AI onboarding for Vrbo hosts
Hypothesis
Vrbo hosts have to write their own headlines and descriptions when listing a new property.
We know this causes friction for hosts and quality issues for Vrbo including:
non-compliant text that results in friction for hosts,
localization at scale is an issue when hosts create text that’s not localizable,
and new hosts feel intimidated writing headlines and descriptions to market their listings.
These obstacles don’t just affect Partners, EG Travelers rely on high-quality listings to make their travel decisions.
Our hypothesis is that by providing hosts with generated headlines and descriptions that include traveler-relevant data
My Role: Content Design Director - product concept, demo, and UX writing
Approach
Me and the PM pitched the SVP of Product on using ChatGPT to generate headline and description text for hosts as a test and learn experiment.
Our learning metrics:
% increase in host onboarding conversion
% increase of headline and description step completion
% of headlines and descriptions adopted with no edits
Demo
To get executive support and align the team quickly on the vision, I created a demo with our PM and UX Designer.
I wrote the story direction, script, and illustration direction for the demo
The demo not only touches on the base functionality of a headline and description feature, but future personalization opportunities using GPT in combination with EG Traveler data and shopping signals
LLM & evals
As the content design director, I worked with an ML and a CD on my team on the prompt and evaluation metrics for this feature including:
text classification for property descriptors
troubleshooting data modeling issues between onboarding data and output text
creating a definition of good for headline and description including training set examples
defining eval metrics and grading outputs on appropriateness, style, granularity, representativeness, accuracy, tone and more
We actually had to explore a structured content (NLP) intermediary step to categorize some of the onboarding data into categories the LLM could use more effectively for relevant outputs.