Not sure where to start? Take a guided tour.
AI Site Grader
AI Query Scanner
llms.txt GeneratorAgreement by query class
Comparison prompts are the most stable query class, while broader best-of and general prompts remain less portable across models.
Cross-model agreement benchmark across major prompt families.
The key numbers.
- Comparison-query agreement
- Comparison prompts produce the highest average agreement.
- 50.4%
- General-query agreement
- General prompts are less stable across models.
- 42.2%
- Best-of high divergence
- Best-of prompts frequently split models.
- 14.8%
The ranked view.
- 150.4%
Comparison queries
The highest average agreement rate in the study.
- 245.3%
How-to queries
More constrained than general prompts, but still not highly converged.
- 344.1%
Alternative queries
Moderate agreement with a smaller sample.
- 443.4%
Best-of queries
Broad buyer-intent prompts still split models materially.
- 542.2%
General queries
The least stable mainstream prompt family in the benchmark.
What changed
Tighter prompts converge more
Constrained comparisons pull models closer together than broad recommendation prompts.
Best-of is still volatile enough to matter
Even widely searched recommendation prompts often split the model set.
| Measure | Result |
|---|---|
| Comparison-query agreement Comparison prompts produce the highest average agreement. | 50.4% |
| General-query agreement General prompts are less stable across models. | 42.2% |
| Best-of high divergence Best-of prompts frequently split models. | 14.8% |
Population and method
Built from 797,644 valid comparisons across 44,088 reports and 8 models, covering 6,439,133 model responses in the observed window.
Where the evidence stops
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”.
Cite this tracker.
Source study: Model Divergence“Comparison prompts are the most stable query class, while broader best-of and general prompts remain less portable across models.”
Mack Grenfell · Agreement by query class
Same Question, Different AI, Different Answers · Trakkr Research · Updated 11 March 2026