Behavioral Intelligence
Recapt's behavioral intelligence is the detection engine that powers self-healing UX. It's a custom ML model (not an LLM) that analyzes every session in real-time to find friction before you know it exists.
The Detection Engine
When a user rage-clicks a button, hesitates on a form, or gets lost in navigation, recapt's behavioral model detects these patterns and flags them as issues. This happens automatically on every session, no manual review required.
The model was trained specifically on user behavior patterns, which gives it capabilities that general-purpose AI lacks:
| Capability | What it means |
|---|---|
| Micro-signal detection | Catches subtle hesitations, tiny scroll adjustments, brief pauses that humans miss in replay |
| Real-time analysis | Every session is scored as it happens, not in batch |
| Deterministic output | Same input always produces the same score, no hallucination |
| Privacy-preserving | Analyzes interaction patterns, not content |
What It Detects
Frustration Signals
- Rage clicks: Rapid, repeated clicks on the same element
- Dead clicks: Clicks on non-interactive elements
- Scroll thrashing: Rapid scrolling up and down
- Backtracking: Repeatedly returning to previous pages
- Micro-frustrations: Small, rapid scroll reversals or brief repeated clicks that happen too fast to notice in playback
Confusion Signals
- Hesitation: Long pauses before taking action
- Erratic movement: Mouse movements that suggest uncertainty
- Path wandering: Taking indirect routes to complete tasks
- Form abandonment: Starting but not completing forms
- Micro-hesitations: Brief pauses and small cursor movements that indicate momentary uncertainty
Engagement Signals
- Confident navigation: Direct paths to goals
- Smooth interactions: Steady, purposeful movements
- Task completion: Successfully finishing intended actions
Behavioral Scores
Every session receives three scores on a 0-1 scale:
| Score | What it measures | High score means |
|---|---|---|
| Frustration | Rage clicks, dead clicks, thrashing, backtracking | User is struggling with the UI |
| Confusion | Hesitation, erratic movement, path wandering | User doesn't know what to do |
| Confidence | Direct navigation, purposeful movement | User is completing tasks smoothly |
These scores feed directly into the self-improvement workflow. When frustration spikes on a page, recapt flags it as an issue and the AI investigates.
From Detection to Action
The behavioral model doesn't just report symptoms. It connects signals to causes:
Symptom: "Users are rage-clicking the checkout button"
Cause: "The button appears clickable but doesn't respond until form validation completes"
Fix: "Add a loading state and disable the button until ready"
This causal understanding is what enables autonomous fixing. The AI doesn't just know something is wrong. It knows why, and can propose a specific fix.
Session Outcomes
Beyond scores, recapt automatically classifies each session's outcome:
| Outcome | What it means |
|---|---|
| Completed | User achieved their goal with minimal friction |
| Struggled | User achieved their goal but with friction |
| Blocked | User couldn't complete their goal |
| Exploring | User was browsing without clear intent |
| Disengaged | User left without meaningful interaction |
| Error | User encountered a technical failure |
Learn more: Session Outcomes
Why Not an LLM?
General-purpose LLMs are powerful, but they're not built for this job:
| Aspect | Recapt's Behavioral Model | General-Purpose LLMs |
|---|---|---|
| Training | Trained on real user behavior patterns | Trained on text data |
| Output | Consistent numerical scores (0-1) | Generated text that can vary |
| Speed | Real-time analysis on every session | Slower, resource-intensive |
| Reliability | Deterministic: same input, same output | Can hallucinate or be inconsistent |
| Focus | Detects subtle behavioral signals | Optimized for language understanding |
The behavioral model handles detection. LLMs (via the MCP server) handle diagnosis and fixing. Each does what it's best at.
Next Steps
- Behavioral Scores: Detailed breakdown of frustration, confusion, and confidence
- Session Outcomes: How sessions are automatically classified
- Self-Improvement Workflow: How detection feeds into autonomous fixing