How Urekata thinks

The evidence behind reactions,
open to inspection.

These three studies informed how we ask products’ questions.
We separate what the research found from what Urekata itself has verified.

The 100 customers are a fixed sample from our open dataset Japan Synthetic Personas (CC BY 4.0), built on NVIDIA Nemotron-Personas-Japan (CC BY 4.0) with household-income and consumer-behavior attributes added.
Fujita, Atsunori; Gong, Vincent; Ogushi, Masaya; Yamamoto, Kotaro; Suhara, Yoshi; Corneil, Dane; Meyer, Yev. Nemotron-Personas-Japan, September 2025.

Current comparison method

The same 100 personas choose among two or three options

Change just one item: price, headline, product description, target customer, or sales channel. Personas and all other product conditions stay fixed.

Comparison method version m2026-10-07.compare.2Response instruction version c2026-10-07.2

The same 100 people see the options side by side and pick one (they can also pick "none of these"). We ask five separate times. The order of the options changes for each group of 10 people and each round, so across the 100 people every option appears in every position about equally often. "None of these" is always last. Each person's choice is the option they picked most often across the five rounds; ties go to the earlier round.

What we ask each round (instruction summary)

Choose the one option you would pay for at the listed price, considering your own life, budget, and shopping habits. If you would not buy any option, choose "none of these". Explain the reason in your own words.

Presentation order, counting, and reasons

The five rounds are not counted as 500 people. Each person’s final choice is their most frequent choice; ties go to the tied option they selected first. If options have equal final counts, the total choices across all five rounds break the tie, followed by A, B, then C if still tied. This is a counting rule, not evidence of a difference.

Stability counts how many rounds had the final leading option as the unique leader among the options. Tied rounds do not count. Five rounds gives "same in all five", four gives "four of five", and three or fewer gives "changes across rounds". "None of these" is excluded from option ranking; a count higher than every option is shown separately. When leaders change across rounds and another option is within five people of the leader, the result is shown as "close".

Reason groups use each person’s reason from the first round in which they selected their final choice. AI summarizes these into short tags for each choice. Each person counts toward at most one tag; the three largest tags are shown. The representative quote comes from a stored reason.

Limits when reading the result

The 100 people are AI customers built from synthetic data, not self-reports or behavior records of real people.

Counts show how many of the 100 picked each option; they are not forecasts of purchase rates or sales.

Confidence is shown as how many of the five rounds had the same top option among the options, never as a percentage or score. A lead for "none of these" is shown separately.

Reason tags are short AI summaries of the reasons people wrote; the count on each tag is how many people wrote that reason.

01Comparing how you tell it

Benomar et al. · 2026

Same product — does the answer change with how you ask?

A study on whether how you prompt AI affects the conclusion when comparing ways to present a new product.

For exampleWhether to pitch the same eco detergent as "time-saving" or "gentle on skin." The study varied the order of the two options and how answers were given.

At Urekata

Urekata presents two or three options side by side to the same 100 personas and asks them to choose five times in changing orders. The study informs how we separate option content from presentation order.

Method details and what Urekata does not do

The study validated virtual customers built on real behavioral data against past A/B tests. Urekata uses 100 AI-created personas. It supports both two- or three-option comparisons and single-option analysis. Its accuracy has not been validated against real behavioral history or sales-test results.

Read the paper Data-Driven Persona-Conditioned Agents for A/B Test Simulation

02How to ask for reactions

Maier et al. · 2025

Does a "5 out of 5" alone capture how people feel?

A study on whether asking for purchase intent in words first works better than asking for a number outright.

For exampleIf someone seeing a new shampoo says "I like the scent, but I’m hesitant about a subscription," that hesitation stays on record.

At Urekata

Urekata references this when designing questions that keep reasons separate from scores. The current questions and scoring definitions are published below.

Method details and what Urekata does not do

The paper’s SSR method embeds free text, compares it to five reference statements, and converts that into a probability distribution over ratings. Urekata does not reproduce SSR and has not validated it for Japanese products.

Read the paper LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings

03How personas are built

Park et al. · revised 2026

Can age and gender alone explain why people buy?

A study on whether simulating a person’s answers takes more than age and gender — everyday life and values too.

For exampleEven within dual-income households, "worried about my child’s skin" and "just want less housework" respond to different messages.

At Urekata

Urekata references this when letting synthetic personas react to a product while considering their life situations and how they choose.

Method details and what Urekata does not do

The study used long interviews and questionnaires with real people. All 100 Urekata personas are synthetic; they do not reproduce interviews or self-reports from real individuals.

Read the paper LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals

The questions are open too.

Standard questions for single-option analysis

These are the four standard questions for reactions to one product. Analyses with custom questions save the questions and method used with the result. Comparisons do not use these four questions or the interest score.

Loading the question definitions…