How our exit-intent ML model actually works: a deep dive into LeadYup's ExitSense

Beyond the Mouse-Out: Why Legacy Exit-Intent Falls Short

For years, exit-intent technology primarily focused on detecting a user's mouse cursor leaving the browser viewport. While a step up from timed popups, this method has significant limitations. It's often inaccurate on touch devices, easily triggered by accidental movements, and doesn't account for genuine browsing behavior.

The Nielsen Norman Group has consistently highlighted that intrusive popups, even exit-intent ones, can create negative user experiences if not deployed intelligently. Our goal with ExitSense was to move past these rudimentary signals and build a more nuanced understanding of user intent.

The 26 Signals Our Popup ML Watches

LeadYup's ExitSense ML model doesn't just watch for a mouse leaving the window; it processes a rich tapestry of 26 distinct behavioral signals. These signals range from basic interactions to more complex patterns that reveal a user's true engagement level. Some examples include:

• Mouse movement speed and direction: Rapid, erratic movements can indicate frustration or an attempt to leave.
• Scroll depth and velocity: A sudden scroll-up after being deep on a page often precedes an exit.
• Time spent on page vs. average: Deviations from typical engagement patterns are key.
• Number of clicks and interactions: Low engagement despite significant time on page can signal disinterest.
• Tab switching frequency: A user cycling through tabs might be looking for alternatives.
• Form field interaction: Abandoned forms are a strong indicator of exit intent.

By combining these granular signals, the model builds a real-time probability score for a user's imminent departure, allowing for precise and effective popup timing. This multi-signal approach is a core part of how our exit-intent ML model actually works.

What We Learned from 10,000+ Popup Impressions 📊

Running LeadYup across thousands of sites has provided invaluable real-world data. We've observed that the average conversion rate for well-timed, relevant popups hovers around 3.09%, a figure consistent with findings from studies like Sumo's 2016 research. However, top-performing popups, often leveraging advanced timing and personalization, can exceed 9.28%.

One key experience-based observation: On the 1,000+ sites running LeadYup popups, exit-intent on mobile typically needs a scroll-up + idle hybrid because traditional mouse-out doesn't fire. Relying solely on desktop-centric triggers for mobile audiences drastically reduces effectiveness. This insight shaped how our ML model prioritizes mobile-specific behavioral cues.

We also learned that content relevance is paramount. Even with perfect timing, a generic offer will underperform a highly personalized one, which brings us to the role of AI in content generation and optimization.

What Modern AI/LLMs Add to How Our Exit-Intent ML Model Actually Works

The integration of AI and Large Language Models (LLMs) transforms how our exit-intent ML model actually works, moving beyond mere trigger detection to intelligent engagement. Here’s what modern AI/LLMs enable:

1. Per-Page Copy Generation: Unlike rule-based legacy tools that rely on static templates, LeadYup uses LLMs to generate unique, contextually relevant popup copy for each specific page a user is viewing. This means the offer and messaging are always tailored to the content, significantly boosting relevance and conversion potential.
2. Thompson Sampling for A/B Testing at Scale: Traditional A/B testing can be slow and resource-intensive, especially for SMBs. We employ Thompson sampling, an advanced Bayesian optimization technique, to intelligently explore different headlines and calls-to-action. This allows the system to quickly identify winning variations and allocate more traffic to them, even with fewer impressions, providing robust optimization capabilities that were once only available to large enterprises. This is a critical component of how our exit-intent ML model actually works.
3. Behavioral Signal Fusion via XGBoost: Our ExitSense ML model leverages algorithms like XGBoost to fuse the 26 behavioral signals. XGBoost is a powerful gradient boosting framework known for its efficiency and accuracy in handling complex datasets. It doesn't just look at signals in isolation; it understands their intricate relationships and weights them dynamically to predict exit intent with high precision, far surpassing what simple threshold-based rules can achieve.

Thompson Sampling Explained for Marketers: Smarter A/B Testing

For marketers, understanding Thompson sampling is key to appreciating the sophistication of modern popup optimization. Imagine you have several headlines for your popup. Instead of splitting traffic equally (A/B/C/D testing) and waiting a long time to see which performs best, Thompson sampling is more agile.

It starts by exploring all options, but quickly starts favoring the variations that show early signs of success. As more data comes in, it allocates an increasing percentage of traffic to the proven winners, while still giving less successful variants a small chance to prove themselves (to avoid getting stuck on a local optimum). This means your campaigns optimize faster, achieving higher conversion rates sooner, and continuously adapt to user behavior without manual intervention. It's a pragmatic approach to continuous improvement, ensuring that the most effective messages are always shown to the maximum number of visitors, enhancing how our exit-intent ML model actually works.

Roman Bootko

Founder & CEO, LeadYup

Roman has built lead-capture products since 2019, serving 1,000+ websites across 12 countries. He writes about exit-intent ML, popup conversion data, and the unsexy reality of growing SaaS from zero.