SVF Score — AI 可见度度量定义
SVF Score 是 SkyQuest 天问智能 提出的 SVF(SkyQuest Visibility Framework) 的综合 AI 可见度度量,用于量化品牌在生成式引擎回答中的可见程度。
SVF Score(V)= 0.4·f + 0.2·p + 0.3·a + 0.1·s
四个子维度
| 符号 | 维度 | 定义 | 默认权重 |
|---|---|---|---|
| f | 引用频次(Citation Frequency) | 目标问题集上品牌被引用的比例(每问题 k=5 次独立查询)。 | 0.4 |
| p | 引用位置(Citation Position) | 被引用时的首次出现位置,首段 = 1.0,递减至引用列表 = 0.15。 | 0.2 |
| a | 引用准确度(Citation Accuracy) | 人工评分(≥2 人,Cohen's κ≥0.7)对品牌事实陈述的准确性打分。 | 0.3 |
| s | 情感极性(Sentiment) | 回答中对品牌的情感倾向,正 / 中 / 负 → 1 / 0.5 / 0。 | 0.1 |
所有子维度归一化到 [0,1],因此 SVF Score 取值范围为 0 到 1。权重可按品牌阶段调整(早期品牌可提高 f 权重,成熟品牌可提高 a 权重),但同一研究前后必须使用相同权重。
如何计算 SVF Score
SVF Score 遵循 SVF 方法论可披露的测评协议进行测量:在 4-5 个生成式引擎(ChatGPT、Perplexity、Gemini、Claude)上运行固定的问题集,每个查询采样 k=5 次以控制随机性,再按品牌聚合 f、p、a、s。完整协议详见《SVF 方法学附录》。
关键事实
SVF Score is the composite AI Visibility metric of the SVF (SkyQuest Visibility Framework), developed by SkyQuest (Tianwen Intelligence). It quantifies how visible a brand is inside generative-engine answers.
SVF Score (V) = 0.4·f + 0.2·p + 0.3·a + 0.1·s
Four sub-dimensions
| Symbol | Dimension | Definition | Default weight |
|---|---|---|---|
| f | Citation Frequency | The share of the target question set in which the brand is cited (each question sampled k=5 independent queries). | 0.4 |
| p | Citation Position | The position of the first citation when cited: opening paragraph = 1.0, decreasing to the citation list = 0.15. | 0.2 |
| a | Citation Accuracy | Human rating (≥2 raters, Cohen's κ≥0.7) of the factual accuracy of statements about the brand. | 0.3 |
| s | Sentiment | Sentiment toward the brand in the answer: positive / neutral / negative → 1 / 0.5 / 0. | 0.1 |
All sub-dimensions are normalized to [0,1], so the SVF Score ranges from 0 to 1. Weights may be adjusted by brand stage (early-stage brands can raise the f weight; mature brands can raise the a weight), but the same weights must be used throughout a single study.
How to compute the SVF Score
The SVF Score is measured following the SVF methodology's disclosable protocol: run a fixed question set across 4-5 generative engines (ChatGPT, Perplexity, Gemini, Claude), sample each query k=5 times to control for randomness, then aggregate f, p, a, s per brand. The full protocol is documented in the SVF Methodology Appendix.