Troubleshooting Guide

Fresh Updated January 2026

Common issues and solutions when dealing with LLM technical phenomena in GEO/AEO strategy.

Quick Diagnostic

Variance Issues

Problem: Different outputs for same prompt

Symptoms:

• Same prompt produces different results
• Inconsistent citations across runs
• Variable content recommendations

Causes:

• KV cache non-determinism (11.67 unique outputs at temp=0)
• MoE routing variance (30 unique outputs in 30 runs)
• Batch composition changes

Solutions:

Solution	Implementation	Effectiveness
Multiple runs	Run 3-5 times, use consensus	High
Lower temperature	Set temp=0 (still varies)	Partial
Consistent timing	Same time of day	Low
Enterprise tier	Dedicated capacity	Medium

Problem: MoE model high variance

Symptoms:

• Extremely different outputs each run
• No consistency even with identical prompts

Solutions:

• Accept variance as inherent to MoE architecture
• Use consensus-based approach (3+ runs)
• Consider dense model alternatives for consistency-critical tasks

---

Citation Issues

Problem: Not being cited on a platform

Diagnostic questions:

Platform-specific fixes:

Platform	If Not Cited	Action
ChatGPT	Missing from training	Build Wikipedia presence, comprehensive docs
Perplexity	Not discoverable	Update content, add schema markup
Claude	Not validated	Get press coverage, academic citations
Gemini	Poor SEO	Traditional SEO, Google News presence

Problem: Cited but poor positioning

Solutions:

1. Check content position (key claims in first 20%)
2. Reduce distractors (each one costs ~15% citation chance)
3. Add unique data/research
4. Strengthen authority signals

Problem: Only 42% citation overlap

This is expected behavior. Different platforms use different source priorities.

Strategy:

• Don't rely on single-platform optimization
• Create platform-specific content layers
• Test on each target platform individually

---

Context Issues

Problem: Quality degradation with long content

Symptoms:

• Answers become less accurate
• Hallucinations increase
• Key information ignored

Diagnostic:

Solutions by severity:

Severity	Solution
Mild (50-100k)	Move key info to first/last 20%
Moderate	Split into multiple documents
Severe (>100k)	Aggressive summarization, chunking

Problem: "Lost in the middle" effect

Symptoms:

• Information in middle 40-60% is missed
• Model focuses on start and end only

Solutions:

1. Restructure content (key points at start/end)
2. Repeat critical information in summary
3. Use clear heading hierarchy
4. Reduce middle section length

---

Alignment Issues

Problem: Content being refused or filtered

Symptoms:

• Model declines to answer
• Overly cautious responses
• Information suppression

Causes:

• RLHF alignment tax (5-15% performance drop)
• Safety training triggers
• Sensitive topic detection

Solutions:

Approach	Implementation
Reframe neutrally	Remove assertive/controversial framing
Add context	Explain legitimate use case
Use third-party	Cite authoritative sources
Split requests	Break into smaller, neutral queries

Problem: Conservative recommendations

This is expected. Claude and other aligned models prioritize safety.

Strategy:

• Accept conservative bias in outputs
• Frame content for conservative models
• Use third-party validation to overcome skepticism

---

Rate Limiting Issues

Problem: Inconsistent results during high volume

Symptoms:

• Variance increases during bulk operations
• Queue delays affect output

Solutions:

1. Smooth request patterns (avoid bursts)
2. Space requests evenly over time
3. Monitor TPM usage
4. Consider enterprise tier for consistency

Problem: Rate limit errors

Solutions:

// Implement exponential backoff
async function requestWithBackoff(prompt, maxRetries = 5) {
  for (let i = 0; i < maxRetries; i++) {
    try {
      return await makeRequest(prompt);
    } catch (e) {
      if (e.status === 429) {
        await sleep(Math.pow(2, i) * 1000);
      } else throw e;
    }
  }
}

---

Monitoring Checklist

Regular checks to prevent issues:

• [ ] Weekly: Quick citation check on top 3 queries per platform
• [ ] Monthly: Full matrix test (all queries, all platforms)
• [ ] Quarterly: Strategy review and platform priority adjustment
• [ ] Ongoing: Monitor for behavioral drift

When to Escalate

Consider professional help when:

• Citation drops >50% without content changes
• Multiple platforms simultaneously affected
• New competitor dominates citations
• Unexplained systematic bias against content

See Also

• Data Points Reference
• SOP-001: KV Cache Non-Determinism
• SOP-002: Context Rot
• SOP-003: Alignment Tax
• Workflow: Multi-Platform Strategy