow us to prioritize improvements that will have the greatest impact on user satisfaction."
— Simulated product manager in the comparative website analysis, April 2025

AI personas vs. Forsa study: Nearly identical results

We surveyed 2,000 simulated personalities on the same topics as a representative Forsa study commissioned by Targobank– with astonishing agreement

Speed limit on highways

Forsa study

63%

Support a speed limit

AI simulation

68.05%

Support a speed limit

Deviation: only 5%

Feature subscriptions in cars

Forsa study

78%

Reject feature subscriptions

AI simulation

76.93%

Reject feature subscriptions

Deviation: only 1.07%

Charging infrastructure at workplaces

Forsa study

75-78%

Support expansion

AI simulation

74.64%

Support expansion

Deviation: only 0.36-3.36%

Methodology & Notes

Our simulation included 2,000 AI personas that we surveyed on the same topics as a representative Forsa survey.

  • While nearly identical results were achieved on some topics, there were also larger deviations on other questions.
  • Without access to the exact Forsa questionnaires, we had to reconstruct the questions, which may lead to methodological differences.
  • The language model used had a knowledge cutoff of November 2023, which may have affected the timeliness of some answers.

Our research and development process

We are continuously working to improve our AI personas. Although the results are already impressive, there is still research needed in the following areas:

  • Even more natural and nuanced response patterns that reflect the full complexity of human reactions
  • Integration of current events and societal developments after the language model's knowledge cutoff
  • Improved simulation of cultural nuances and regional peculiarities in different markets

Despite these ongoing optimizations, our studies and customer projects already show today: The AI simulation is "good enough" for most marketing and product development purposes - while offering significant advantages in terms of speed, cost and scalability.

Experience the difference with AI personas

Try our interactive demo or secure Early Access