Keller Williams /
Using AI to personalize the online home-finding experience.

Objective: Reimagine what finding a home online could look like in the future.

timeline & Deliverables
  • Timeline: 3 weeks
  • Deliverables: High fidelity proof of concept, and a pitch deck.
Jump to the solution
Team
  • Cameron Koczon - Creative Director
  • Myself - UX Designer
01
Challenge

For the past 15 years, real estate company websites have looked the same.

Our challenge was to think beyond just enhancing a standard real estate website and create a vision for what finding a home online could look like in the near future.

All major real estate companies use the same configuration for their search page: a search field and filters above a split screen of listings in a grid and a map.

02
Discovery

Several of my colleagues at Fictive Kin had gone through the home buying process recently, so we got started by sharing our experiences using real estate websites and quickly uncovered some universal pain points.

"I don't want to miss a new listing, so I go to the site and search at least once a day, sometimes multiple times."
"Email notifications from my saved searches never make me feel like I haven't missed something. I always end up back on the site."
"I change up my search parameters as I'm browsing because I'm not looking for something specific. There's some fuzziness."
"I wish I could hide the listings I've seen and am not interested in so that I never have to see them again."

It was clear to us that the traditional real estate website search method is broken in two specific ways:

Endless pages of listings creates anxiety.

The urge to make sure you've seen every listing stems from a lack of confidence in the search results.

It lacks the control and fluidity users want.

Using too many filters returned very few listings, but not using enough meant hours of sifting through the pile.

03
Process

We began by investigating the search-centered experience.

The anchoring feature of real estate websites is the ability to save a search. Saved searches are meant to reduce labor of the user by allowing them to quickly repeat a set of search parameters.

However, we knew from our group discussion that saving a search didn't do much in terms of reducing labor and that it possibly created more labor.

We mapped out a search-centered experience to get a better look:

The part of the experience where the user spends the most time is browsing listings—a tedious experience with or without the back-and-forth action of adjusting your filter settings.

Saving a search merely allows you to exactly repeat the tedious experience again at another time.

Can we solve any problems with a better interface?

Our ultimate goal was to mitigate the lack of control and fluidity, which meant we needed to accommodate users who have very specific needs as well as those who are all-around flexible or are still figuring out their needs.

We listed example search parameters of the two extremes on a spectrum of flexible to constrained:

With these constraints and flexibilities in mind, I took a shot at creating a new search/browse interface:

  1. If searching in multiple cities, quickly jump between each one via tabs.
  2. Vertical search panel allows us to use 1-click settings like toggles and sliders.
  3. Like or hide a listing. Liking a listing allows you to immediately share it with your search partner or realtor.
  4. Adjustable layout options that allow for larger listing cards reduces the need to click into the listing page.

The UI exploration gave the user control and ease in the search experience, but we felt there was potential to reduce the labor of the user in the browse experience.

Browsing is an action we take nearly every day, so we looked at the products we browse through most often.

One major insight we found was that a lot of the time the browse experience isn't proceeded by the search experience—we just open the app and browse.

AI has changed what we expect from online experiences.

Behind all of our favorite services is a profile of us being quietly built in the background.

04
Solution

To deliver an experience that was both expected and delightful, we had to get personal with the users.

We wanted the experience to center around browsing, not searching. This meant that in a personalized home search experience the user would only need to configure their search once. 

After that, Keller Williams must get to know each individual user, learn what they are looking for beyond just their constraints and flexibilities, and surface listings that matched their needs and taste.

We mapped out what a person-centered experience might look like:

With a personalized dashboard at the center of the experience, relevant listings are brought to the user, lightening the effort and creating a more enjoyable and efficient experience.

To teach a passive artificial intelligence system which listings the user wants to see, we need to provide the user opportunities to interact with the AI.

Now that we knew we wanted a personalized dashboard, we had to figure out how it would be populated. We started by defining how the user would train the algorithm, because that would inform what content we had to work with and how it could be organized.

Post-search, the user trains the AI by telling it exactly what it wants to see, through:

  1. Liking and hiding listings
  2. “Following” specific home features
  3. Answering direct preferential questions

Liking and Hiding Listings

"Inbox zero" concept

"Following" specific home features

Examples of surfacing followable features
A home style often has several home features within it that a user may not follow. For example, a user loves Mid-century modern-style homes, partially because they love floor-to-ceiling windows. But because any home—regardless of the style—may have these kinds of windows the user can follow the “Floor-to-ceiling Windows” home feature to be shown any home with this kind of window treatment.

Features are mapped onto a photo collage with dots. Clicking on a dot reveals the feature and a follow button.

Answering Preferential Questions

Example of a multiple choice question sequence
We can ask the user if or how much they like something to know more about what to show them. In this example, someone can say it “depends” on something whether they like it or not. Now we know it isn’t a dealbreaker and can show listings with that feature.

The more the AI learns about a home buyer's preferences, the more intelligent its suggestions will be.

The culmination of AI training and collected data results in hyper-personalized modules that predict what a user might want to see or add to their search.

Smart suggestions come in the form of:

  1. Bespoke collections of homes
  2. Homes that are perfect, but slightly above price range.
  3. Home features they're likely to want.
  4. Surrounding cities where they have a high chance of finding a home match.

These modules can be mixed in with the training modules to make up the content of the personalized dashboard.

Smart Suggestions

AI-generated modules based on user behaviors
Collections can also be made by mixing two home features together. For example, “Homes with a rooftop deck and a finished basement” or “Rooftop decks with a view”.

In this style, listings in a collection are presented one-by-one, forcing the user to make a binary decision before seeing the rest of the collection.
Like Amazon’s feature “These items are often bought together”, we can show a feature that is followed by other users who also follow a feature this user follows. For example, I follow “Good Views”, and many users who also follow “Good Views” follow “Rooftop Decks” as well, therefore I might want to follow it, too.
After learning a little about what a user is looking for, we can suggest they look into nearby cities they may be able to commute from. This works particularly well if a user is searching in a major city.

Sometimes users are looking in smaller cities where “commute from” doesn’t apply very well. In this case we can show the user cities that may be a little further away (2-5 hour drive) that share characteristics with their searched city.

Filling in the rest

We all knew from experience that searching for a home is more than just browsing listings. It’s also learning about neighborhoods, keeping a close eye on the market, picking up on real estate terms and learning the process, and going to open houses and first-time buyer workshops.

We brainstormed other modules that we could intersperse into a user’s feed that would round out the experience and further build the user’s relationship with the Keller Williams brand.

Housing Market Insights

Examples of interactive modules
We can use the average days on the market and the average amount above or below asking price to determine if a city's market is a Seller's market or a Buyer's market. This can help the user know if the market is competitive and if it is a good time to buy.

With additional data, we can offer forecasting in the way Kayak forecasts whether the price of a plane ticket may drop soon.

Realtor Engagement

Examples of supplemental modules

The personalized dashboard changes as the user interacts with it, eliminating the need to repeat a search or adjust filter settings.

Each time a user logs in, they'll see a stack of new listings, market updates, and various interactive modules that make smart suggestions and collect insights based on what the user likes, follows, or hides.

Final Thoughts

As someone who is excited by AI and all of its potential, I loved working on this blue-sky project for Keller Williams. I'm delighted by the recommendations I see every time I open my Spotify app and or log onto Netflix, and it was so fun see how that kind of algorithm could be applied to a real estate website.

Something I was thinking about while working on this, and something I'm still unsure of, is the societal effect an algorithm can have when it plays a key role in helping someone find a home.

Every major city in the U.S. faces the issue of gentrification and rising home prices, and without careful testing and implementation, an "innocent" personalization algorithm could possibly have unintended effects.

Go back home