At a glance
Timeline: August 2025 – January 2026 (~5 months end-to-end; ~2 months active design)
Platform: ChatGPT app, built on the MCP iframe layout
Outcome (since Feb 2026 tracking): 170K visitors, 140K searches, 20K click-throughs to Udemy.com, 2,350 buyers, $110.5K GMV
The problem
What is ChatGPT actually good at that Udemy.com isn't, and where does that overlap with a real learner problem?
Udemy’s vast catalog made keyword search insufficient, especially for learners with broad or ambiguous goals who needed contextual course recommendations.
We had no data on how users were actually querying ChatGPT about Udemy or about learning more broadly - The v1 product was designed as a discovery and vetting surface in ChatGPT, instrumented to learn how users actually ask AI about learning to guide future iterations.
Those two observations reframed the product brief. The product wasn't "Udemy on ChatGPT." It was a discovery and vetting surface for Udemy's catalog, designed around ChatGPT's conversational affordances, and instrumented to teach us what learners actually ask AI about learning. That reframe became the thesis the rest of the team aligned around.
Discovery
There was no formal user research before design. This was a tradeoff I'll return to in the reflection. Instead, I worked from two sources of grounding: deep familiarity with Udemy's catalog and platform behavior, and direct exploration of ChatGPT's affordances and constraints as a host environment.
The catalog work was the more important of the two. Running a handful of representative queries against Udemy's own search made the density problem concrete: tens to hundreds of viable courses for nearly any topic, with ranking driven by signals (popularity, recency, instructor reputation) that don't always map to a given learner's needs. The implication was that any v1 of this product had to help users *vet* options, not just see them.
The ChatGPT exploration shaped the form factor. ChatGPT's MCP integration provides iframe-based layouts inside an otherwise conversational interface, meaning you can render real UI, but the user's mental model is still a conversation. That dual nature became the design constraint that organized everything else.