Validation comes first · calibrated to real transactions

See a market before you move it.

Thousands of AI-agent shoppers, calibrated against real sales — not surveys. We simulate what people buy, not what they say, so you can test any price, product, or promotion before you commit.

0.64held-out share-recovery r on real supermarket sales it never saw
−2.51simulated price elasticity — matched the real elasticity exactly
revealedcalibrated to what people bought, not what they told a survey
How it works

A calibrated digital twin of your category — believable agents, anchored to real data.

01
📥
Inject your data

Your category sales — products, prices, units. Or public/syndicated data. Even a few weeks is enough to start.

02
🧑‍🤝‍🧑
Thousands of agents shop

LLM-agent shoppers weigh brand, attributes and price and choose — the way real buyers do, one decision at a time.

03
📊
Emergent market shares

Individual choices aggregate into shares, elasticities and substitution — the structure of your market, simulated.

04
🎯
Calibrated decisions

We tune the model to your real sales, then you run any what-if: price, assortment, promo, AI-channel — decision-grade.

Does it actually work?

We tested the engine against real market data with a held-out design, so it can't be curve-fitting.

0.64
held-out share-recovery r on real supermarket data (markets it never saw)
−2.51
simulated price elasticity — matched the real elasticity exactly
1 wk
of your data is enough for the sim-prior to beat a pure data fit
2
independent datasets validated (Nevo cereal + dunnhumby retail)
DatasetElasticity sim vs realRaw sim+ calibration (held-out)
Nevo cereal (IO textbook)−2.5 vs −3.9r = −0.35r = +0.44
dunnhumby real supermarket (pizza)−2.51 vs −2.51r = +0.05r = +0.64

The simulation supplies the price mechanism (worth the most where you have no data); calibration supplies the brand levels from your data. Together: decision-grade.

What you get

One engine, three answers your team makes decisions on.

🏷️
Pricing & promotion

Own- and cross-price elasticity, promo lift, and exactly who you steal share from — or lose it to.

🧩
Assortment & launch

Add, drop, or launch a SKU — with cannibalization, including products that have no sales history yet.

🔎
AI-channel visibility

Are you found when shoppers and agents ask AI what to buy — and what agentic commerce does to your demand.

Simulate what people buy — not what they say

The synthetic-consumer field asks agents what they'd say in a survey. We calibrate agents to what people actually bought — the difference between believable and accurate.

 Others (Simile · Aaru)Avanti
Anchored tostated preference — surveys, interviewsrevealed preference — real sales
Validated byreproducing survey answers (self-retest)recovering real market shares it never saw (r 0.64)
Founder DNACS / HCI / MLagents + microeconomics (demand estimation)
Outputopinion distributionsdecision-grade price & share counterfactuals

The field's own critique — "individually believable, collectively wrong" — is only answered by calibrating to what people actually bought.

Run your category through it.

Drag a price. Watch thousands of agents re-decide. See who wins.

Open the live simulator →