Favorites: An AI powered vision in six weeks

Context

Favorites, the saved homes at the heart of Zillow's save-and-return loop, needed a vision, and nobody had one. Product was largely unopinionated about the space; the problem definition itself was up for grabs. And like everything at Zillow, the surface touched many teams, which meant any real vision would have to work across multiple organizations, not just mine.

I was asked to create that vision. I gave myself six weeks.

How I ran it

I drove the foundation personally: problem definition, success metrics, the first rounds of research, and AI-built prototypes that got real feedback in front of users fast. With that foundation solid, I brought in senior and principal teammates for object-model mapping workshops and principles creation, then ran a final round of moderated studies to pressure-test the direction.

The methodological bet was object-oriented UX. Partnering with the IA team, I mapped the actual objects in the space (homes, collections, searches, the relationships between them) before designing any screens. That model, not the mockups, is what made the vision durable: it gave every team touching the surface a shared vocabulary, and it's now rippling across the design org as an emerging standard.

Screens are opinions. Object models are agreements. Agreements are what survive contact with seven workstreams and three orgs.

From model to evidence

With the model and principles set, research told us where to drive. A first-round card sort with 20 high-intent movers established the strategic frame in one sentence: Favorites is a decision space, not a discovery surface. Comparing and narrowing ranked as the top job of the phase, at 4.9 out of 5. Collaboration should be the default, but adapt gracefully for solo shoppers. And users decisively voted several candidates off the island — vague recommendations, tour management, and, notably, agent presence.

Three concept directions then went in front of users in moderated studies, ranging from a familiar collections-first organizer to a fully collaborative decision hub built around alignment between co-shoppers.

The moderated round produced the vision's most energizing finding: a pros-and-cons feature that surfaced each home's trade-offs was spontaneously praised in 13 of 23 interviews. "The model that showed the flaws was the best model," as one participant put it. It also produced the most honest one: doing pros and cons well requires a depth of preference understanding Zillow is still building. That dependency didn't kill the feature; it became a named workstream on the roadmap, sequenced for when the capability lands.

AI as force multiplier — with receipts

Six weeks was only possible because I used AI aggressively: to teach myself unfamiliar territory, to enhance the thinking, and to compress prototyping from weeks to days. But I used it with two non-negotiable rules. First, I verified sources: AI accelerated my research; it never replaced my judgment about what was true. Second, I labeled what was AI-made when sharing it, because the fastest way to burn a team's trust in AI-augmented work is to let them discover the provenance themselves.

That transparency turned out to be a feature, not a disclaimer. It modeled a way of working that colleagues could adopt without feeling deceived by it.

Outcome

Six weeks in, we had a tested vision, a shared object model, design principles, and a six-month roadmap that sequenced the work and named exactly when we'd need dependency-creating teams. The vision in one sentence: turn Favorites from a static bookmark into a decision engine, for solo shoppers and co-shoppers alike. Product and SLT bought in, and three of the seven workstreams have already kicked off.