CPA Optimization Redesign

We launched two CPA-based ML models:

• Cost per Event Category — e.g., optimizing toward "Landing Page" as a category
• Cost per Action (per specific event) — optimizing toward a user-selected event

The DSP is flexible: events can be associated at both campaign and order level. That flexibility is powerful, but it can be confusing without structured guidance.

The initial experience allowed users to:

• Select CPA optimization
• Add multiple events for measurement
• Never explicitly associate what the model should optimize toward

Additionally, the legacy model optimized toward all tracked events, which diluted the signal and reduced efficiency. The system technically worked, but structurally encouraged misconfiguration.

Before redesigning the flow, I used SQL to understand real usage:

• Which events were used most frequently
• How often advertisers used campaign vs. order-level associations
• How often multiple measurement events were present without a primary optimization target

This analysis confirmed we should design for the most common patterns while preserving advanced flexibility.

I owned the end-to-end UX redesign:

• Mapped ML model requirements to configuration steps
• Identified fragmentation and signal-dilution failure points
• Designed a guided setup flow with conditional logic
• Partnered with engineering to ensure feasibility within existing constraints

Instead of redesigning visuals, I restructured decision sequencing to ensure the model always receives a valid signal during order creation.

1. If CPA per event category → require category selection + pixel selection (implicit when advertiser has a single pixel)
2. If CPA per specific event → require at least one optimization event selection
3. Render only the fields needed for the chosen goal: conditional + progressive disclosure
4. Ensure no extra post-creation steps are needed for effective optimization

The redesign made the dependency explicit:

• Measurement events are not the same as the optimization target
• CPA optimization requires a clear event signal at setup
• The flow enforces this requirement at the moment it matters

The solution required aligning UX logic with backend constraints:

• Conditional rendering tied to model selection and advertiser configuration (e.g., pixel count)
• Handling event hierarchy across campaign and order levels
• Validation states that prevent incomplete or ineffective configurations
• Preserving API contracts without expanding engineering scope