{
  "kind": "answer",
  "studySlug": "model-divergence",
  "slug": "are-general-and-best-of-prompts-more-volatile-than-comparisons",
  "title": "Are general and best-of prompts more volatile than comparisons?",
  "description": "Yes. Comparison prompts average 50.4% agreement, while general prompts average 42.2% and best-of prompts carry a 14.8% high-divergence rate.",
  "lastUpdated": "2026-03-11",
  "lastTested": "2026-03-11",
  "sourceStudyUrl": "/trakkr-research/model-divergence",
  "sourceStudyTitle": "Same Question, Different AI, Different Answers",
  "claimIds": [
    "model-divergence:comparison-agreement",
    "model-divergence:general-agreement",
    "model-divergence:bestof-divergence"
  ],
  "relatedSlugs": [
    "answer:what-does-an-average-top-three-overlap-of-two-point-eight-mean",
    "answer:should-you-use-one-model-as-a-proxy-for-all-ai-visibility",
    "fact:comparison-prompts-are-the-most-stable-query-class",
    "tracker:query-class-agreement-tracker"
  ],
  "methodologySummary": "Built from 797,644 valid comparisons across 44,088 reports and 8 models, covering 6,439,133 model responses in the observed window.",
  "limitations": [
    "Agreement is measured across recommendation outputs, not across hidden reasoning or retrieval context.",
    "Average agreement can hide large differences between query classes and model pairs.",
    "The study measures overlap, not which answer was objectively “right”."
  ],
  "keywords": [
    "model divergence",
    "AI agreement",
    "ChatGPT vs Claude",
    "Gemini vs Perplexity",
    "best of prompt volatility",
    "general query divergence"
  ],
  "schemaHints": {
    "pageType": "Article",
    "includeDataset": true
  },
  "question": "Are general and best-of prompts more volatile than comparisons?",
  "directAnswer": "Yes. Comparison prompts average 50.4% agreement, while general prompts average 42.2% and best-of prompts carry a 14.8% high-divergence rate.",
  "answerSummary": "Open-ended market questions leave more room for models to choose different evidence and ranking logic, requiring broader monitoring strategies.",
  "keyFacts": [
    {
      "label": "Comparison-query agreement",
      "value": "50.4%",
      "detail": "Comparison prompts produce the highest average agreement.",
      "claimId": "model-divergence:comparison-agreement"
    },
    {
      "label": "General-query agreement",
      "value": "42.2%",
      "detail": "General prompts are less stable across models.",
      "claimId": "model-divergence:general-agreement"
    },
    {
      "label": "Best-of high divergence",
      "value": "14.8%",
      "detail": "Best-of prompts frequently split models.",
      "claimId": "model-divergence:bestof-divergence"
    }
  ],
  "evidenceTable": [
    {
      "label": "Comparison-query agreement",
      "value": "50.4%",
      "note": "Comparison prompts produce the highest average agreement."
    },
    {
      "label": "General-query agreement",
      "value": "42.2%",
      "note": "General prompts are less stable across models."
    },
    {
      "label": "Best-of high divergence",
      "value": "14.8%",
      "note": "Best-of prompts frequently split models."
    }
  ],
  "whyItMatters": "Operators must allocate resources differently based on query type, as high-divergence categories require multi-model optimization rather than single-platform focus.",
  "whatToDo": [
    "Track visibility across multiple models instead of using one platform as a proxy for the whole market.",
    "Prioritize query classes where disagreement is highest because that is where share can move fastest.",
    "Treat consensus as a benchmark, but treat divergence as the operating reality."
  ],
  "faqs": [
    {
      "question": "Which prompt type produces the highest agreement across models?",
      "answer": "Comparison prompts produce the highest average agreement at 50.4%."
    },
    {
      "question": "How often do best-of prompts cause models to split?",
      "answer": "Best-of prompts carry a 14.8% high-divergence rate, frequently splitting models."
    }
  ]
}