Skip to content
TruthVouch Docs

Narrative Contamination & Risk Scoring

Narrative contamination occurs when false or outdated information about your brand spreads from one AI system to others. A single hallucination in ChatGPT can contaminate Claude, Gemini, and other models if that false information enters their training data.

How Contamination Happens

The Contamination Cycle

Week 1: One engine mentions false claim
  ChatGPT: "TruthVouch was founded in 2018"
  (Source: old blog post in training data)


Week 2-3: Other engines pick it up
  Claude crawls a webpage that quoted ChatGPT's response
  Now Claude: "Founded in 2018"


Week 4-5: Spreads further
  Gemini's training data includes a forum post citing ChatGPT
  Now Gemini: "Founded in 2018"


Week 6+: Entrenched
  5 engines now repeat the false claim
  It's harder to correct because multiple "sources" now exist

Why It Happens

  1. Training data includes AI responses

    • LLM training data includes web content
    • Web content increasingly includes AI outputs
    • AI systems cite each other
    • False information spreads through this citation cycle

  2. Web sources are authoritative

    • If false info appears on 3+ websites, AI systems treat it as fact
    • A single hallucination can create “evidence” by being quoted

  3. No self-correction

    • AI systems don’t fact-check each other
    • No mechanism to say “that other AI was wrong”
    • Each system independently learns from training data

Contamination Risk Scoring

Brand Intelligence scores the contamination risk for each inaccurate claim:

Contamination Risk = (Prevalence × 0.40) + (Growth Rate × 0.35) + (Source Authority × 0.25)


Prevalence:
- 1 engine: 20% risk
- 2 engines: 40% risk
- 3 engines: 60% risk
- 4+ engines: 80%+ risk


Growth Rate:
- Stable (not spreading): Low risk
- Growing 1 engine/week: Medium risk
- Growing 2+ engines/week: High risk


Source Authority:
- Obscure web source: Low risk
- Major publication: High risk
- Your website: Very high risk

Risk Levels

LOW RISK (0-30%)

Characteristics:

  • Inaccuracy mentioned by only 1 engine
  • Not growing (no other engines picking it up)
  • Source is obscure

Example:

  • ChatGPT says “Founded in 2019” (but you founded in 2020)
  • No other engine mentions this
  • Source is a single old blog post

Action: Monitor for 2 weeks. If it doesn’t spread, low priority.

MEDIUM RISK (30-60%)

Characteristics:

  • Inaccuracy mentioned by 2-3 engines
  • Starting to grow
  • Or source is semi-authoritative

Example:

  • ChatGPT says “500 customers”
  • Claude picked it up (now 2/5 engines)
  • Growing at 1 engine per week

Action: Prioritize fixing. Counter-messaging needed within 2 weeks.

HIGH RISK (60-80%)

Characteristics:

  • Inaccuracy mentioned by 3+ engines
  • Growing rapidly (2+ engines per week)
  • Multiple sources

Example:

  • “Founded in 2018” mentioned by 4 engines
  • Gained 2 engines this week
  • Now appearing on 5+ websites quoting each other

Action: Urgent. Major counter-messaging and website updates required.

CRITICAL RISK (80%+)

Characteristics:

  • Mentioned by 4+ engines
  • Rapidly spreading
  • High-authority sources (news, Wikipedia, major websites)

Example:

  • False claim about your product capabilities
  • Mentioned by 5/5 monitored engines
  • In major publication or Wikipedia
  • Growing daily

Action: Crisis mode. Immediate intervention needed.

Detecting Early Contamination

Brand Intelligence alerts you when contamination risk increases:

  • Alert Type: “Contamination Risk Rising”
  • Trigger: Same inaccuracy spreads to a 2nd engine (30% risk threshold)
  • Action: You’re notified immediately

Example alert:

CONTAMINATION ALERT
Claim: "TruthVouch was founded in 2018"


Week 1: ChatGPT (1 engine, 20% risk) → No alert
Week 2: Claude picks it up (2 engines, 40% risk) → ALERT SENT


Risk Level: MEDIUM
Current Spread: 2 engines
Growth Rate: +1 engine per week at current rate


Recommended Action:
1. Create content clarifying founding year
2. Update relevant website pages
3. Monitor weekly for further spread

Contamination vs. Accuracy

Different but related concepts:

Accuracy:

  • Does AI say something true about you?
  • Measured per-claim
  • Updated weekly as AI systems change

Contamination:

  • Is a false claim spreading?
  • Measured by count of engines affected
  • Growing threat over time

Example:

Claim: "Founded in 2019"
Actual: "Founded in 2020"


Week 1 Accuracy: 60/100 (1 engine wrong, 4 right)
Week 1 Contamination: LOW (only 1 engine affected)


Week 3 Accuracy: 40/100 (3 engines wrong, 2 right)
Week 3 Contamination: MEDIUM (spreading to 3 engines)


Week 6 Accuracy: 0/100 (5 engines wrong)
Week 6 Contamination: CRITICAL (all engines wrong)

Preventing Contamination

Immediate (Stop Spread)

  1. Publish corrected information on your website

    • Clear, prominent statement of correct fact
    • Include date when you made the correction
    • Example: “TruthVouch was founded in April 2020. [Some sources incorrectly state 2018; this was corrected January 2024.]”

  2. Create news/press release

    • Official source documenting correction
    • Gives AI systems newer, authoritative source

  3. Update public profiles

    • LinkedIn company page
    • Wikipedia (if listed)
    • Crunchbase
    • Any directory listing you control

Short-term (Reduce Impact)

  1. Monitor spread weekly

    • Is the false claim still growing?
    • Check back every Monday
    • Alert if growth continues

  2. Create counter-content

    • Blog post: “Clarifying Our Founding Story”
    • FAQ: “When was TruthVouch founded?”
    • Makes your correct version more visible in search/AI

  3. Reach out to sources

    • If false info is on a major website you know about, contact them
    • Request correction
    • Most will update if you provide evidence

Long-term (Replace False Narrative)

  1. Wait for model updates

    • AI models retrain periodically
    • ChatGPT: Every 3-4 months
    • Claude: Every 3-6 months
    • Gemini: Every 2-3 months
    • When they retrain with new training data (including your corrected website), the false claim should disappear

  2. Publish more content

    • The more you write about your founding, the better
    • AI systems weight more recent, authoritative sources heavier
    • Your website should dominate results

  3. Build backlinks

    • Quality external links to your website increase its authority
    • AI training data gives more weight to linked content

Tracking Contamination Over Time

Navigate to Brand Intelligence → Contamination Risk for:

  • Contamination Timeline: Tracks each false claim’s spread
  • Trend Chart: Shows prevalence over time (growing/stable/declining)
  • Per-Engine Breakdown: Which engines have the false claim
  • Source Analysis: Where the false claim originated

Example Timeline:

Claim: "AI Visibility Score affects search rankings"
[This is FALSE; GEO measures AI discovery, not SEO]


Timeline:
Jan 15: ChatGPT mentions this (1 engine, 20% risk)
Jan 20: Perplexity mentions it (2 engines, 40% risk)
Jan 25: Claude mentions it (3 engines, 60% risk) ← HIGH RISK ALERT
Jan 30: Gemini mentions it (4 engines, 80% risk) ← CRITICAL ALERT
Feb 5:  Still 4 engines (stable, 80% risk) → Plateau


Action Timeline:
Jan 20: You published blog clarifying the difference
Jan 25: Contacted sources that quote the false claim
Feb 1:  Created FAQ "GEO vs SEO: What's the Difference?"
Feb 5:  Noticed growth has stopped
Feb 15: Gemini updated and dropped the false claim (back to 3 engines)
Mar 1:  All engines corrected (back to 20% or lower)

Managing Critical Contamination

If a claim reaches CRITICAL risk (80%+):

Immediate Actions (Today)

  1. Create official correction on your homepage
  2. Send media inquiry to major sources citing the false claim
  3. Post on social media with correction
  4. Notify your team to mention correction in customer conversations

Short-term (This Week)

  1. Publish detailed blog post explaining correct info
  2. Update all website pages that might perpetuate the false claim
  3. Create FAQ addressing the false claim directly
  4. Monitor daily for further spread

Medium-term (This Month)

  1. Reach out to Wikipedia (if listed) to correct entry
  2. Contact major publications that cited false claim
  3. Build PR campaign around correct narrative
  4. Implement schema markup on your website to help AI systems understand your facts

Case Studies

Case 1: Prevented Contamination (Quick Action)

False Claim: "TruthVouch charges per API call"
Actual: "TruthVouch charges per month, flat rate"


Week 1: ChatGPT mentions pay-per-call pricing
Action: You immediately publish "Our Pricing Model" blog post
Week 2: Claude picks it up, but your blog post appears at top of search
Result: Other engines cite your blog instead of ChatGPT
Outcome: Contamination stopped at 2 engines before spreading further

Case 2: Late Intervention (Harder Fix)

False Claim: "TruthVouch is expensive"
Actual: Competitive pricing vs. alternatives


Week 1: ChatGPT mentions pricing as high
Week 2: Claude and Gemini pick it up
Week 3: Perplexity mentions it (now 4/5 engines)
Week 4: You create "ROI Calculator" tool


Result: Your ROI content helps, but claims already entrenched
Timeline to full recovery: 3-4 months (waiting for model updates)
Lesson: Act at week 1, not week 4

Tools for Contamination Management

In TruthVouch dashboard:

Contamination Monitoring

  • Automated tracking of claim spread
  • Alerts when claims reach medium/high risk
  • Per-engine tracking
  • Source attribution

Recommended Actions

  • Auto-generated content suggestions
  • FAQ templates addressing false claims
  • Blog post outlines
  • Counter-messaging frameworks

Risk Forecasting

  • Trend projection: “At current rate, will reach 5 engines in X weeks”
  • Growth rate monitoring
  • Early warnings

Next Steps