UGC Moderation Assistant

Overview

This skill provides a structured moderation framework for user-generated content (UGC) across e-commerce platforms — product reviews, ratings, Q&A, community forums, and user-submitted photos/videos. It balances brand protection with authentic consumer voice, applying regulatory, legal, and brand-safety filters while preserving genuine feedback that drives conversion.

Authentic reviews increase conversion rates by 15-30%. The goal is not to suppress negative feedback but to ensure all published content is genuine, compliant, and safe.

When to Use

Processing incoming product reviews before publication on DTC or marketplace storefronts.
Moderating Q&A submissions on product detail pages.
Screening user-submitted photos and videos for brand safety.
Auditing existing review corpuses for policy violations or fake review patterns.
Generating moderation reports for brand and legal teams.
Responding to or flagging reviews that require brand intervention.

Required Inputs

| Input | Description | Example | |---|---|---| | ugc_content | The user-generated text, rating, and metadata | { "text": "...", "rating": 4, "author": "...", "date": "..." } | | product_info | Associated product name, category, and claims | Product record | | moderation_policy | Brand-specific moderation rules and thresholds | Policy document | | channel | Platform where UGC will appear | "DTC Shopify", "Amazon Vine", "Bazaarvoice" | | media_attachments | URLs to any user-uploaded images or videos | List of URLs | | escalation_contacts | Contacts for legal, PR, and customer service escalation | Email/Slack channels | | historical_patterns | Known fake review patterns or repeat offenders | Pattern library |

Methodology

Step 1 — Content Classification

Classify each piece of UGC into a moderation category:

| Category | Definition | Default Action | |---|---|---| | Approved | Genuine, policy-compliant, and relevant | Publish | | Approved with Edit | Minor issues (PII, formatting) that can be auto-corrected | Publish after edit | | Flagged for Review | Ambiguous content requiring human judgment | Queue for manual review | | Rejected — Policy | Violates moderation policy (profanity, hate speech, threats) | Block with reason code | | Rejected — Compliance | Contains regulated claims or legal risk | Block and escalate | | Rejected — Authenticity | Suspected fake, incentivized, or competitor-planted review | Block and log |

Step 2 — Authenticity Assessment

Apply the Review Authenticity Scorecard to detect inauthentic content:

| Signal | Weight | Indicators | |---|---|---| | Linguistic Analysis | 25% | Generic language, excessive superlatives, template patterns, lack of product-specific detail | | Behavioral Patterns | 25% | Review velocity (multiple reviews in minutes), reviewer history, geographic anomalies | | Purchase Verification | 20% | Verified purchase flag, order-to-review timing (< 24hrs suspicious, > 90 days stale) | | Sentiment-Rating Alignment | 15% | 5-star rating with negative text, or vice versa — indicates manipulation | | Duplicate Detection | 15% | Near-duplicate text across products, accounts, or time periods |

Authenticity Score: 0-100. Content below 40 is auto-rejected. 40-65 is flagged for human review. Above 65 passes authenticity gate.

Step 3 — Safety & Compliance Screening

Screen content against multiple safety layers:

Legal Safety:

PII detection: email addresses, phone numbers, full names of non-public individuals, addresses.
Defamation risk: false statements of fact about competitors or individuals.
Intellectual property: copyrighted text, trademarked terms used inappropriately.

Regulatory Compliance:

Health/medical claims in reviews: "This cured my diabetes" — flag for disclaimer or removal.
Off-label use descriptions for regulated products.
Adverse event reports (FDA-reportable for supplements, OTC drugs, cosmetics).

Brand Safety:

Profanity and hate speech (zero tolerance).
Violent or sexually explicit content.
Competitor promotion or spam links.
Politically divisive or discriminatory language.

Step 4 — Adverse Event Detection (CPG-Specific)

For FDA-regulated categories (supplements, OTC, cosmetics, food), identify potential adverse event reports:

Scan for medical terminology: "allergic reaction," "rash," "hospitalized," "side effect."
Classify severity: mild (discomfort), moderate (required medical attention), serious (hospitalization, life-threatening).
Serious adverse events trigger mandatory escalation to regulatory affairs within 24 hours.
Log all adverse event mentions in the pharmacovigilance tracking system.

Step 5 — Sentiment Analysis & Brand Intelligence

Extract actionable intelligence from approved UGC:

Sentiment Distribution: Positive / neutral / negative ratio by product and time period.
Topic Clustering: Group reviews by theme (packaging, taste, efficacy, value, shipping).
Emerging Issues: Detect sudden spikes in negative sentiment on specific topics.
Competitive Mentions: Track competitor name mentions and comparative sentiment.
Feature Requests: Identify recurring requests for product improvements.

Step 6 — Response Prioritization

Rank reviews requiring brand response by urgency:

| Priority | Criteria | Response SLA | |---|---|---| | P0 — Crisis | Safety issue, viral potential, adverse event | 2 hours | | P1 — Urgent | 1-star verified purchase with specific product defect | 24 hours | | P2 — Important | Detailed negative review with actionable feedback | 48 hours | | P3 — Standard | Positive review warranting thank-you response | 72 hours | | P4 — Monitor | Neutral review, no action needed | No response required |

Output Specification

output:
  moderation_decision: string        # "approved" | "approved_with_edit" | "flagged" | "rejected"
  rejection_reason: string           # Policy code if rejected
  authenticity_score: float          # 0-100
  safety_flags: list[string]        # Specific safety issues found
  adverse_event_detected: boolean
  adverse_event_severity: string     # "mild" | "moderate" | "serious" | null
  sentiment: string                  # "positive" | "neutral" | "negative"
  topics: list[string]              # Extracted themes
  response_priority: string          # P0-P4
  suggested_response: string         # Draft brand response if P0-P3
  edits_applied: list[string]       # Auto-corrections made (PII redaction, etc.)
  pii_redacted: boolean

Analysis Framework

UGC Health Dashboard Metrics (aggregate across all moderated content):

| Metric | Healthy Range | Alert Threshold | |---|---|---| | Approval Rate | 85-95% | < 75% (overly restrictive) or > 98% (under-moderated) | | Fake Review Rate | < 5% | > 10% | | Avg Authenticity Score | > 75 | < 60 | | Adverse Event Rate | < 0.1% | > 0.5% (potential product issue) | | Avg Response Time (P0/P1) | < 12 hrs | > 24 hrs | | Sentiment Trend | Stable or improving | 3+ consecutive weeks declining |

Examples

Input Review: "TERRIBLE product!! Gave me a horrible rash all over my arms. Called my doctor and he said it was a chemical burn. DO NOT BUY. Contact me at jane.smith@email.com to join the lawsuit."

Analysis:

Moderation Decision: Rejected — Compliance
Authenticity Score: 78 (genuine language, specific details)
Safety Flags: PII detected (email), potential legal threat
Adverse Event: Detected — Severity: Moderate (required medical attention)
Response Priority: P0 — Crisis
Actions: (1) Redact email address. (2) Escalate adverse event to regulatory affairs immediately. (3) Escalate legal threat to legal team. (4) Draft empathetic response with customer service contact.

Guidelines

Never suppress genuine negative reviews — they build trust and provide product intelligence.
Adverse event detection is a legal obligation for FDA-regulated products. Err on the side of over-reporting.
PII must be redacted before publication, never after.
Incentivized reviews (samples, discounts) must be disclosed per FTC guidelines.
Maintain consistent moderation standards regardless of rating — do not apply stricter standards to negative reviews.
All moderation decisions must be logged with timestamps and rationale for audit trails.

Validation Checklist

[ ] Content is classified into the correct moderation category.
[ ] Authenticity score is calculated using all five signal dimensions.
[ ] PII is detected and redacted before any publication decision.
[ ] Adverse events are identified, classified by severity, and escalated per protocol.
[ ] Safety screening covers legal, regulatory, and brand-safety layers.
[ ] Response priority is assigned based on urgency criteria.
[ ] Suggested brand response is drafted for P0-P3 reviews.
[ ] Moderation decision is logged with full audit trail.
[ ] Fake review patterns are cross-referenced against historical data.
[ ] Aggregate moderation metrics fall within healthy ranges.