Synthetic consumers, built from real consented data. Simulate the market before you spend.

Synthetic consumer simulation

Test your campaign on 1,000 synthetic consumers, before you spend a dollar.

Pebbles builds AI consumers from real, consented zero-party data. Simulate how a market responds to a product, price, or message, segment by segment, before the budget is committed.

Grounded in peer-reviewed choice science.

1,000 agents · simulating
Campaign B/A
Value seekers72%
Early adopters58%
Brand loyal41%

Predicted positive response by segment

0

Synthetic consumers per simulation

0%

Parity with human test-retest reliability

0%

Agentic task success rate

0.00

Segment stability (bootstrap Jaccard)

The problem

Brands spend first and learn second.

Traditional market research is slow, costly, and small-sample, so most launches fly blind and the real feedback only arrives once the media budget is already gone. Pebbles moves that learning upstream: simulate the market, then spend.

How it works

One loop: capture, simulate, activate.

Layer 1 · Capture

Showkey earns the data others guess at.

Showkey is an AI shopping assistant people actually want to use. In exchange for genuinely useful help, shoppers state real preferences, giving Pebbles consented first-party signal that behavioral tracking can never see.

  • Zero-party data: stated, not inferred
  • Consent-first, privacy-preserving capture
  • Preference depth, not click exhaust

Showkey

What matters most for your next serum?

Fast absorptionFragrance-freeUnder $40Clean brand

Zero-party · consented

Layer 2 · Simulate

1,000 AI consumers, shaped by real preferences.

Those preferences initialize a population of synthetic consumers. Pose a scenario, a new SKU, a price change, a campaign line, and a market-shaped response emerges segment by segment, in minutes instead of weeks.

  • A population, not a single persona
  • Segment-level response, not one blended average
  • Minutes per scenario, run as many as you like

Scenario · price −10%

Positive response61%

Layer 3 · Activate

Carry the winner into the real campaign.

The simulation does not replace the market, it tells you where to point before you pay for reach. Push the winning scenario downstream into live campaigns and commerce with MachimShop.

  • Decide before the media spend
  • Test messages, prices, and products
  • From simulation to launch in one loop

MachimShop · recommended

“Fragrance-free. Absorbs in 30 seconds.”

Predicted lift vs. control82%
Confidence74%

Ready to launch

The engine

A simulation engine, not a chatbot wrapper.

Pebbles treats a large language model as a calibration input to choice science, conjoint and discrete-choice modeling, not as an oracle you ask and trust.

01

Grounded in real data

Synthetic consumers are initialized from consented zero-party preferences, not model imagination.

02

LLM as calibration, not answer

The model anchors and augments a choice-science core, so estimates stay consistent instead of drifting into bias.

03

A decision model underneath

Each consumer is a POMDP: belief seeded by Bayesian priors, anchored by RAG over real responses, explored with MCTS.

04

Private by design

Differential privacy (ε = 1.0) and k-anonymity at capture. No raw personal data enters the simulation.

Inside one consumer

Bayesian belief

seed priors from ZPD

RAG anchoring

ground on real responses

MCTS search

explore consumer actions

Aggregate

segment-level response

End to end

From a real preference to a market decision.

01

Capture

Consented zero-party preferences via Showkey.

02

Model

Build a belief state for each synthetic consumer.

03

Simulate

Run 1,000 agents against your scenario.

04

Predict

Aggregate a segment-level market response.

05

Activate

Hand the winning direction to your campaign.

Validated

Measured against human ceilings, not marketing math.

We do not predict the future. We reproduce how consumers actually respond, and we measure that against the reliability of human panels themselves.

Slot F1

90%

Preference extraction

Intent F1

90%

Intent classification

Response fidelity

85%

vs. human test-retest

Segment stability

0.85

Bootstrap Jaccard

Task success

85%

Agentic retail tasks

Targets are set at the human ceiling: matching the reliability of human panels, not claiming to beat it. Figures are internal targets, verified by an accredited test lab report.

Why Pebbles

Built to be trusted, not just impressive.

Grounded in real, consented data

Every synthetic consumer traces back to a real stated preference, captured with consent, not scraped or inferred.

Peer-reviewed science, not vibes

The method builds on published choice-science and generative-agent research, not a one-off demo trick.

Privacy engineered in, from capture on

Differential privacy and k-anonymity live in the pipeline by default, not bolted on after the fact.

Grounded in research

Standing on peer-reviewed work.

Pebbles is built on published science, not a proprietary black box. A few of the results the method relies on:

Fusing LLM output with real data is a consistent estimator: it cuts the sample needed by 24.9–79.8% at equal accuracy, while naive replacement adds bias.
Marketing Science, 2025
Generative agents reproduce individual human survey and experiment responses at roughly 85% accuracy.
Stanford, 2024
“Silicon sampling”: language models conditioned on real profiles act as statistically faithful proxies for human populations.
Argyle et al., Political Analysis, 2023
Human test-retest reliability itself sits near 90%, which reframes our 90% as human-ceiling parity, not a superhuman claim.
Simulation reliability, 2025

See it run

Simulate your market before you spend on it.

Tell us what you would test, a product, a price, or a message, and we will show you a Pebbles simulation on your category.

  • Zero-party by design, no tracking pixels
  • Your captured data stays yours
  • A working simulation, not slides