What You'll See in the Dashboard
On paid plans, the Intelligence tab at einstein.clickstream.com shows real-time behavioral scores for active visitors. This series explains what each score measures, what it means for your business, and how it works under the hood.
Why Behavioral Scoring at the Edge?
Traditional analytics tools collect events and ship them to a server for batch processing. By the time insights emerge, the user is long gone. ClickStream takes a fundamentally different approach: every behavioral model runs inside a Cloudflare Worker, scoring user behavior in real time at the nearest edge node as each event batch arrives -- the full 26-model pass is held under 3 ms p95 per event by a CI-enforced benchmark.
This architecture yields three critical advantages:
- Latency: Scores are available before the next page render, enabling real-time personalization and intervention.
- Privacy: PII is scrubbed server-side at the edge before storage. Analytics storage holds scrubbed events and computed scores with hashed identifiers only.
- Cost: Edge compute is dramatically cheaper than maintaining GPU clusters for real-time ML inference.
The 26 Behavioral Models
ClickStream computes 26 distinct behavioral scores for every active session. Most scores are normalized 0–100 values (session momentum runs -100 to +100; emotional state and next action are categorical), updated incrementally as new events arrive. Together, they form a comprehensive behavioral signature that powers personalization, alerting, and predictive analytics.
| # | Model | What It Measures | Series Part |
|---|---|---|---|
| 1 | Intent Score | Purchase/conversion likelihood based on navigation patterns | Part 1 |
| 2 | Frustration Score | User friction signals (rage clicks, dead clicks, error encounters) | Part 1 |
| 3 | Engagement Score | Depth and quality of interaction with content | Part 1 |
| 4 | Value Estimator | Predicted monetary value of the session/visitor | Part 2 |
| 5 | Anomaly Score | Statistical deviation from normal behavioral patterns | Part 2 |
| 6 | Confusion Score | Navigation confusion and information-finding difficulty | Part 3 |
| 7 | Emotional State | Inferred emotional valence from interaction dynamics | Part 3 |
| 8 | Decision Confidence | How confident the user appears in their purchase decision | Part 4 |
| 9 | Regret Risk | Likelihood of post-purchase regret or return | Part 4 |
| 10 | Churn Prediction | Probability of visitor/customer churn | Part 5 |
| 11 | LTV Score | Predicted lifetime value based on behavioral signals | Part 5 |
| 12 | Abandonment Score | Real-time cart/form/page abandonment detection | Part 6 |
| 13 | Purchase Timing | Proximity to purchase decision and urgency signals | Part 6 |
| 14 | Content Affinity | Topic and content-type preferences | Part 7 |
| 15 | Form Friction | Field-level form interaction difficulty | Part 7 |
| 16 | Next Action Prediction | Most likely next user action with confidence score | Part 7 |
| 17 | Scroll Depth Intelligence | Scroll regression, content zone engagement, reading patterns | Part 9 |
| 18 | Session Quality | Overall quality and substance of the session's engagement | — |
| 19 | Conversion Readiness | Real-time likelihood of completing a conversion goal | Part 9 |
| 20 | Bot Detection Score | Behavioral biometrics-based bot detection | Part 10 |
| 21 | Navigation Pattern | Classification of how the visitor moves through the site | — |
| 22 | Return Visitor | Return likelihood based on session patterns and engagement | Part 10 |
| 23 | Campaign Response | Behavioral response to the campaign or source that brought the visitor | — |
| 24 | Device Engagement | Engagement patterns specific to the visitor's device class | — |
| 25 | Time-of-Day Affinity | Visit-timing patterns and time-of-day preferences | — |
| 26 | Loyalty Trajectory | Loyalty and return-propensity trend across sessions | Part 10 |
Alongside the 26 models, the scoring orchestrator also computes three inline metrics on every pass: session momentum, navigation entropy (Shannon entropy of page-transition patterns), and attention. Part 8 covers all three.
The Three-Phase Pipeline
Every incoming event flows through a three-phase pipeline before scores are updated. This pipeline is designed for incremental computation -- no event requires rescanning the full session history. (The RawEvent, BehavioralFeatures, and SessionContext interfaces below are simplified illustrations of the pipeline's logical shape, not the shipped platform types.)
Phase 1: Event Ingestion & Normalization
Raw browser events (clicks, scrolls, mouse movements, form interactions, page transitions) arrive as JSON payloads at the edge worker. The ingestion phase normalizes timestamps, deduplicates rapid-fire events, and enriches each event with session context.
Phase 2: Feature Extraction
Normalized events are fed into the feature extraction layer, which maintains running statistics and computes derived features. This layer produces a BehavioralFeatures object that serves as input for all 26 models.
Phase 3: Model Scoring
The scoring phase takes the BehavioralFeatures object and the current SessionContext, runs all 26 models, and produces the final score set. Each model is a pure function: given the same features and context, it always returns the same scores.
The clickstream_scores Dataset Schema
All computed scores are written to the clickstream_scores Analytics Engine dataset -- a dedicated score store separate from the raw event stream -- which powers the dashboard, alerts, and exports.
How Scores Flow Through the System
The pipeline operates as an event-driven loop. Here is the full lifecycle of a single behavioral event:
- Browser SDK captures a user interaction (click, scroll, page transition, form field interaction).
- Edge Worker receives the event at the nearest Cloudflare PoP (200+ locations globally).
- Session State: Scoring runs synchronously at ingest -- a pure computation with no external I/O -- while live-session snapshots are maintained in Durable Objects.
- Feature Update: The new event updates the running feature counters and statistics.
- Model Scoring: All 26 models re-score using the updated features. Each model is a deterministic function with no external dependencies.
- Score Emission: Updated scores are written to the
clickstream_scoresdataset as events are processed. - Real-time Actions: Your page code can read the resulting scores through the Signals SDK (developer preview) and apply personalization rules with
applySignalseffects.
The scoring pass itself -- all 26 models on one event -- is held to a p95 under 3 ms by a CI-enforced benchmark that fails the build if it regresses. Scoring happens synchronously at ingest, so scores are updated as each event lands.
Design Principles
Several principles guide the design of every model in this series:
Incremental Computation
No model ever requires scanning the full event history. Every score can be updated from the previous score plus the new event. This keeps compute costs O(1) per event regardless of session length.
Weighted Signal Decay
Recent events carry more weight than older ones. Most models apply an exponential decay to older signals -- so a rage-click 30 seconds ago matters more than one from 5 minutes ago. (The decay constants shown in this series are illustrative; the exact values are tunable implementation details.)
Normalization to 0–100
Numeric scores are normalized to a common 0–100 scale (session momentum runs -100 to +100, and a few outputs are categorical labels). The normalization curves shown in this series are illustrative. This makes scores directly comparable across models and easy to reason about in business rules.
Cross-Model Interaction
As a design principle, models are meant to inform each other rather than operate in isolation. For example, a high frustration score should temper how you read a high intent score (frustrated users are less likely to convert), while strong engagement reinforces value estimates. Each part of the series discusses these interactions as design rationale for interpreting the scores together.
Configurable Thresholds
Every model exposes thresholds that can be tuned per-site. Defaults ship with the platform, and every weight and threshold can be overridden for your specific use case.
Series Index
This series is organized into ten parts, each covering two to four closely related models. The groupings reflect how models interact in practice -- understanding them together gives you a much richer picture than treating each in isolation.
| Part | Models Covered | Key Topics |
|---|---|---|
| Part 1 | Intent, Frustration, Engagement | Signal weights, intent categories, 9 frustration signals, tab visibility, 4 behavioral archetypes, e-commerce walkthrough |
| Part 2 | Value Estimator, Anomaly Detection | E-commerce vs SaaS scoring, dual baseline comparison, bot detection formula, false positive mitigation |
| Part 3 | Confusion, Emotional State | 7 confusion signals, 4 confusion types, 8 emotional states, mouse dynamics, typing cadence, ethical considerations |
| Part 4 | Decision Confidence, Regret Risk | 7 confidence signals, convergence analysis, speed-research mismatch, price sensitivity, intervention strategies |
| Part 5 | Churn Prediction, LTV Score | 9 churn signals, SaaS vs e-commerce, 4 LTV tiers, priority matrix with actions, real-time intervention timeline |
| Part 6 | Abandonment, Purchase Timing | 8 abandonment signals, exit intent formula, latency breakdown, proximity scoring, urgency classification |
| Part 7 | Content Affinity, Form Friction, Next Action | 6 affinity types, 5 friction types, sequential pattern mining, personalization trinity |
| Part 8 | Session Momentum, Navigation Entropy, Attention Score | Progression velocity, page-transition randomness, tab focus, reading patterns, bot vs human signals |
| Part 9 | Conversion Probability, Hover Intent, Scroll Depth Intelligence | Real-time conversion likelihood, CTA hover patterns, scroll regression, content zone analysis |
| Part 10 | Price Sensitivity, Loyalty, Micro-Conversion, Bot Detection | Comparison shopping, return propensity, funnel micro-steps, behavioral biometrics fraud detection |
Prerequisites
This series assumes familiarity with:
- Basic statistics (mean, standard deviation, sigmoid functions)
- Web analytics concepts (sessions, pageviews, events)
- TypeScript/JavaScript (code examples use TypeScript interfaces)
- SQL (for dataset schema and query examples)
No machine learning expertise is required. While the models are inspired by ML techniques, they are implemented as deterministic scoring functions that can be reasoned about directly.
Getting Started
Start with Part 1: Intent, Frustration & Engagement to understand the three foundational models, or jump directly to the model group most relevant to your use case using the series navigation above.
Each part is self-contained -- you can read them in any order -- but the cross-model interactions will make more sense if you have read the earlier parts first.