How our exit-intent ML model actually works: A Candid Look Under the Hood for 2026

Q1: What exactly is 'exit-intent' and why is it so effective?

Exit-intent is a website visitor's predicted intention to leave a page without completing a desired action. It's effective because it presents a last-chance offer – a discount, a lead magnet, or a subscription prompt – at the moment a visitor is disengaging. Traditional research, like Sumo's 2016 study, found that the average popup conversion rate is around 3.09%, with top performers reaching over 9.28%.

The power lies in context: you're not interrupting an engaged user, but rather re-engaging someone who was already on their way out. This makes the intervention less intrusive and more likely to be perceived as helpful, especially when the offer is relevant. By understanding how our exit-intent ML model actually works, marketers can leverage this moment effectively.

Q2: What are the 26 signals our popup ML watches to predict exit?

Our ExitSense ML model doesn't just look for a mouse moving toward the browser's close button. It actively monitors 26 behavioral signals, creating a comprehensive profile of a user's engagement level. These signals range from basic cursor movements to more subtle interactions.

• Mouse/Pointer Dynamics: Speed, acceleration, trajectory towards the top of the browser window, sudden changes in direction, and 'jitter' (small, erratic movements).
• Scroll Behavior: Upward scrolling velocity, repeated scrolling up and down, sudden stops in scrolling, and the total scroll depth reached.
• Engagement Metrics: Time spent on page, idle time (no activity), tab switching (focus changes), and interactions with page elements (clicks, hovers).
• Form Interaction: Abandoned form fields, typing speed, and cursor position within form inputs.
• Session Context: Number of pages visited in the current session, referrer information, and device type.

These signals are fed into our machine learning model, which learns to identify patterns indicative of an impending exit, far beyond what simple rules can achieve.

Q3: What did we learn from 10,000 popup impressions and beyond?

Analyzing data from tens of thousands of popup impressions across diverse client sites has provided invaluable insights. We've seen that the 'perfect' timing isn't universal; it varies significantly by industry, audience, and even the specific page content. For instance, on e-commerce sites, a quick 'back-to-cart' popup upon exit intent can recover significant revenue, while for SaaS trials, a content upgrade or demo offer performs better.

One critical observation: 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 as reliably on touch devices. This led us to refine our mobile exit-intent detection significantly. We also learned that overly aggressive popups, even with perfect timing, can lead to negative user sentiment if the offer isn't genuinely valuable, reinforcing Nielsen Norman Group's UX guidelines on modal dialogs.

Q4: How does Thompson sampling pick winning headlines for these popups?

Once the ExitSense ML model identifies an exit-intent moment, the next challenge is presenting the most effective message. This is where Thompson sampling comes into play for selecting winning headlines, offering a more efficient alternative to traditional A/B testing, especially for SMBs and indie SaaS founders with less traffic.

Thompson sampling is a 'multi-armed bandit' algorithm. Instead of rigidly splitting traffic (like A/B testing) and waiting for statistical significance, it dynamically allocates more impressions to the headlines that are performing better, faster. It explores new options while simultaneously exploiting the best-performing ones. This means that over time, the algorithm automatically prioritizes the headlines with the highest conversion probability, maximizing your results from day one, rather than waiting for a test to conclude. It’s how our popup builder ensures continuous optimization.

Q5: What modern AI/LLMs add to how our exit-intent ML model actually works

Modern AI and Large Language Models (LLMs) significantly enhance LeadYup's approach to how our exit-intent ML model actually works, moving beyond the capabilities of legacy, rule-based tools. Here’s what sets us apart:

1. Per-Page Copy & Headline Generation: Traditional tools require manual copy creation. LeadYup leverages LLMs to generate highly relevant, per-page popup copy and headlines. The language model analyzes the page content and user intent to suggest compelling text tailored to that specific context, drastically reducing manual effort and improving relevance.
2. Adaptive Behavioral Signal Fusion via XGBoost: Our ExitSense ML model uses advanced algorithms like XGBoost to fuse the 26 behavioral signals. Unlike simple 'if-then' rules, XGBoost can identify complex, non-linear relationships between signals and adapt its predictions in real-time based on new data. This allows for far more accurate and nuanced exit-intent detection.
3. Thompson Sampling for SMB Scale: As mentioned, Thompson sampling makes continuous optimization viable even for sites with moderate traffic. LLMs assist here by generating a diverse initial set of headlines for the bandit to test, ensuring a strong starting point for optimization. This means smaller businesses can achieve sophisticated A/B testing results without needing massive traffic volumes or dedicated CRO specialists.

These AI-driven capabilities mean popups are not just timed perfectly, but also personalized and optimized for maximum impact automatically.

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.