Stay AI
2026
Increasing save rates 4.5x through offer personalization
I rebuilt an acquired retention tool’s editor from scratch so merchants could see, trust and tailor their cancellation flows. Save rates rose 4.5x.
Overview
Deals closing, merchants switching, save rates climbing
Stay AI acquired RetentionEngine, a tool which let merchants build a survey to show customers trying to cancel offers like discounts or changing products an utilized machine learning to identify the highest performing offers. It was acquired for the machine learning, however the tool itself was difficult to use and was a separate platform to maintain. I owned the redesign end to end, from research through launch and migration, and defined the user goals no one else had set.
The Impact
Average save rates rose from 8.67% to 38.97% for merchants who switched. 90% of merchants migrated within 90 days, and Sales now points to Cancellation Surveys in deals worth $626.9k ARR.
Problem
The machine-learning wasn’t the problem. The editor was.
Business Problem
Stay was paying to host and maintain two platforms, and losing deals to competitors whose tools looked better in demos. Leadership set goals for cost and competition. Generally it was known that we should make the tool 'better' too, but no one had set a save-rate target or even pulled an average.
User Problem
Merchants had two logins and an editor with a hidden preview state, confusing terminology, and a logic system that was difficult to trace. Only 7.5% of editing sessions ended in a saved change. A PM had logged 18 types of recurring complaints from merchants.
Before: hidden preview, single survey, and difficult to trace logic.
Research
Identifying themes across research methods
I wanted to understand RetentionEngine from every angle, so I pulled from multiple sources.
I reviewed our three main competitors screen by screen, along with CX tickets, Slack threads, a Claude + Gong scan of CX calls, and 60 days of FullStory sessions and Intercom tickets. I then sorted every issue into UX fixes, platform constraints, and future requests.
Every one pointed to the same set of problems: merchants couldn't properly see what they were building, they kept running into limits, and so they left under-performing surveys alone.
Competitive analysis of direct market competitors and similar tools in different markets.
Findings
The tool was confusing, but it's rigid limits also frustrated merchants.
The system was a black box.
The flow editor didn't show which path a customer would actually take. Merchants got lost moving between side-panel edit states. They couldn't see how the ML picked offers. Some didn't believe new offers were being tested fairly.
Edits didn't feel safe.
The preview state was hidden and there was no clear confirmation that a change had saved. Only 7.5% of editing sessions ended with a save, and the median took 9+ minutes.
The tool was too rigid.
Merchants were limited to showing one offer before cancel and setting conditions was hidden and repetitive.
Subscribers felt it too.
Subscribers hit dead ends and generic messages because merchants couldn't figure out how to personalize what they saw. A discount means nothing when you said you didn't like the product.
I turned these findings into three user goals: merchants should see what the subscriber sees, tailor offers to cancel reasons and user groups, and make the resulting tailored paths easily traceable.
Prototyping & User Testing
I tested my assumptions and was surprised
I expected a guided form to beat a flow builder, because an open canvas seemed like it would cause choice paralysis and our user base was already fluent in other form based tools in our app. I built both as prototypes and ran an Insiders session with 4 merchants, pairing the prototypes with a quiz on how subscribers would move through each flow.
Testing a flow builder against a form, with a quiz on subscriber outcomes.
What testing changed
All four merchants chose the flow builder and they predicted subscriber outcomes with complete accuracy. They wanted conditions inside each node rather than as separate steps. They also asked to see one cancel reason at a time. I’d revisit that call, because the prototype couldn’t pan or zoom, which skewed the feedback.
Designs
Visualize the path a user will travel with confidence
The final designs addressed each of the core user problems: visibility, trust in edits, and flexibility.
Merchants can now show multiple offers in succession when one is declined, segment flows with AND/OR conditions so that different user groups see different offers. We also grew the offer types from 2 to 8, including swaps, skips, delays and education.
Libraries of every screen type, so merchants have a bird's eye view of what they have made and they see exactly what customers see.
Merchants could now see a live preview side by side with the inputs as they created offers and pages for their survey. No flipping back and forth, the preview is always right there.
Simple controls replaced CSS, so exact brand details carry to every offer and page.
The survey went from a modal to full screen, something merchants had consistently asked for. Every question and offer is written by the merchant, so the base design had to be flexible and minimal.
Outcomes
Merchants built better surveys which saved more subscribers
Business Outcomes
Both business goals were met: 90% of merchants migrated within 90 days, and RetentionEngine was retired. Cancellation Surveys became the #1 source of positive sentiment in sales calls, and in the following six months 31 closed deals worth $626.9k ARR cited it as a top-3 reason for choosing Stay AI.
User Outcomes
Within 30 days of launch, merchants were building better surveys, and building them faster. Six months after launch, the average save rate had risen from 8.67% to 38.97%, and editing sessions ending in a save went from 7.5% to 33%.
4.5x
Save rate, six months post-launch
4.4x
Editing sessions ending in a save
90%
Merchants migrated in 90 days
$626.9k
ARR from 31 closed sales deals
