Expedite home discovery with Zillow Bot
There is only a 4% conversion rate in the home buying journey
In a typical home buying journey, the drop-off rate between exploration (researching the market, window shopping, etc.) to becoming more serious (saving/sharing listings, creating criteria, etc.), and then to logistic planning (contacting the agent/going on tours, applying for mortgages, etc.), becomes increasingly greater down the funnel.
What if we increase the conversion rate by combining exploration and serious stages so that potential buyers can expedite the journey by seeing homes they can realistically afford and love?
Results
• Delightfulness: 8 out of 10
• Intuitiveness: 9 out of 10
• Trustworthiness: 6 out of 10 (wary of AI and Zillow in general)
“This one is a lot better than the previous test. The questions were more straightforward and clear. The neighborhood selection experience makes a lot of sense having 4 of them comprising as one neighborhood vs before, when only 1 photo represented a neighborhood.
I can definitely see people who have no idea where to start to use this feature.”
My Role
Conversation designer
Product Designer
Prototyper
Timeline
3 months
September - December 2024
Scope
Conversation Design
AI & Machine Learning
UX/UI Design
User Research
Interaction Design
The Zillow Bot helps the user speed up exploration and research process by matching the user with their perfect neighborhood, searching for homes based on travel times, and calculating their budget.
This information helps the user narrow down their scope, as well as show personalized listings that accurately match their criteria, which turns into saves and eventually, conversions.
Success Metrics
• 30% of the homes shown by the Zillow Bot saved
• 20% increase in clicking on “Request a tour” or “Contact agent”
• 80% of users indicate high trustworthiness in customer satisfaction score (CSAT) for the bot
• 25% adoption rate for Zillow Bot ( = 50% of first-time home buyers adopting this)
Discovery started with defining the bot’s personality as well as understanding the feasibility of my potential use cases.
Part 1 - Bot Personality
Knowing that this bot is to be a companion for the user on the platform, its personality would be close to that of a real estate agent. But what kind of agent does a home buyer like?
But since this is Zillow, one of the main brand voices was delightful, so without a doubt it was incorporated too.
Part 2 - Use Cases
To create compelling use cases, the interaction goals should be efficient, flexible, accurate, and personalized, which were key words I uncovered during research for the bot’s personality
For this class project, we did have a constraint that the interaction must be mostly natural language understanding (NLU) based, and must require a couple of turn-taking
This automatically disqualified my ideas in using a chatbot to ask for listings using more unique criteria that aren’t already on Zillow, and generating pros and cons for listings based on the customer’s preferences to help them make a sound decision
Then I saw a BIG OPPORTUNITY IN FIRST-TIME HOME BUYERS (or new rental hunters) based on intuition after playing around with Zillow, and top competitors like Realtor.com, Redfin, and more.
Therefore, my main use cases are
Find a neighborhood match based on the user’s criteria
Search based on commute / travel times
Calculate budget & understand home-related costs
Knowing that image recognition, insight extraction, and data analysis can be utilized for my use cases, I began to explore the scripts.
To minimize conversation turns and drop-off rates, I opted to ask the most crucial questions relating to each use cases first in the most logical order.
01 Neighborhood Match
⚡️ Note
To incorporate neighborhood vibe check, I thought about using machine learning to target a couple of communities in the vicinity of potential neighborhoods based on the user’s criteria so far.
Then, based on the option they clicked on, more options will be shown again for the user to pick until 1 clear winner is shown.
02 Search by Commute / Travel Times
03 Calculate Budget & Home-Related Costs
⚡️ Note
I wrote 2 different scripts, one in which the user answers the bot directly in the chat, and another one linking to the BuyAbility feature since it already exists and is more robust. However, I learned from my test participants that they would prefer talking directly to the bot, and additionally, prevents them from abandoning the current experience.
Acknowledgment/Confirmation (As Needed) + Context + Cue is the anatomy of the bot’s prompt that makes the dialogue friendly and easy to comprehend, as well as ensures the flow’s efficiency in collecting the needed information from the user
Below is a sample repair mechanism that exemplifies the voice of the Zillow Bot:
Re-prompting the user to take action by being accommodating to include alternative ways to answer the question
These scripts, and eventually flow diagram, led to prototyping on Voiceflow using intent classification and prompt engineering to create output parsers that significantly reduced AI token usage for entity recognition from over 5,000 to less than 5 tokens per turn
The biggest takeaway from the 1st round of testing is to limit the ways a user can answer the bot to eliminate the number of possibilities that have to be accounted for, especially for an NLU-based bot
Implementing a semi-closed prompt from the bot gives the user a better idea of how to answer:
Additionally, I was really surprised that people STILL don’t read even when it’s a chat-based experience. Therefore, some prompts that ask for multiple inputs or are not easily comprehensible had to be improved.
It is also a good idea to explore ways to convince the user why a certain neighborhood is the perfect fit!
Had to include the city name since neighborhood names aren’t as well-known
Included a map to show where it is located geographically
What this neighborhood offers that matches the user’s criteria as disclosed during the conversation
The interface experience was only conceptualized after most of the conversations had been finalized to take into consideration the use cases to create a seamless, modular, and scalable user experience.
01 Homepage Experience
To strategically introduce natural language and AI into Zillow, I brought natural language search into the homepage search bar from the mobile app, and embedded intent keywords packaged as AI chat UI to serve as entry points into this feature.
Same goes for the search bar itself since I had assumed users will most likely notice this feature when trying to search for a home
02 Search Page Experience
Considering that a user may only want to search for the use cases designed for the bot after becoming frustrated while looking at the listings, entry points into the Zillow Bot are available in the search bar as well as a standalone button next to it
03 Zillow Bot Introductions
I designed 2 variations of Zillow Bot’s introduction in the chat itself depending on how the user accessed it. This ensures the bot does not repeat itself and cause confusion when the user comes in from clicking on the intent keywords shortcut versus selecting an intent within the chat box.
Next Steps
Improve the entry point to the chat experience on the UI front to accommodate for different search behaviors
Improve the conversations to allow the greatest flexibility for users who are able to communicate from the get-go all the criteria they have on their minds
Key Learnings
01 Dig deep and understand the inner workings of generative AI and machine learning
02 Connect that knowledge to serve as a basis for this product experience
03 Self-learn and continue through trial and error to figure out how to build an output parser using AI on Voiceflow, and utilize it as the logic step to carry on the conversation flow to ensure success and flexibility for the testing participants
Because of my relentless drive to solve this issue, I was invited by my lecturer to teach my class on this topic with the software :D