I built a 'Synthetic Market' to predict the Soda Wars (and it actually worked)

Published: (April 30, 2026 at 03:48 PM EDT)
2 min read
Source: Dev.to

Source: Dev.to

Most market research is a lie. People say they’ll buy a product in a survey, but then they abandon their cart in real life.

Yesterday, I launched Sediman to bridge that gap. Instead of asking humans what they think, I built a behavioral engine that spawns AI agents with rich, census‑weighted personas—diverse incomes, ages, and psychological biases—and lets them “live” through a market simulation.

The Experiment

I ran 100 agents through a 2026 Cola War simulation.

Result: A 56.8% – 43.2% split for Coke vs. Pepsi.

Logic: The agents weren’t just picking names. High‑income personas clustered at Costco; lower‑income personas prioritized value at Aldi. They behaved like real people under financial stress.

The Tech Stack

Persona Layer

Deep psychographic prompting to ensure “Avery, 28, $135k” doesn’t think the same way as a retiree in Ohio.

Decision Engine

Separating the Identity from the Evaluator to prevent LLM brand bias.

Web Mode (WIP)

Moving from surveys to actual site‑navigation audits.

Why I’m Building This

In 2026, we shouldn’t wait six weeks for a focus group. We should be able to “unit test” our marketing strategy.

I’m currently building out the Web Agent to let these personas browse real D2C sites to find conversion leaks.

Check out the full Day 1 results here.

0 views
Back to Blog

Related posts

Read more »

The smarter the model, the more it saves.

The Myth: Smarter Models Will Make Plugins Redundant Since WOZCODE launched, many Claude Code power users have whispered that the plugin’s advantage will disap...