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

The Limitations of Legacy Exit-Intent

Traditional exit-intent popups often rely on simple mouse-out events, primarily detecting when a user's cursor leaves the browser viewport. While effective to a degree, this method has significant limitations. For instance, on mobile devices, there's no cursor, rendering mouse-out detection useless. Moreover, even on desktop, a user might accidentally swipe their mouse off-screen without intending to leave, leading to a premature and potentially annoying popup.

According to Sumo's research, the average conversion rate for popups hovers around 3.09%, with the top 10% achieving 9.28% or more. The gap between average and top performance often lies in the sophistication of the trigger mechanism. Simply put, a poorly timed popup can be as detrimental as no popup at all, disrupting the user experience and potentially increasing bounce rates.

Unpacking ExitSense: The 26 Signals Our ML Watches

At the core of LeadYup's ExitSense model are 26 distinct behavioral signals. These signals range from subtle mouse movements to scroll patterns, typing activity, and even idle time. We don't just look for one 'tell' of departure; instead, our machine learning model, often a gradient-boosted tree like XGBoost, synthesizes these signals in real-time to build a probabilistic score of user intent.

Some key signals include:

• Mouse Velocity & Direction: Rapid movement towards the top of the browser window, especially after a period of inactivity, is a strong indicator.
• Scroll Depth & Speed: A sudden, rapid scroll to the top or bottom of a page, or a lack of scrolling after a significant time on page, can signal disengagement.
• Tab Switching & Focus Changes: Has the user clicked away to another tab, or has the browser window lost focus?
• Form Field Interaction: Abandonment of a partially filled form is a clear sign of potential exit.
• Idle Time: While a simple signal, extended idle time, especially on content-heavy pages, can precede an exit.

On the 1,000+ sites running LeadYup popups, our team has observed that exit-intent on mobile typically needs a scroll-up + idle hybrid because mouse-out doesn't fire. This insight informed the development of specific mobile-centric signals within ExitSense, moving beyond desktop-centric assumptions.

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

The integration of modern AI and large language models (LLMs) fundamentally transforms how our exit-intent ML model actually works, distinguishing LeadYup from older, rule-based systems. Here’s what sets us apart:

1. Per-Page Copy Generation: Unlike legacy tools that require manual copy creation, LeadYup uses LLMs to generate highly relevant, per-page popup copy. The language model analyzes the content of the specific page a user is on and crafts compelling headlines and body text, dramatically increasing personalization and conversion potential.
2. Thompson Sampling for Dynamic A/B Testing: For SMBs and even agencies, traditional A/B testing can be slow and resource-intensive. LeadYup employs Thompson sampling, an advanced Bayesian approach, to dynamically pick winning headlines and offers in real-time, even with limited traffic. This ensures that the most effective variations are shown more frequently, accelerating optimization without manual intervention. This is crucial for understanding how our exit-intent ML model actually works efficiently.
3. Sophisticated Behavioral Signal Fusion: Rule-based systems often use simple 'if-then' logic for a handful of signals. LeadYup's ExitSense ML, powered by advanced algorithms, can fuse all 26 behavioral signals to predict intent with much higher accuracy. This complex signal fusion via techniques like XGBoost allows for nuanced understanding of user behavior, minimizing false positives and maximizing conversion opportunities, which is a key part of how our exit-intent ML model actually works.

What We Learned From 10,000+ Popup Impressions

Through analyzing millions of user interactions across our platform, we’ve gathered significant insights into effective popup strategies. One critical learning is that the 'perfect' timing isn't universal. What works for an e-commerce store with high-impulse purchases may not work for a B2B SaaS signup process.

We also observed that intrusive, full-screen takeovers often lead to higher bounce rates if not perfectly timed. Conversely, a subtle, well-timed slide-in that offers genuine value (e.g., a relevant discount code or a content upgrade) tends to perform better in terms of long-term engagement and conversion without irritating users. Wisepops' industry benchmarks consistently show that relevancy and timing are paramount for popup success.

Another key takeaway is the power of personalized copy. A generic 'Sign Up' popup converts significantly less than one that addresses the user's current page context, a capability directly enabled by LeadYup's LLM-driven copy generation. This reinforces the need for dynamic, adaptable solutions when considering how our exit-intent ML model actually works in practice.

Honest Tradeoffs: What Doesn't Work (and Why)

No technology is a silver bullet, and understanding the tradeoffs is essential. Relying solely on a single behavioral signal, no matter how strong, often leads to suboptimal results. For example, triggering a popup purely on 'mouse moving towards the close button' can be too late, as the user has already made the decision to leave. The strength of ExitSense lies in its predictive, multi-signal approach, aiming to intervene before that decision is firm.

Another common pitfall is ignoring frequency capping. Even the most perfectly timed popup can become annoying if it appears on every page visit. Our system incorporates intelligent frequency capping to ensure a positive user experience. Lastly, poorly designed popups — those with illegible fonts, irrelevant offers, or confusing calls to action — will underperform regardless of how sophisticated the exit-intent trigger is. UX best practices, as highlighted by Nielsen Norman Group, are still foundational.

LeadYup Editorial

Product & growth team

Hands-on operators behind LeadYup's popup engine, ExitSense ML model, and A/B infra. We write what we ship, not what we wish.