Propensity to Buy
“Which website visitors will make a purchase in the next 7 days?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which website visitors will make a purchase in the next 7 days?
E-commerce and SaaS companies drive millions of site visits, but fewer than 3% convert to a purchase. Marketing teams blast the same promotions to everyone, wasting ad spend on visitors who were never going to buy and under-investing in visitors on the verge of purchasing. Without visitor-level propensity scores, personalization engines, ad bidding, and on-site merchandising operate blind.
Quick answer
Propensity-to-buy models predict which website visitors will make a purchase within a defined time window (typically 7 days). The best models learn from browsing sequences, session depth, traffic source, and cross-visitor behavioral patterns rather than simple page-view counts. Visitor-level propensity scores lift conversion rates by 2.5x when used for personalized offers and real-time bid adjustments.
Approaches compared
4 ways to solve this problem
1. Rule-Based Triggers
Fire conversion offers based on rules: 'if cart page visited AND session > 3 min, show popup.' The default in most CRO tools (Optimizely, VWO).
Best for
Simple ecommerce sites with well-defined conversion funnels and limited traffic complexity.
Watch out for
Rules are static and one-dimensional. A visitor who spent 4 minutes reading a blog post triggers the same rule as one who spent 4 minutes comparing products on the pricing page. The intent is completely different, but the rule cannot distinguish them.
2. Logistic Regression on Session Features
Train a simple classifier on session-level features: pages viewed, time on site, traffic source, device type. Score each visitor by conversion probability.
Best for
Teams that want a quick ML baseline. Easy to implement and interpret. Runs fast enough for real-time scoring.
Watch out for
Limited to session-level aggregates. Cannot capture page-view sequences (pricing then cart is different from cart then pricing) or cross-visitor patterns (visitors from the same campaign converting at different rates based on landing page).
3. Gradient Boosted Trees on Behavioral Features
Train XGBoost on hand-crafted behavioral features: page view sequences, scroll depth, click patterns, time on pricing page, cart interactions. The current industry standard.
Best for
Teams with ML engineers and behavioral analytics infrastructure. Good accuracy when features capture the right behavioral signals.
Watch out for
Feature engineering is the bottleneck. Each new behavioral signal (scroll depth, hover patterns, cross-device sessions) requires a new feature pipeline. The model treats each visitor independently, missing cross-visitor signals like 'visitors from this campaign who viewed this product convert at 3x the base rate.'
4. KumoRFM (Graph Neural Networks on Relational Data)
Connects visitors, page views, and orders into a temporal relational graph. Learns from browsing sequences, cross-visitor patterns, and traffic source signals automatically. Scores update continuously as new page views stream in.
Best for
High-traffic ecommerce and SaaS sites where visitor-level propensity scores drive real-time personalization, bid adjustment, and on-site merchandising.
Watch out for
Requires page-level or event-level visitor data with timestamps. If your analytics only tracks session-level aggregates, the model has fewer behavioral signals to work with.
Key metric: Visitor-level propensity scores lift conversion rates by 2.5x. SAP SALT benchmark: 91% accuracy for multi-table relational models vs 75% for single-table approaches.
Why relational data changes the answer
Visitor V001 arrived from paid search, viewed a product page for 45 seconds, spent 120 seconds on the pricing page, and visited the cart. A flat model sees these as four features and scores the conversion probability. But the relational graph captures the sequence and context: the pricing-then-cart sequence is a stronger signal than cart-then-pricing. The 120-second pricing page dwell time is 2.4x the average, indicating serious comparison, not casual browsing. And other visitors from the same paid search campaign who followed this exact sequence converted at 4.1x the base rate.
These cross-visitor patterns are invisible to models that score each visitor independently. The graph neural network propagates information from converted visitors to unconverted ones with similar browsing patterns, traffic sources, and product interests. The WHERE clause in PQL filters to visitors with 3+ page views in the last 7 days, focusing on visitors with enough behavioral signal to score meaningfully. On the SAP SALT benchmark, relational models achieve 91% accuracy vs 75% for single-table models. For ecommerce propensity scoring, the relational advantage translates directly to higher conversion rates: 2.5x lift when propensity scores drive personalized offers, retargeting bid adjustments, and real-time on-site merchandising.
Scoring visitors with session-level features is like a store greeter deciding who is likely to buy based only on how long they have been in the store. A relational model is like a seasoned sales associate who notices the visitor is carrying a competitor's shopping bag (traffic source), went straight to the premium section (page sequence), and resembles the customers who bought during last weekend's promotion (cross-visitor patterns). The time in store matters, but the relational context is what drives the recommendation.
How KumoRFM solves this
Relational intelligence for smarter acquisition
Kumo ingests VISITORS, PAGE_VIEWS, and ORDERS into a temporal relational graph. The model learns sequences and cross-entity patterns — like 'visitors who viewed 5+ pages including pricing, from a paid source, within a session that lasted over 4 minutes' — and combines them with relational signals from other converting visitors. The WHERE clause filters to visitors with recent engagement, ensuring predictions are actionable. Scores update continuously as new page views stream in.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
VISITORS
| visitor_id | source | device | first_seen |
|---|---|---|---|
| V001 | paid_search | desktop | 2025-11-10 |
| V002 | organic | mobile | 2025-11-11 |
| V003 | desktop | 2025-11-12 | |
| V004 | direct | tablet | 2025-11-12 |
PAGE_VIEWS
| view_id | visitor_id | page_url | duration_sec | timestamp |
|---|---|---|---|---|
| PV01 | V001 | /product/shoes | 45 | 2025-11-10 |
| PV02 | V001 | /pricing | 120 | 2025-11-10 |
| PV03 | V001 | /cart | 30 | 2025-11-11 |
| PV04 | V002 | /blog/guide | 90 | 2025-11-11 |
| PV05 | V003 | /product/jacket | 60 | 2025-11-12 |
| PV06 | V003 | /pricing | 85 | 2025-11-12 |
ORDERS
| order_id | visitor_id | amount | timestamp |
|---|---|---|---|
| O801 | V001 | $149 | 2025-11-12 |
| O802 | V003 | $225 | 2025-11-14 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT COUNT(ORDERS.*, 0, 7, days) > 0 FOR EACH VISITORS.VISITOR_ID WHERE COUNT(PAGE_VIEWS.*, -7, 0, days) > 3
Prediction output
Every entity gets a score, updated continuously
| VISITOR_ID | TIMESTAMP | TARGET_PRED | True_PROB |
|---|---|---|---|
| V001 | 2025-11-10 | True | 0.92 |
| V002 | 2025-11-11 | False | 0.08 |
| V003 | 2025-11-12 | True | 0.79 |
| V004 | 2025-11-12 | False | 0.15 |
Understand why
Every prediction includes feature attributions — no black boxes
Visitor V001 — paid_search / desktop
Predicted: True (92% probability)
Top contributing features
Visited cart page within 24 hours of product view
True
32% attribution
Time on pricing page > 90 seconds
120 sec
26% attribution
Source — paid_search (highest-converting channel)
paid_search
20% attribution
3+ page views in last 7 days
3 views
14% attribution
Desktop device (higher AOV segment)
desktop
8% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about propensity to buy
What conversion rate improvement can propensity scoring deliver?
Visitor-level propensity scores typically lift conversion rates by 2-3x when used for personalized offers and real-time bid adjustments. The improvement comes from concentrating marketing spend on high-probability converters and personalizing the experience for visitors on the cusp of purchasing.
Can propensity models work for anonymous visitors?
Yes. Graph models score visitors based on their behavioral patterns (page sequences, session depth, traffic source) without requiring login or identity. The model learns from the full visitor-page-view-order graph, and anonymous visitors with strong behavioral signals can score just as high as known users.
How fast do propensity scores need to update?
For real-time personalization and on-site merchandising, scores should update with each page view (sub-second latency). For retargeting bid adjustment and email personalization, hourly or daily updates are sufficient. The highest ROI comes from real-time scoring because the conversion window for most visitors is minutes, not days.
What is the relationship between propensity scoring and retargeting?
Propensity scores make retargeting dramatically more efficient. Instead of bidding the same amount for all site visitors, you bid high for visitors with 80%+ propensity (who need one more nudge) and bid low or exclude visitors with under 10% propensity (who were never going to convert). This typically reduces retargeting cost-per-acquisition by 40-60%.
Bottom line: Visitor-level propensity scores lift conversion rates by 2.5x when used for personalized offers, retargeting bid adjustments, and on-site merchandising — turning anonymous traffic into attributable revenue.
Related use cases
Explore more acquisition use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
