# Mercer vs Radford: which do AI models recommend for compensation benchmarkin, September 2026

HR AI Recommendation Index, September 2026 Edition, Compensation benchmarking data. Zero of twelve models named Mercer first on the direct prompt; zero named Radford. Page: https://hr-ai-index.com/rewards/compensation-benchmarking-data/mercer-vs-radford/

| | First-choice share | Rank | Negative rate | Labels | Models naming it |
|---|---|---|---|---|---|
| Mercer | 6% | #4 of 8 | 43% | 28 | 10 of 12 |
| Radford | 4% | #5 of 8 | 37% | 27 | 11 of 12 |

## The direct prompt, model by model

- GPT-5.4 mini: neither first, one named (first choices: Payscale) (alternatives: Radford, Salary.com)
- Gemini 3.5 Flash: neither first, one named (first choices: Pave) (alternatives: Carta Total Comp, Figures, Payscale, Radford, Ravio, Salary.com)
- Kimi K2: neither first, one named (first choices: Payscale, SalaryCube) (alternatives: Mercer, Radford, Salary.com)
- GLM 4.7 FlashX: neither first, one named (first choices: Pave) (alternatives: Aon Radford, CaptivateIQ, Ficstar, Mercer, OpenComp, Payscale Ascent, Ravio, Salary.com CompAnalyst)
- Claude Haiku 4.5: neither named (first choices: Payscale) (alternatives: Lattice, SalaryCube, beqom)
- Perplexity Sonar: neither named (first choices: Payscale) (alternatives: Ravio, SalaryCube)
- Grok 4.1 Fast: neither named (first choices: Payscale) (alternatives: Pave, Salary.com)
- Mistral Small: neither named (first choices: SalaryCube) (alternatives: Pave, Payfactors, Salary.com)
- DeepSeek V4 Flash: neither named (first choices: Payscale) (alternatives: ERI, Pave, Ravio, Salary.com)
- Llama 4 Maverick: neither named (first choices: Payscale)
- Qwen 3.7 Flash: neither named (first choices: Payscale) (alternatives: Pave, Payfactors, Salary.com, SalaryCube)
- MiniMax M2.5: neither named (first choices: Payscale) (alternatives: Comprehensive.io)

## What the models said about Mercer

- "Avoid over-investing in Mercer/WTW unless you have global complexity" (Kimi K2, paraphrase prompt, hard negative)
- "Avoid traditional surveys (Mercer/Radford: $10k+/year) unless scaling." (Grok 4.1 Fast, budget prompt, hard negative)
- "high-end, legacy enterprise surveys like Mercer, Radford (Aon), or Willis Towers Watson (WTW)—which can easily cost $20,000 to $40,000+ annually—are generally out of reach" (Gemini 3.5 Flash, budget prompt, soft negative)
- "Buy one major survey (like Robert Half or Mercer) to ensure you are paying within the 25th–75th percentiles" (Qwen 3.7 Flash, paraphrase prompt, first choice)
- "The strongest fit from the results is Mercer or Salary.com/Compdata for broad benchmarking" (Perplexity Sonar, paraphrase prompt, first choice)
- "Especially strong in pay and total rewards benchmarking, plus HR strategy." (GPT-5.4 mini, comparative prompt, first choice)

## What the models said about Radford

- "What to Avoid on a Limited Budget ... enterprise-grade pricing with far more data than most small/mid-market companies need" (Kimi K2, budget prompt, hard negative)
- "Avoid traditional surveys (Mercer/Radford: $10k+/year) unless scaling." (Grok 4.1 Fast, budget prompt, hard negative)
- "it requires active participation (submitting your own data) and can be expensive and labor-intensive to manage compared to automated tools" (Gemini 3.5 Flash, paraphrase prompt, soft negative)
- "I'd typically recommend starting with Radford if you're in tech/SaaS, or Mercer for broader B2B sectors" (Claude Haiku 4.5, paraphrase prompt, first choice)
- "Mercer or Radford-style benchmark data is often the strongest foundation" (GPT-5.4 mini, paraphrase prompt, first choice)
- "Tech-First / Specialized Platforms (Radford, Pave, CompAnalyst, Option Impact) — Best for: Scalability, ease of use, and equity/tech-heavy roles." (Qwen 3.7 Flash, scale prompt, alternative)

Share is the count of first choices across the direct, paraphrase, budget and scale prompts over all twelve models, for a mid-market B2B company; rank is within the category. Comparisons are drawn for the top eight products in each category. Published under CC BY 4.0; the output is the models' output, and nothing here is a recommendation by the index.