SourceBench - Shopping: leaderboard
Metric: Weighted source-quality score on a 20 to 100 scale: every web page an answer cites is rated 1 to 5 on eight rubrics (content relevance, factual accuracy and objectivity weighted 1.5; freshness, ownership accountability, author accountability, domain authority and layout clarity weighted 1) by a few-shot Qwen-Plus evaluator calibrated on human labels, and the weighted mean is scaled so 100 is the maximum; the 20 product search queries (VA-COS NLQ); search-equipped LLMs answer real user queries with their own web search, and the URLs in the final answer are scored; higher is better. Source: arxiv.org. Saturation forecast: Around June 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 85.55 | #91 |
| 2 | Grok 4.1 Fast (Non-reasoning) | 84.49 | #208 (Grok 4.1 Fast) |
| 3 | Gemini 3 Pro | 81.36 | #77 |
| 4 | Gemini 2.5 Flash | 81.02 | #237 |
| 5 | GPT-4o | 80.88 | #333 |
| 6 | Claude Sonnet 4.5 | 80.1 | #138 |
| 7 | Sonar Pro | 79.39 | #268 |
| 8 | Gemini 3 Flash | 78.73 | #93 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=sourcebench-shopping · How It Works · Data refreshed daily, snapshot 2026-10-11.