.jpg)
When a technology emerges that promises to solve the fundamental constraints of an industry, I pay attention. When that technology is built on fundamentally flawed assumptions, I speak out.
The latest buzz? Using GPT to generate "synthetic personas" for search and personalization. The pitch sounds compelling: feed product catalogs into an LLM, generate artificial users with fabricated preferences, simulate their behaviors, and—voilà!—you've magically solved the cold start problem in recommendations.
Let me be unequivocally clear: this approach is intellectual snake oil.
As someone who's spent years optimizing recommendation systems at scale (with the patents to prove it), I understand the allure. Data sparsity is the Achilles' heel of personalization. But simulating human behavior with LLMs fundamentally misunderstands what drives conversion.
Here's the inconvenient truth: human purchasing behavior isn't deterministic text generation. Human behavior—especially in commerce—isn't neatly scripted. Customers aren’t GPT-generated automatons faithfully following logical pathways. They’re unpredictable, moody, irrational, and brilliantly chaotic. No matter how sophisticated your LLM-driven "agents" become, modeling the subtleties of human motivation and randomness at scale remains fundamentally out of reach.
At Sequen, we don’t fall for shiny toys. Our LEM (Large Event Model) was built precisely because we respect the complexity of genuine user interactions. Instead of approximating reality with synthetic personas, LEM is trained on actual nuanced user behaviors—real clicks, real purchases, real hesitation, real irrationality. Our proprietary technology doesn't rely on guesswork; it captures the authentic complexity of human engagement.
Rather than hallucinating what users might do, Sequen has built a system that adapts in real-time to what they actually do. Our technology delivers 3B+ recommendations/month with p99 response times under 25ms, enabling behavior optimization that synthetic data simply cannot match.
The proof? Our first customer cohort generated an additional USD 162M in annualized revenue over just seven months. That's not theoretical—it's measured impact.
While competitors chase the mirage of synthetic personas, we're focused on changing how users act, not just what they see. Our LEMs overcome data sparsity by leveraging billions of actual user interactions across digital properties, delivering personalization that responds to genuine behavior patterns rather than fabricated ones.
The most sophisticated recommendation engines fail without sufficient user data—this is true. But the solution isn't to manufacture artificial data. It's to build systems that can extract maximum value from minimal real data, then adapt and learn with each interaction.
The future of digital experiences isn't in simulated behavior—it's in optimized actual behavior. And at Sequen, we're building that future today, one real interaction at a time.
Because in the end, what matters isn't how clever your synthetic data generation is—it's whether you can actually drive measurable business outcomes. And on that metric, our LEMs speak for themselves.
— Zoe Weil
Cofounder & CEO