Most analytics teams report what happened. I built a system that predicts what will happen next — and identifies which visitors are worth targeting before they leave.

The Challenge: Data is Everywhere, but Intent is Hidden

Most e-commerce businesses look at Google Analytics 4 (GA4) with a focus on what happened: How many sessions? What is the bounce rate? What is the conversion rate?

While these metrics are important for reporting, they are reactive. As a data scientist with a background in statistical modeling, I believe the real value lies in being proactive — identifying a visitor’s intent while the session is still active to trigger the right intervention at the right time.

This post walks through a complete analytical pipeline I built on a high-fidelity synthetic GA4 dataset representing 25,000 sessions from a fashion e-commerce store. The pipeline includes:

  • Exploratory data analysis
  • Two-model propensity scoring architecture
  • K-Means behavioral segmentation
  • Segment-specific business recommendations

The Dataset: Synthetic but Realistic

Unlike toy datasets, this one includes granular behavioral signals that mirror real GA4 event streams:

Feature Group Variables
Engagement Scroll depth, session duration, pages viewed
Product Interaction Product views, wishlist adds, site search, promo clicks
Funnel Progression Add-to-cart, checkout steps reached
Context Device, traffic source, country, hour, day of week

Baseline conversion rate: 7.39%

Predictive Modeling: Two-Model Architecture

I trained two distinct Random Forest models to predict the likelihood of conversion. The reason for two models — not one — is deliberate and strategically important.

Model A: Full-Session Propensity

This model uses all available behavioral signals, including late-funnel variables such as add_to_cart and checkout_step_reached.

Purpose: Operational scoring — rank all sessions by conversion probability for post-session remarketing, email targeting, and audience creation.

Conversion Lift: Top 10% Propensity Group vs Site Average Figure 1: The top 10% of sessions ranked by propensity score converted at 24.16% — a 3.27x lift over the 7.39% site average.

Key results:

Metric Value
ROC-AUC 0.8721
Average Precision 0.5634
Top 10% Conversion Rate 24.16%
Top 10% Lift 3.27x
Top 20% Lift 2.68x

Top predictive features:

Feature Importance — Full Session Propensity Model Figure 2: Engagement score, add-to-cart behavior, and checkout progression dominate feature importance — confirming that funnel depth is the strongest conversion signal.

Model B: Pre-Checkout Propensity (The Strategic Model)

To enable earlier intervention — before a user adds to cart or reaches checkout — I built a second model that deliberately excludes all late-funnel signals:

Excluded variables: add_to_cart, reached_checkout, checkout_step_reached, checkout_progress_ratio

Purpose: Real-time personalization, smart popup triggering, promo eligibility scoring, and PDP optimization — all triggered before the user signals explicit purchase intent.

Key results:

Metric Value
ROC-AUC 0.5812
Average Precision 0.0998
Top 10% Conversion Rate 11.84%
Top 10% Lift 1.60x
Top 20% Lift 1.48x

The drop in predictive power compared to Model A is expected and honest. When late-funnel variables are removed, the model has less information. But it still identifies a top-decile audience converting at 1.60x the baseline using only pre-checkout behavior — which is commercially actionable.

“When late-funnel variables were removed, the model’s predictive power dropped as expected — but it still identified a top-decile audience converting at 1.6x the baseline rate using only pre-checkout behavior.”

Top pre-checkout signals:

  • engagement_score_precheckout
  • session_duration_sec
  • product_view_rate_per_page
  • session_duration_per_page
  • product_views
  • hour
  • scroll_depth_pct

Behavioral Segmentation: One Size Does Not Fit All

Using K-Means clustering (k=6, selected by silhouette score), I identified six distinct visitor segments based purely on behavioral features — without using the conversion label.

Customer Segments — Behavioral Segmentation Figure 3: Four key segments with distinct behavioral profiles and conversion rates. Each segment requires a different intervention strategy.

The key insight: not all non-converting sessions are the same.

1. Mobile Low-Intent Skimmers

  • Sessions: 9,323 (37% of traffic)
  • Conversion Rate: 4.46%
  • Mobile share: 91%
  • Behavior: Low product interaction, short sessions, minimal scroll depth

This is the largest segment and the weakest converter. The primary opportunity is not acquiring more traffic — it is reducing mobile friction.

Recommended actions:

  • Simplify mobile landing pages
  • Improve page load speed
  • Add sticky CTAs and trust badges
  • Test first-session micro-incentives

2. Desktop Window Shoppers

  • Sessions: 5,593
  • Conversion Rate: 6.94%
  • Behavior: Moderate engagement, desktop-dominant, no strong purchase signal

Recommended actions:

  • Browse abandonment email sequences
  • Recently viewed products widget
  • Wishlist nudges

3. Search-Driven Browsers

  • Sessions: 3,236
  • Conversion Rate: 6.03%
  • used_search = 1.00 (all users in this segment used site search)
  • Behavior: Clear intent, but not converting

These users know what they want but may not be finding it fast enough.

Recommended actions:

  • Onsite search UX audit
  • Autocomplete and synonym tolerance
  • No-results page optimization
  • Filter and facet improvements

4. Product Explorers / Cart Hesitators

  • Sessions: 2,614
  • Conversion Rate: 6.24%
  • product_views: 4.26 average
  • add_to_cart: 0.30
  • reached_checkout: 0.01

High product interest, low checkout progression. This is a high-value friction point.

Recommended actions:

  • Cart abandonment recovery
  • Shipping threshold communication
  • Exit-intent offers
  • Mini-cart UX improvements

5. High-Intent Buyers

  • Sessions: 2,620
  • Conversion Rate: 22.56%
  • Revenue per session: $33.23
  • add_to_cart: 1.00, reached_checkout: 1.00

The goal here is not persuasion — it is zero friction.

Recommended actions:

  • Streamline checkout to minimum steps
  • Optimize payment UX
  • Surface upsell and bundle offers
  • Prioritize checkout abandonment recovery

The Predictive Growth Engine: End-to-End Architecture

Predictive Growth Engine Workflow Figure 4: The full pipeline — from raw GA4 behavioral data through feature engineering, dual propensity models, and segmentation — to segment-specific business actions.

The architecture connects five layers:

  1. GA4 Behavioral Data → raw event stream
  2. Feature Engineering → session-level aggregation, engagement scoring, device/time flags
  3. Propensity Models → full-session (operational) + pre-checkout (real-time)
  4. Segmentation → K-Means behavioral clusters
  5. Business Actions → personalization, CRO prioritization, remarketing, checkout optimization

From Insights to Revenue: The Business Case

By targeting the top 20% of high-intent sessions identified by the model with a personalized intervention — a time-sensitive offer, dynamic social proof, or a smart nudge — a conservative 12% relative lift in their conversion rate generates meaningful incremental revenue.

In this 25,000-session sample:

Metric Value
Top 20% audience size 1,250 sessions
Observed conversions in top 20% 137
Extra conversions at +12% lift ~16
Average Order Value $143.53
Estimated incremental revenue ~$2,300

At production scale — 500,000 sessions/month — this translates to $46,000+ in incremental monthly revenue from a single model-driven intervention.

Summary: Data Science as a Growth Engine

Modeling conversion is not just about the AUC score. It is about resource allocation.

By knowing which sessions have the highest probability of conversion, businesses can:

  • Optimize ad spend toward high-intent audiences
  • Trigger onsite personalization only when it is statistically likely to work
  • Prioritize UX improvements based on segment-specific friction points
  • Reduce wasted retargeting budget on low-intent visitors

“Segmentation showed that not all non-converting sessions were the same. Some users were low-intent mobile skimmers, others were search-driven evaluators, and another group displayed strong product interest without progressing to checkout. This suggests that conversion optimization should be segment-specific rather than one-size-fits-all.”

The combination of a two-model propensity architecture with behavioral segmentation gives businesses a framework that is both statistically rigorous and operationally actionable — which is exactly where data science creates the most value.

What’s Next

In the next post, I will connect this pipeline to a real GA4 property using BigQuery export, build the feature engineering layer in Python, and demonstrate how to score live sessions in near real-time.

If you are working on a similar problem or want to discuss how this approach could apply to your business, I would be happy to connect.

Connect on LinkedIn