How our exit-intent ML model actually works: an honest critique for marketers

Beyond the Click: The 26 Signals Our ML Watches

When we first started building LeadYup, the conventional wisdom around exit-intent was largely reactive: detect a mouse moving towards the 'X' button or outside the viewport. While a good starting point, this approach misses a rich tapestry of pre-exit behaviors. Our how our exit-intent ML model actually works is built on observing 26 distinct behavioral signals, far beyond a simple mouse-out. These signals include micro-scrolling patterns, acceleration and deceleration of cursor movement, time spent on specific page elements, active vs. inactive tab status, and even idle time after significant interaction.

For instance, a user might scroll rapidly to the bottom of a pricing page, then quickly scroll back to the top before hovering over the browser's close button. Each of these actions contributes to a 'propensity to exit' score. The ExitSense ML model isn't just looking for one smoking gun; it's fusing these disparate signals to build a real-time probability curve of a user's imminent departure. The goal is to intervene at the precise moment a user is considering leaving, but before they've mentally checked out.

We learned early on that relying on a single signal, like a mouse moving out of the viewport, led to either premature popups that annoyed users or delayed ones that missed the opportunity entirely. The rich data set from these 26 signals allows for a much more nuanced prediction, significantly improving conversion rates compared to static, rule-based systems. This granular observation is key to delivering a timely, relevant message.

Thompson Sampling Explained for Marketers: Dynamic Headline Optimization

One of the persistent challenges in popup conversion rate optimization (CRO) is headline fatigue. What works today might underperform tomorrow, and what converts on one page might fail on another. This is where Thompson sampling shines for marketers, moving beyond traditional A/B testing's slower iterations and fixed traffic allocation. Instead of splitting traffic equally and waiting for statistical significance, Thompson sampling dynamically allocates more impressions to better-performing headlines as data accrues.

Imagine you have five headline variations for a popup. Thompson sampling doesn't commit to a fixed percentage for each. It acts like a curious explorer, trying each headline a few times, then gradually increasing exposure for those showing promise, while still occasionally testing underperforming ones to ensure it hasn't missed a trend or context shift. This 'explore-exploit' balance means that over thousands of impressions, your audience is predominantly seeing the best-performing headlines, while the system continuously learns and adapts.

For indie SaaS founders or SMB e-commerce owners with smaller traffic volumes, this dynamic allocation is a game-changer. You don't need millions of page views to find winning headlines. Thompson sampling allows for faster optimization and higher cumulative conversions even with moderate traffic, making it a powerful tool for maximizing the impact of every popup impression. It’s about more than just finding a winner; it’s about consistently serving the current best performer.

What We Learned from 10,000 Popup Impressions: The Good, The Bad, and The Mobile

Our data from observing over 10,000 popup impressions across various industries has yielded critical insights. The average conversion rate for popups, according to Sumo's 2016 study, was around 3.09%, with top performers reaching over 9%. Our own data aligns with this, showing that highly targeted, context-aware popups significantly outperform generic ones. The biggest 'bad' is undoubtedly poorly timed or irrelevant popups. Nielsen Norman Group's UX research consistently highlights user frustration with disruptive interfaces, and popups are no exception if not handled delicately.

One experience-based observation that stands out: On the 1,000+ sites running LeadYup popups, exit-intent on mobile typically needs a scroll-up + idle hybrid because mouse-out doesn't fire. Mobile users don't have a mouse cursor to 'exit' the viewport. Instead, signals like rapid scrolling upwards (indicating a potential 'back' gesture or trying to reach browser controls), combined with a period of inactivity, are far more reliable indicators of exit intent. Simply relying on back button presses often misses a significant portion of departing users.

What generally works? Popups with clear value propositions, concise copy (often generated per-page by our language model), and a single call to action. What doesn't? Overly complex forms, aggressive timing, or offers that don't align with the page content. Honesty about these tradeoffs is crucial for effective CRO. A well-designed how our exit-intent ML model actually works is about respect for the user experience, not just about capturing leads at any cost.

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

Modern AI, particularly Large Language Models (LLMs) and advanced machine learning techniques, fundamentally changes how our exit-intent ML model actually works compared to legacy systems. Rule-based popup tools from a few years ago were limited to IF-THEN statements: IF mouse leaves viewport, THEN show popup. This is static and brittle. LeadYup's approach is dynamically adaptive and personalized.

Firstly, per-page copy generation with a language model allows for hyper-relevant messaging. Instead of a single generic headline across an entire site, our LLM analyzes the content of the specific page a user is viewing and crafts a popup headline and body copy tailored to that context. This significantly boosts engagement because the offer feels personalized and timely. For example, a user on a 'pricing' page might see an offer for a 'demo' while a user on a 'features' page might get a 'download guide' offer.

Secondly, the integration of Thompson sampling for headline selection, as discussed, is an ML-driven approach to continuous optimization. Legacy systems often require manual A/B test setup and analysis, limiting the number of variations and slowing down learning. LeadYup automates this, ensuring that the best-performing creative is almost always displayed without manual oversight.

Finally, the core ExitSense ML model itself is a testament to modern AI. Leveraging techniques like XGBoost or similar ensemble methods, it can fuse the 26 behavioral signals with far greater accuracy than simple thresholds. This allows for probabilistic timing – not just 'yes/no' but 'how likely is this user to exit in the next 5 seconds?' – enabling a far more precise and less intrusive intervention. This level of predictive analytics is simply not possible with older, rule-based systems, offering a significant competitive advantage for LeadYup users.

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.