Compass · Yaron Schoen

Compass Agent Search

Goal: Address low adoption of search and inventory capabilities.
Role: Principal designer. IC work, plus managing the Search and Inventory team of 4 designers. Reported to the CPO.
Approach: Redesign the system to accommodate agents' previously learned behavior, and regional discrepancies.
Outcome: Increased adoption in every region it was implemented. In DC, daily search use rose over 30% after launch.

Compass agent search as a full map of priced pins

The problem

Agents learned to search on their regional MLS. The Compass agent search was built off of the consumer experience instead.

That gap between their mental model and the product caused errors in how they input search criteria. The queries were incorrect, the results unexpected, so they kept going back to the MLS and missed out on the Compass core features.

And because search is the gateway to all other agent workflows, those workflows were starting on other platforms.

Zillow search results: a map of Brooklyn beside listing photos — Whether intentional or not, the MLS systems were designed in rather consistent ways, unlike a normal consumer facing search experience.

Redfin search results: a map of Brooklyn beside listing photos — Whether intentional or not, the MLS systems were designed in rather consistent ways, unlike a normal consumer facing search experience.

Paragon MLS search criteria: a long form of fields with operators — The forms agents knew, and the search bar they were given. As a result, the queries were incorrect, causing them to retrieve unexpected results.

Bright MLS search criteria: a dense form of fields and checkboxes — The forms agents knew, and the search bar they were given. As a result, the queries were incorrect, causing them to retrieve unexpected results.

A map of the United States covered in dots, one per MLS — There are 597 MLSs in the United States, each with its own criteria.

The approach

We started by listening. In 2018 we flew to SF for a research session with local agents. We learned that agents expect search to work like the tools they already know. And those tools change by region: New York agents were used to typing into one box, everywhere else to filter forms.

So the idea was to stop bending a consumer search to fit agents, and design around how they already work. I wrote it up as a design vision, plus a framework that broke the work into parts we could research, design, and build.

The team and I pitched that vision to the CEO, COO, and CPO. It was approved, funded, and became the largest project in the roadmap.

Four people smiling for a selfie by the water in San Francisco — 2 PMs, a researcher and a designer sat at a bar...

The first page of a document titled Agent Search Design Vision 2020, written by Yaron Schoen — The design vision: an inventory platform that makes surfacing relevant listing data, and taking action on it, flexible enough to accommodate every imaginable agent workflow, user preference, and regional need.

A table of four columns: entry points, inputs and criteria, results, exit points and actions — So many unknowns, we needed a framework in which we can uncover them. It divided the experience into groups that could be researched, designed, and built: entry points, criteria, results, and exit points.

How the parts connect: a workflow leads to an entry point, criteria and results feed each other, and it all ends in an action. An agent's region and preferences shape every step.

A low fidelity wireframe: search criteria beside a results table — The framework as flows and wireframes: entry, criteria, results, actions, exit.

A low fidelity wireframe: a results table beside a map — The framework as flows and wireframes: entry, criteria, results, actions, exit.

Filters first

The framework gave us four parts to work on. I suggested starting with filters, because that's how agents input their queries. If they can't ask the question their way, they won't see the right results.

The catch: the 597 MLSs had tens, even hundreds, of different filters each. One consolidated UI that accounted for all of them was impossible at the time.

So I proposed the opposite: replicate the filter panel per region, with labeling, grouping, layout, and interactions similar to the MLS. Agents' mental models were just too hard to break, so we designed for them instead. We launched region by region, starting with Boston.

A line chart: the share of DC agents using search per day over twelve months, jumping after the launch date — The DC launch: daily search use up by more than 30% while the rest of the company stayed flat, and searches per agent up by roughly 70%, above target. It was a huge success, prompting the company to significantly increase headcount focused on filters.

A line chart: average searches per DC agent per day against 2018 and a target, rising above both after the launch date — The DC launch: daily search use up by more than 30% while the rest of the company stayed flat, and searches per agent up by roughly 70%, above target. It was a huge success, prompting the company to significantly increase headcount focused on filters.

Draw the rest of the owl

This was a huge project, so I'll stop the case study here, before it turns into a book. Get in touch if you want the long version.

But with the filters in, the rest of the framework followed: results, customization, and the exit points into an agent's other workflows.

The experience was implemented in all regions on a rolling basis through user feedback and iteration, and adoption increased in every single one. In DC, daily search use rose by more than 30% in the weeks after launch, and searches per agent by roughly 70%.