Ask a generic AI chatbot "Will my premium skincare brand sell in Vietnam?" and you get a generic guess. AI Market Twin runs simulations on a persona pool that's actually grounded in real government and public statistics — Vietnam's Bộ Y Tế food registration process, Indonesia's BPS household income distribution, Japan's 厚生労働省 cosmetics regulations, and 21 more.
Income distribution, profession-level wages, household composition, consumption patterns, and core regulations — measured data pulled from each country's government and public statistics offices is the ground truth for persona generation.
Refreshed annually via an automated GitHub Actions pipeline · Additional countries can be onboarded in 4–6 weeks on request
You'll get an answer either way. The question is whether the answer is something you can actually take into a boardroom.
The most reliable way to reduce LLM hallucination is external grounding plus self-evaluation. AI Market Twin runs public-data anchors — universal ones for every origin, plus per-origin national-data providers — and a 5-metric self-scoring pipeline.
Origin-agnostic by design: each home country plugs in its national equivalents of listed-company financials and food/drug regulatory data. Korea, the United States, Japan, and the United Kingdom are wired today; more markets extend the same interface.
The back-test showed the true winning market is often decided by brand-level factors an existing footprint or a distribution deal (Shake Shack → UAE via a licensing group) that no macro anchor can detect. Two signals surface them; live-only, weighted into the country ranking.
Confidence via dominance. A STRONG label is not a mere plurality — it requires the winning market to dominate: agreement across independent LLM providers and a clear vote-share margin over the runner-up. This calibration came from the N=19 back-test, where a plurality-only STRONG was over-confident (right ~55%). The rule makes STRONG mean "bet on this", MODERATE "shortlist it", WEAK "don't bet on this alone".
The 5-metric scoring runs as open, auditable logic. Every code change is re-measured on the same product set (multiple K-product fixtures) automatically, and results are verified with paired t-tests for statistical significance. The measurement-improvement cycle runs weekly.
Every sim result lists the grounding anchors used and a per-metric score breakdown directly in the PDF. Accuracy improvements roll out to existing customers automatically — no separate upgrade cost per release.
From wizard input to PDF output in 5–7 minutes on average. Each stage uses a different LLM model tuned to its role.
Validates product info and pre-allocates 200 persona slots from category-specific profession pools.
Inspects sales bans and labelling rules per country and category. Markets where launch is structurally impossible are auto-excluded.
Reuses matching personas from the workspace pool; only new slots get fresh generation. Each persona's first-person voice quote is generated alongside.
Aggregates intent distribution, rejection factors, and trust signals across 200 personas to score each market on demand, CAC, and competitive intensity.
Three parallel pricing simulations with median selection eliminate single-call variance. Per-market price sensitivity broken out separately.
Vision model analyses any uploaded creative. Self-critique pass automatically verifies macro consistency (best-country alignment, etc.) before finalising the result.
An anonymised K-food 5-market validation simulation. PDF, no signup.