Scribe Detection Reference

This document provides authoritative reference information on how Scribe detects, scores, and weights visitor preferences.

Activation Thresholds

Scribe becomes “ready” when all three conditions are met:

Threshold	Value	Purpose
Session age	≥10 minutes	Ensures visitor is genuinely browsing, not just passing through
Properties viewed	≥3 properties	Provides sufficient data to detect meaningful patterns
Average confidence	≥50%	Ensures preferences are reliable enough to show visitor

Until these thresholds are met, the floating action button displays a “Start exploring” message.

Signal Sources & Confidence

Each behavioral signal is assigned a confidence range based on its reliability:

Signal	What It Reveals	Confidence Range	Weight
Explicit filters	Direct user intent (price, beds, location filters)	95%	Highest
Return visits	Strong interest in specific properties	90%	Very High
Favorites saved	Strong positive preferences	90%	Very High
Properties viewed	Price range, size, property type from listings	30-85%	Variable
Photo engagement	Specific amenities of interest	30-80%	Variable
Search queries	Location and feature priorities from natural language	50-85%	High
Quick bounces	Deal-breakers and dislikes	30-70%	Variable

Note

Preferences below 30% confidence are not displayed to visitors.

Engagement Scoring

Every property interaction is scored on a 100-point scale to weight how much it influences preferences:

Interaction	Max Points	Full Credit Threshold
Time on page	30	5+ minutes
Scroll depth	20	100% of page
Photos viewed	20	All photos in carousel
Mortgage calculator	10	Used calculator
PDF download	10	Downloaded flysheet
Tour request	10	Submitted request
Share action	5	Shared property
Favorite action	5	Saved to favorites

Total possible: 100 points per property view

Scoring Formula

engagement_score = (time_points + scroll_points + photo_points +
                   action_points) / 100

Properties with higher engagement scores have proportionally more influence on preference detection.

Partial Credit

Points are awarded proportionally:

• Time on page: Linear scale from 0-5 minutes (6 points per minute)
• Scroll depth: Linear scale from 0-100% (0.2 points per percent)
• Photos viewed: Points per photo viewed (5 points each for 4-photo carousel)
• Actions: Full points awarded for any completion

Return Visit Detection

When a visitor views the same property multiple times:

Visit Count	Confidence Boost	Weight Multiplier
First visit	Normal (30-85%)	1.0x
Return visit	Boosted to 90%	2.0x
Third visit	Remains 90%	3.0x
Fourth+ visit	Remains 90%	Visit count × 1.0

Tracking Mechanism

Return visits are tracked in browser localStorage:

{
  "idx_property_views": {
    "listing_12345": [
      "2024-01-15T10:30:00Z",
      "2024-01-16T14:20:00Z",
      "2024-01-17T09:15:00Z"
    ]
  }
}

Timestamps persist across sessions within the same browser.

Negative Preferences

Scribe tracks what visitors don’t want by detecting quick bounces.

Bounce Detection Criteria

Metric	Threshold	Meaning
Time on page	<5 seconds	Too quick to be interested
Scroll depth	<20%	Didn’t engage with content
Bounce count	3+ properties	Consistent pattern detected

Tracked Attributes

Negative preferences are inferred for:

• Property type (single_family, condo, townhome, etc.)
• Price bucket (under $300k, $300k-$500k, $500k-$750k, etc.)
• Specific features (HOA, no garage, no yard, etc.)

Display

Negative preferences appear with a “Not interested” label and are excluded from:

• Agent matching algorithms
• Search suggestions
• Automated property alerts

Photo Feature Detection

When visitors browse property photos, Scribe tracks engagement at the photo level.

Dwell Time Tracking

Dwell Duration	Weight	Notes
<500ms	0 points	Considered a quick scroll-past
500ms-2s	1 point	Casual viewing
2s-5s	2 points	Moderate interest
5s+	3 points	Strong interest

Zoom Interactions

Zooming into a photo adds:

• +5 points to that photo’s engagement score
• +2% confidence to any features detected in that photo

Feature Recognition

Features become preferences when:

• Viewed for 5+ seconds total across all photos
• Detected in 2+ properties
• Average confidence ≥40%

Common detected features:

• Kitchen: granite countertops, island, stainless appliances, pantry
• Bathroom: soaking tub, walk-in shower, double vanity
• Exterior: pool, covered patio, mountain views, 3-car garage
• Interior: fireplace, vaulted ceilings, hardwood floors, built-ins

Search Query Analysis

Natural language searches are analyzed for feature keywords.

Keyword Extraction

"modern home with pool near downtown"
       ↓           ↓        ↓
   style pref  feature  location

Confidence Assignment

Mention Count	Confidence	Example
1 mention	40%	“pool” mentioned once
2 mentions	65%	“pool” mentioned in 2 searches
3+ mentions	85%	“pool” mentioned in 3+ searches

Keywords must appear in 2+ searches to become inferred preferences.

Feature Synonym Mapping

The system recognizes synonyms:

• “pool” = “swimming pool” = “outdoor pool”
• “granite” = “granite counters” = “granite countertops”
• “garage” = “car garage” = “parking”

Confidence Score Ranges

Each preference includes a confidence score (0-100%):

Confidence	Source	Display Behavior
100%	User confirmed or manually added	Always shown, cannot be overridden
85-95%	Explicit filter applied by user	Always shown, high priority
60-85%	Return visit or strong signal	Shown with “detected” label
30-70%	Inferred from passive browsing	Shown with confidence %
Under 30%	Insufficient evidence	Not shown to visitor

Confidence Decay

Preferences decay over time if contradicted by new behavior:

• Per-session decay: 2% per contradicting property view
• Minimum floor: Inferred preferences cannot decay below 20%
• User-confirmed: No decay (remains at 100%)

Example: Visitor initially viewed $400k-$500k properties (65% confidence), then viewed 5 properties at $600k-$700k. Original preference decays to 55%, new range emerges at 70%.

Match Score Calculation

Agent matching uses cosine similarity between embeddings:

Formula

similarity_score = 1 - (criteria_embedding <=> preference_embedding) / 2
match_percentage = similarity_score * 100

Match Labels

Score	Label	Threshold
≥80%	Excellent Match	Strong semantic alignment
≥60%	Great Match	Good fit for preferences
≥40%	Good Match	Relevant experience
<40%	Available	General availability

Fallback Behavior

If fewer than 3 agents score ≥40%:

1. Add brokerage’s primary agent with 50% score
2. Fall back to first eligible agents by join date
3. Return empty list only if brokerage has zero agents

Browser Storage Schema

Scribe uses localStorage for client-side state:

Key	Type	Purpose	Expiry
scribe_disabled	boolean	Permanent opt-out flag	Never
scribe_dismissed	timestamp	Temporary snooze until time	24 hours
idx_property_views	object	Return visit tracking	Session

Data Format

{
  "scribe_disabled": false,
  "scribe_dismissed": "2024-01-15T14:30:00Z",
  "idx_property_views": {
    "listing_12345": ["2024-01-15T10:00:00Z"]
  }
}

Session-to-Client Sync

When a visitor converts, session data syncs to CRM if confidence ≥70% or confirmed:

Session Data	Client Field	Sync Condition
Price preferences	budget_min, budget_max	Confidence ≥70% or confirmed
Location preferences	preferred_areas	Confidence ≥70% or confirmed
Property types	property_types	Confidence ≥70% or confirmed
Bedroom preferences	beds_min	Confidence ≥70% or confirmed
Bathroom preferences	baths_min	Confidence ≥70% or confirmed
Freeform notes	notes (appended)	Always synced
Assigned agent	assigned_agent_id	Always synced

Note

Only high-confidence preferences transfer to avoid polluting CRM with weak signals.

Related References

• Understanding Scribe — Conceptual overview
• Using Scribe — Practical usage guide
• Scribe API — API endpoints and integration