A FOUNDATION MODEL FOR JAPANESE CONSUMER BEHAVIOR

What consumers buy
shows in behavior, not in words.
So we simulate the
behavior directly.

A synthetic-consumer platform built for the Japanese market.
1,000 AI agents generated from the Japan Census validate concepts, packaging, and pricing before launch.
Correlation with real purchase data: r = 0.78–0.92. First report delivered in as little as 72 hours.

r=0.92vs. real purchase 72hfirst draft N=1,000synthetic consumers NDA · No training use
LIVE · SIMULATING T+000
PREDICTED PURCHASE INCIDENCE
0.0%
N=1,000 · RTD BEVERAGE
VALIDATION · r=0.92 0 / 1000 AGENTS
QUESTIONS YOU CAN ASK SAMPLE PROMPTS · 4 OF ∞
PACKAGE TEST
Among women aged 30–40, which scores higher — Package A or B? Break down by attribute.
→ Top2Box by segment / preference rationale
PRICING
If the suggested retail price moves from ¥198 → ¥248, by how much does purchase intent fall?
→ Price sensitivity curve / PSM
CONCEPT
How much does acceptance of this functional-beverage concept differ between Tokyo and Osaka?
→ Regional variance / segment fit
SWITCHING
If this new product launches, which existing brands will customers switch from?
→ Switching sources / cannibalization
VALIDATION FIDELITY
r=0
Upper-bound correlation vs. real purchase / survey data
SAMPLE POPULATION
N=0
Census-aligned, evaluated in parallel
TIME TO READOUT
0h
1/40 of a traditional panel survey
UNIT ECONOMICS
1/0
~20% of an equivalent-size panel
THE THESIS

Between words and behavior,
there is a structural gap.

Surveys, focus groups, and Top2Box purchase-intent scores are well-honed methods that remain essential in the final validation phase. But the people who say "I'll buy it" do not always buy — a structural gap known as the Say-Do Gap. Launch Simulation is a complementary method that fills this gap by simulating behavior — not preference — in the screening and shortlisting stages.

STATED INTENT · self-reported
100people say "I'll buy it"
Surveys measure intent. Intent is not decision.
SAY DO
REVEALED BEHAVIOR · at the shelf
100actually carry it to the register
Up to 25% drop off. That is the Say/Do Gap — the root cause of 85% of launch failures.
01 / SAY/DO GAP
25%

Intent does not equal behavior

Up to 25% of respondents who say "I'll buy it" never actually purchase. No amount of survey-design refinement removes the structural error inherent in self-report.

02 / SUNK COST
¥0.5-1.5M

Traditional cost per concept

A single quantitative panel (N=1,000, 20–25 questions) typically runs ¥0.5–1.5M (close to ¥2M with heavy video/image or multi-SKU comparisons). Cycle time of 4–8 weeks structurally limits how many concepts you can test.

03 / MARKET REALITY
85%

New-product failure rate

Share of new products that exit the market within two years of launch. Source: Nielsen/NIQ 2023, ESOMAR 2024. Even with thorough research, this number has not improved.

VALIDATION · PEER-REVIEWED + INTERNAL

Not claims —
verifiability.

The accuracy of synthetic consumers is determined by the depth of training data and how well it fits a benchmark. Three peer-reviewed studies support the validity of this approach, and we have validated against real purchase data in the beverage category. We can reproduce your past surveys under a Blind Validation protocol and prove fit with a quantitative reconciliation report.

STUDY 01

Accuracy validation of digital twins

Large-scale validation using data from over 2,000 real consumers. AI digital twins demonstrated high accuracy in economic and social domains such as purchase decisions and price sensitivity.

Toubia et al. (2025)
Marketing Science / Columbia
STUDY 02

Reproducing purchase intent with synthetic consumers

Compared against 57 real consumer studies (9,300 respondents), AI synthetic consumers achieved 90% of human test-retest reliability. Scoring natural-language responses by semantic similarity is the key.

Maier et al. (2025)
PyMC Labs × Colgate-Palmolive
STUDY 03

Persona depth determines accuracy

Equipping AI personas with detailed life context, values, and purchase habits improves alignment with real surveys by up to 87%. We embed life context derived from the Japan Census and Household Survey.

Park et al. (2025)
Persona-Based Simulation

Internal validation (beverage category)

Reconciliation against real purchase data is ongoing. Through Blind Validation, your team can independently confirm reproducibility on your own data.

N=12 CONCEPTS · INTERNAL
0.00
Brand share prediction
0.00
Purchase drivers by segment
0.00
Price sensitivity
0.00
Past-concept reproduction
PROTOCOL · 72H

A decision back
in 72 hours.

Census-based population generation × Population-True Approach × Three-layer judgment architecture

Hand off your existing concept materials as-is. We generate AI personas, run simulations in parallel, and reconcile the results internally — then deliver a report shaped to your team's decision format.

STEP.01
01 / INPUT

Hand off your
existing materials as-is

Product name, target audience, price range, positioning, concept board. PDF, PPT, or Word — any format. No reformatting required.

REQUIRED · ≤ 1H
STEP.02
02 / SIMULATE

1,000 synthetic consumers
reproduce the decision

1,000 agents aligned with demographics, household spending, and category purchase history simulate behavior at the shelf — each grounded in their own context (family, budget, loyalty to existing brands).

PROCESSING · 6-24H
STEP.03
03 / DELIVER

Delivered as a report
your team can act on

Purchase probability, Top2Box by segment, price-sensitivity curves, main switching sources, and improvement implications. Delivered in a format ready to drop into your internal approval workflow.

DELIVERY · ≤ 72H
LIVE SIMULATION · 5 PERSONAS DEEP DIVE

Your product, judged
by 5 people from 5 lives.

Enter product details below and press "Run Simulation". The Claude API simulates 5 real Japanese consumers, evaluating each one's life context, inner voice, and price acceptability. The full version delivers the same resolution across 1,000 personas, aggregated by segment.

Simulation Settings

5 PERSONAS · LIVE API
SIMULATING · LIVE API

5 AI personas are standing at the shelf.

Generating attribute distribution from Japan Census & Household Survey
Building 5 personas (age, occupation, income, lifestyle)
Evaluating the concept in each persona's context
Generating inner voice, peer dialogue, and price acceptability
Constructing rationale logic for GO / HOLD / NO-GO
↑ Enter product details and press "Run Simulation".
The API typically takes 20–40 seconds to respond.
* Generated in real time via the Claude API. 5-persona sample. The full version runs 1,000 personas in parallel, broken out by Top2Box per segment, price elasticity, and switching sources.
ECONOMICS

Higher decision quality
at lower cost.

Not a replacement for traditional panel surveys — an upstream layer that feeds into them. We increase the density of decisions made at the screening stage, while panel surveys continue to anchor the final validation phase.

METRIC
Traditional concept test (complemented)
Launch Simulation (upstream)
Cost
¥0.5–1.5M / concept
From ¥150,000 / concept
Lead time
4–8 weeks
Within 72 hours
Stage in the funnel
Pre-screening narrowing handled separately
Purpose-built for upstream narrowing
Test volume
A few times per year
Any number of concepts compared in parallel
Say/Do Gap
Self-report based (category-dependent)
Complemented by behavior-based prediction
Iteration cycle
One concept at a time, rigorously
Multiple concepts validated in parallel
Accuracy validation
Performed in the final validation phase
Pre-confirmed via independent Blind Validation
USE CASES · WHO BUYS THIS

Used behind decisions
like these.

Product planning for beverages, confectionery, and household goods; agency pitch phases; the upstream layer of research firms. Core use cases: launch go/no-go, concept narrowing, pre-launch price-sensitivity testing, and renewal decisions.

★ PRIMARY

Food, beverage, household goods
Product planning / brand managers

Owners of launch decisions and screening leads
  • Narrow down winning concepts before the quantitative panel
  • Compare N concepts in parallel within a category to focus development resources
  • Understand price sensitivity and segment-level Top2Box before launch
  • Plug into short-cycle decisions: limited editions, SKU extensions, renewals
  • De-risk launches in high-unit-price categories where failure is expensive
Agency / consulting

Marketing strategy
Strategy / planning

Client pitches / strategy development
  • Add quantitative backing to proposals quickly
  • Pre-test message A/B before campaign launch
  • Compress research-outsourcing lead time and raise win rates
  • Differentiate existing offerings with synthetic-consumer validation
Research firms / insights teams

OEM / offering extension
BizDev / strategy

Upstream extension of an existing panel business
  • Add a synthetic-consumer layer upstream of existing panel surveys
  • Build a new offering for existing clients
  • Provide leading-edge capabilities without building an LLM stack in-house
  • Confirm accuracy via Blind Validation before deciding to integrate
REACH · HARD-TO-PANEL SEGMENTS

Reach segments
panels rarely cover — at the same accuracy.

SUPPLEMENTARY · NON-EXHAUSTIVE
REGIONAL · SENIOR
Seniors in rural areas
Hard to scale in online panels skewed toward major metros. Region, household composition, and disposable income are reconstructed from the Japan Census.
CARE-GIVER
Mothers on parental leave / raising children
High panel drop-off rate due to time constraints. Daily routines and purchase timing are reproduced in context.
SMB OWNER
SMB owners and sole proprietors
Industry panels are expensive to recruit and have low cooperation rates. We model the decision context of those who run a B2C business alongside their work.
HIGH-INCOME
High-income / affluent households
Low incidence in panels; reproducing lived reality matters more than surveys. Behavior is derived from disposable income and spend structure.
PRICING · Transparent

From ¥150,000 per concept,
raise your decision quality.

Quantitative panel surveys in Japan (N=1,000, 20–25 questions) typically run ¥0.5–1.5M per concept. Launch Simulation starts at ¥150,000 per concept. Generating and running 1,000 AI personas is dominated by server cost, so unit cost does not scale linearly with volume — the more concepts you test in parallel, the lower the unit price.

PLAN 01 · One-off trial

Single Test

Validate a single concept with 1,000 AI personas. Delivered in as little as 72 hours.

¥ 150K + / concept
  • 1,000 AI personas built (category-specific)
  • Top2Box purchase intent & cross-tabs by segment
  • Price sensitivity analysis (PSM)
  • Qualitative comments (150 extracted)
  • Report (PDF / raw data)
PLAN 03 · Enterprise

Enterprise

For product development teams, agencies, and research firms. Annual contracts / API / OEM.

ASK
  • Annual allocation of 20–100 concepts
  • Direct API integration / Slack notifications
  • Custom AI persona tuning
  • White-label delivery (OEM)
  • Quarterly review and model updates
  • SLA / domestic-processing option

* Pricing varies with question count, category, and lead time. A firm quote is provided after NDA.
* Start with Blind Validation (free). Confirm accuracy against your past tests before moving to paid work.
* Source for traditional benchmarks: published rate cards from Research Boutique, Rakuten Insight, Macromill, and Asmarq (2024–2025).

FAQ

Frequently Asked Questions

CONTACT · GET STARTED

Start with a 30-minute conversation.

No contract required up front. We'll learn about your situation and recommend which next step fits best — Blind Validation, Pilot Test, or Full Engagement. Confidentiality applies from the first 30 minutes; you can talk to us without sharing any sensitive information.

01
RECOMMENDED · Reply within 24 hours

Book a 30-minute consultation

A senior team member will hear out your product, your challenge, and your accuracy expectations, then propose the best path forward.

  • Duration: 30 min (online)
  • NDA can be signed before the first call
  • Quote can follow immediately after the proposal
02
PILOT · First 3 companies only

Blind Validation (1 case free)

Run the simulation blind to your past test results. We deliver a quantitative reconciliation report showing how closely we hit them.

  • Only the concept brief is shared
  • Reconciliation report delivered
  • No fee if accuracy expectations are not met
03
DOC · Instant download

Download the service deck

Accuracy benchmarks, pricing plans, and case studies (24-page PDF). Built for internal sharing and approval workflows.

  • Detailed delivery process
  • Three pricing plans
  • Three anonymized case studies
RESPONSE
Within 24 hours (business days)
▶ 30-min consultation