From 601e448fa82cd85692fe053693913a076efe24f4 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Mon, 20 Jul 2026 15:20:22 +0800
Subject: [PATCH 1/2] feat(seo): lead compare/blog titles with search phrase,
 stat-led + trimmed meta descriptions
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Improve SERP click-through on page-1 compare pages and blog posts.

- Compare <title> now leads with the GPU pair (the actual query, e.g.
  "B200 vs B300: GLM-5 Inference Benchmark | InferenceX") via title.absolute,
  bypassing the long root template; falls back "Inference Benchmark"→"Benchmark"
  for long pairs. New compareModelSeoName (single natural version name) +
  compareSeoTitle in compare-slug.ts.
- Compare meta description is now stat-led from interpolated head-to-head
  numbers at the default interactivity target (tok/s/GPU + $/M-tok), ≤155 chars,
  gracefully falling back to compact boilerplate for sparse-data pairs.
  New compareMetaDescription / computeCompareStat in compare-ssr.ts.
- Blog <title> uses the shorter "| InferenceX" suffix (title.absolute).
- Blog meta/OG/Twitter descriptions use blogDescription = seoDescription ??
  smartTruncate(subtitle, 155) (word-boundary cut, no mid-sentence truncation).
  Added optional seoDescription frontmatter to 9 long-subtitle posts.
- Mirrored every change on /zh (compareMetaDescriptionZh, zh titles, zh
  seoDescription on the 9 translated posts).
- Unit tests for compareModelSeoName, compareSeoTitle, compareMetaDescription
  (+ zh structural port), smartTruncate, and blogDescription.

中文：让 /compare 与博客页面的 SEO 元数据以搜索短语开头，提升谷歌搜索结果页点击率。
compare 的 <title> 改为以 GPU 对（真实搜索词，如“B200 vs B300: GLM-5 Inference
Benchmark | InferenceX”）开头，使用 title.absolute 绕过冗长的根模板，长组合时将
“Inference Benchmark”缩短为“Benchmark”；新增 compareModelSeoName（单一自然版本名）
与 compareSeoTitle。compare 的 meta 描述改为在默认交互性目标下由插值出的正面对比
数据（tok/s/GPU 与 $/M-tok）驱动，控制在 155 字符内，数据稀疏时优雅回退到精简样板
文案；新增 compareMetaDescription / computeCompareStat。博客 <title> 改用更短的
“| InferenceX”后缀。博客 meta/OG/Twitter 描述改用 blogDescription =
seoDescription ?? smartTruncate(subtitle, 155)（按词边界截断，避免句中截断），并为
9 篇长副标题文章新增 seoDescription frontmatter。所有改动均在 /zh 侧同步（新增
compareMetaDescriptionZh、中文标题、9 篇译文的中文 seoDescription）。补充相应单元测试。
---
 ...nvfp4-vs-h200-fp8-3-6x-perf-per-dollar.mdx |   1 +
 ...vllm-nvfp4-vs-h100-fp8-perf-per-dollar.mdx |   1 +
 ...h200-int4-kimi-k2-vllm-perf-per-dollar.mdx |   1 +
 ...b200-disagg-deepseek-r1-fp4-dynamo-trt.mdx |   1 +
 ...nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4.mdx |   1 +
 ...max-open-source-inference-benchmarking.mdx |   1 +
 ...deepseek-v4-pro-sglang-110x-in-26-days.mdx |   1 +
 ...x-glm5-fp8-sglang-40-cheaper-than-b200.mdx |   1 +
 ...i355x-qwen3-5-sglang-v0-5-12-up-to-17x.mdx |   1 +
 ...nvfp4-vs-h200-fp8-3-6x-perf-per-dollar.mdx |   1 +
 ...vllm-nvfp4-vs-h100-fp8-perf-per-dollar.mdx |   1 +
 ...h200-int4-kimi-k2-vllm-perf-per-dollar.mdx |   1 +
 ...b200-disagg-deepseek-r1-fp4-dynamo-trt.mdx |   1 +
 ...nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4.mdx |   1 +
 ...max-open-source-inference-benchmarking.mdx |   1 +
 ...deepseek-v4-pro-sglang-110x-in-26-days.mdx |   1 +
 ...x-glm5-fp8-sglang-40-cheaper-than-b200.mdx |   1 +
 ...i355x-qwen3-5-sglang-v0-5-12-up-to-17x.mdx |   1 +
 packages/app/src/app/blog/[slug]/page.tsx     |  13 +-
 packages/app/src/app/compare/[slug]/page.tsx  |  36 ++--
 packages/app/src/app/zh/blog/[slug]/page.tsx  |  19 +-
 .../app/src/app/zh/compare/[slug]/page.tsx    |  25 ++-
 packages/app/src/lib/blog.test.ts             |  56 ++++++
 packages/app/src/lib/blog.ts                  |  36 ++++
 packages/app/src/lib/compare-slug.test.ts     |  65 +++++++
 packages/app/src/lib/compare-slug.ts          |  41 +++++
 packages/app/src/lib/compare-ssr-zh.ts        |  71 +++++++-
 packages/app/src/lib/compare-ssr.test.ts      | 170 +++++++++++++++++-
 packages/app/src/lib/compare-ssr.ts           | 146 +++++++++++++++
 .../app/src/lib/compare-variant-ssr.test.ts   |   1 +
 30 files changed, 666 insertions(+), 31 deletions(-)

diff --git a/packages/app/content/blog/b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar.mdx b/packages/app/content/blog/b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar.mdx
index 217217f6..38592c9a 100644
--- a/packages/app/content/blog/b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar.mdx
+++ b/packages/app/content/blog/b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 vs H200 FP8 on GLM-5: Up to 3.65x Better Performance per Dollar with SGLang MTP'
 subtitle: 'Both SKUs run SGLang EAGLE MTP; the Blackwell generation lifts perf/$ by ~1.2x at the peak and the NVIDIA GLM-5-NVFP4 checkpoint on FlashInfer TRT-LLM sparse MLA stacks another ~2.4–3.0x on 8K/1K'
+seoDescription: 'B200 NVFP4 delivers up to 3.65x better perf/$ than H200 FP8 on GLM-5 with SGLang MTP, stacking Blackwell gains and a sparse-MLA checkpoint.'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:
diff --git a/packages/app/content/blog/b200-minimax-m2-5-vllm-nvfp4-vs-h100-fp8-perf-per-dollar.mdx b/packages/app/content/blog/b200-minimax-m2-5-vllm-nvfp4-vs-h100-fp8-perf-per-dollar.mdx
index e0e1f710..24b36501 100644
--- a/packages/app/content/blog/b200-minimax-m2-5-vllm-nvfp4-vs-h100-fp8-perf-per-dollar.mdx
+++ b/packages/app/content/blog/b200-minimax-m2-5-vllm-nvfp4-vs-h100-fp8-perf-per-dollar.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 vs H100 FP8 on MiniMax-M2.5: Up to 8.2x Better Performance per Dollar with vLLM'
 subtitle: 'vLLM PR #36307 unlocks the trtllm-gen FP8 MoE kernel for MiniMax on B200; combined with NVFP4, perf/$ scales from 4.0x at 22 tok/s/user to 8.2x at 110 on 8K/1K'
+seoDescription: 'B200 NVFP4 delivers up to 8.2x better perf/$ than H100 FP8 on MiniMax-M2.5 with vLLM — trtllm-gen FP8 MoE plus NVFP4 on 8K/1K.'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:
diff --git a/packages/app/content/blog/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar.mdx b/packages/app/content/blog/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar.mdx
index 4fb1167c..538aa54f 100644
--- a/packages/app/content/blog/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar.mdx
+++ b/packages/app/content/blog/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 vs H200 INT4 on Kimi K2.5/K2.6: Up to 2.95x Better Performance per Dollar'
 subtitle: "On vLLM 8K/1K the NVFP4 path on B200 is 2.71x–2.95x cheaper per million tokens than H200 INT4 across the entire 30–90 tok/s/user serving band, and 2.45x–2.74x cheaper than B200 INT4 on the same silicon. Both factors decompose cleanly into B200's HBM bandwidth, HBM capacity, and NVFP4 tensor cores"
+seoDescription: 'B200 NVFP4 runs Kimi K2.5/K2.6 up to 2.95x cheaper per token than H200 INT4 on vLLM 8K/1K, across the full 30–90 tok/s/user serving band.'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:
diff --git a/packages/app/content/blog/gb200-nvl72-vs-b200-disagg-deepseek-r1-fp4-dynamo-trt.mdx b/packages/app/content/blog/gb200-nvl72-vs-b200-disagg-deepseek-r1-fp4-dynamo-trt.mdx
index a65b8a21..aa4bb2ff 100644
--- a/packages/app/content/blog/gb200-nvl72-vs-b200-disagg-deepseek-r1-fp4-dynamo-trt.mdx
+++ b/packages/app/content/blog/gb200-nvl72-vs-b200-disagg-deepseek-r1-fp4-dynamo-trt.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'GB200 NVL72 vs B200 on DeepSeek R1 670B: Up to 4.4x Throughput per GPU at 125 tok/s/user'
 subtitle: "DeepSeek R1 FP4 1k/1k. NVL72's 72-GPU NVLink scale-up fabric lets decode run wide EP up to EP=32, where B200's 8-GPU NVLink island caps out at EP=8 over RoCEv2"
+seoDescription: 'GB200 NVL72 hits up to 4.4x more per-GPU throughput than B200 on DeepSeek R1 670B at 125 tok/s/user — 72-GPU NVLink enables wide EP=32.'
 date: '2026-05-23'
 publishDate: '2026-05-23'
 tags:
diff --git a/packages/app/content/blog/gb300-nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4.mdx b/packages/app/content/blog/gb300-nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4.mdx
index 4f8fac9e..fd310844 100644
--- a/packages/app/content/blog/gb300-nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4.mdx
+++ b/packages/app/content/blog/gb300-nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'GB300 NVL72 vs GB200 NVL72 Inference Performance & Perf per Dollar - on DeepSeek-V4-Pro 1.6T: Up to 2.83x Throughput'
 subtitle: "DSv4-Pro FP4 8K/1K, Dynamo+vLLM, disaggregated on both racks. GB300's 50% extra HBM (288 vs 192 GB/GPU) unlocks a wider prefill+decode recipe GB200 can't fit — lifting middle-of-curve perf/$ by 2.31x despite a 20% per-GPU TCO premium."
+seoDescription: 'GB300 NVL72 adds 50% more HBM, lifting DeepSeek-V4-Pro perf/$ up to 2.31x over GB200 NVL72 on vLLM FP4 8K/1K — up to 2.83x throughput per GPU.'
 date: '2026-05-27'
 publishDate: '2026-05-27'
 tags:
diff --git a/packages/app/content/blog/inferencemax-open-source-inference-benchmarking.mdx b/packages/app/content/blog/inferencemax-open-source-inference-benchmarking.mdx
index 4b6da177..32217ffa 100644
--- a/packages/app/content/blog/inferencemax-open-source-inference-benchmarking.mdx
+++ b/packages/app/content/blog/inferencemax-open-source-inference-benchmarking.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'InferenceMAX: Open Source Inference Benchmarking'
 subtitle: 'NVIDIA GB200 NVL72, AMD MI355X, Throughput Token per GPU, Latency Tok/s/user, Perf per Dollar, Cost per Million Tokens, Tokens per Provisioned Megawatt, DeepSeek R1 670B, GPTOSS 120B, Llama3 70B'
+seoDescription: 'InferenceMAX: open-source inference benchmarks across NVIDIA GB200 NVL72 and AMD MI355X — throughput, latency, cost per token and perf per watt.'
 date: '2025-10-09'
 publishDate: '2025-10-09'
 tags:
diff --git a/packages/app/content/blog/mi355x-deepseek-v4-pro-sglang-110x-in-26-days.mdx b/packages/app/content/blog/mi355x-deepseek-v4-pro-sglang-110x-in-26-days.mdx
index b4e9c850..ed001196 100644
--- a/packages/app/content/blog/mi355x-deepseek-v4-pro-sglang-110x-in-26-days.mdx
+++ b/packages/app/content/blog/mi355x-deepseek-v4-pro-sglang-110x-in-26-days.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'MI355X DeepSeek-V4-Pro on SGLang: 110.5x Throughput per GPU in 26 Days'
 subtitle: 'The amd/deepseek_v4 side branch shipped TileLang attention indexer, Triton sparse MLA, fused RoPE/Hadamard, FlyDSL MoE, and FP4 weights across 31 performance optimizations PRs — lifting first-light 20 tok/s/GPU at 2.4 tok/s/user into 2,256 tok/s/GPU at 9.4 tok/s/user on 8K/1K, with both throughput and interactivity climbing together'
+seoDescription: 'MI355X ran DeepSeek-V4-Pro 110.5x faster per GPU in 26 days on SGLang — from 20 to 2,256 tok/s/GPU via TileLang, sparse MLA and FP4 on 8K/1K.'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:
diff --git a/packages/app/content/blog/mi355x-glm5-fp8-sglang-40-cheaper-than-b200.mdx b/packages/app/content/blog/mi355x-glm5-fp8-sglang-40-cheaper-than-b200.mdx
index 353df6b4..b4212a6a 100644
--- a/packages/app/content/blog/mi355x-glm5-fp8-sglang-40-cheaper-than-b200.mdx
+++ b/packages/app/content/blog/mi355x-glm5-fp8-sglang-40-cheaper-than-b200.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'AMD MI355X GLM-5 Inference: Up to 40% Cheaper per Million Tokens than B200 on SGLang FP8'
 subtitle: '14 weeks after GLM-5 launched, AMD landed both MTP and non-MTP SGLang FP8 recipes on MI355X — fused MLA + FP8 KV cache via TileLang flips the single-node FP8 cost curve in AMD favor across most of the performance Pareto'
+seoDescription: 'AMD MI355X runs GLM-5 up to 40% cheaper per million tokens than B200 on SGLang FP8, with fused MLA and FP8 KV cache via TileLang.'
 date: '2026-05-25'
 publishDate: '2026-05-25'
 tags:
diff --git a/packages/app/content/blog/mi355x-qwen3-5-sglang-v0-5-12-up-to-17x.mdx b/packages/app/content/blog/mi355x-qwen3-5-sglang-v0-5-12-up-to-17x.mdx
index 329e4479..94983de0 100644
--- a/packages/app/content/blog/mi355x-qwen3-5-sglang-v0-5-12-up-to-17x.mdx
+++ b/packages/app/content/blog/mi355x-qwen3-5-sglang-v0-5-12-up-to-17x.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'AMD MI355X Qwen3.5 397B-A17B Inference: Up to 19x Throughput per GPU in 3 Months on SGLang FP8'
 subtitle: 'From v0.5.8 (Feb) → v0.5.10rc0 (Apr) → v0.5.12 (May), three AITER kernel landings on MI355X plus a TP=8 → TP=2/TP=4 retune push Qwen3.5 8k/1k peak from 1.3k to 6.4k tok/s/GPU and extend the curve out to 75 tok/s/user'
+seoDescription: 'AMD MI355X lifted Qwen3.5 397B-A17B up to 19x per-GPU throughput in 3 months on SGLang FP8 — 1.3k to 6.4k tok/s/GPU via AITER kernels.'
 date: '2026-05-25'
 publishDate: '2026-05-25'
 tags:
diff --git a/packages/app/content/blog/zh/b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar.mdx b/packages/app/content/blog/zh/b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar.mdx
index 36d30314..ec6db67a 100644
--- a/packages/app/content/blog/zh/b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar.mdx
+++ b/packages/app/content/blog/zh/b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 对比 H200 FP8 运行 GLM-5：SGLang MTP 下性价比提升高达 3.65 倍'
 subtitle: '两款 GPU 均运行 SGLang EAGLE MTP；Blackwell 世代在峰值处带来约 1.2 倍的性价比提升，NVIDIA GLM-5-NVFP4 检查点搭配 FlashInfer TRT-LLM 稀疏 MLA 在 8K/1K 场景下再叠加约 2.4–3.0 倍优势'
+seoDescription: 'B200 NVFP4 在 SGLang MTP 下运行 GLM-5，性价比比 H200 FP8 最高提升 3.65 倍——叠加 Blackwell 世代增益与稀疏 MLA 检查点。'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:
diff --git a/packages/app/content/blog/zh/b200-minimax-m2-5-vllm-nvfp4-vs-h100-fp8-perf-per-dollar.mdx b/packages/app/content/blog/zh/b200-minimax-m2-5-vllm-nvfp4-vs-h100-fp8-perf-per-dollar.mdx
index 73da4323..a889133c 100644
--- a/packages/app/content/blog/zh/b200-minimax-m2-5-vllm-nvfp4-vs-h100-fp8-perf-per-dollar.mdx
+++ b/packages/app/content/blog/zh/b200-minimax-m2-5-vllm-nvfp4-vs-h100-fp8-perf-per-dollar.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 vs H100 FP8 运行 MiniMax-M2.5：vLLM 下每美元性能最高提升 8.2 倍'
 subtitle: 'vLLM PR #36307 为 MiniMax 在 B200 上解锁了 trtllm-gen FP8 MoE 模块化内核；结合 NVFP4，在 8K/1K 负载下性能/成本从 22 tok/s/user 时的 4.0 倍扩大到 110 tok/s/user 时的 8.2 倍'
+seoDescription: 'B200 NVFP4 在 vLLM 下运行 MiniMax-M2.5，每美元性能比 H100 FP8 最高提升 8.2 倍——trtllm-gen FP8 MoE 结合 NVFP4（8K/1K）。'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:
diff --git a/packages/app/content/blog/zh/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar.mdx b/packages/app/content/blog/zh/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar.mdx
index 38e84357..04c09450 100644
--- a/packages/app/content/blog/zh/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar.mdx
+++ b/packages/app/content/blog/zh/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'B200 NVFP4 对比 H200 INT4 运行 Kimi K2.5/K2.6：性价比提升高达 2.95 倍'
 subtitle: '在 vLLM 8K/1K 工作负载下，B200 NVFP4 路径在 30–90 tok/s/user 推理区间内每百万 tokens 成本比 H200 INT4 低 2.71x–2.95x，比同一 B200 硬件上的 INT4 低 2.45x–2.74x。三个因素——B200 的 HBM 带宽、HBM 容量和 NVFP4 张量核心——可清晰分解该优势'
+seoDescription: '在 vLLM 8K/1K 下，B200 NVFP4 运行 Kimi K2.5/K2.6 每 token 成本比 H200 INT4 最多低 2.95 倍，覆盖 30–90 tok/s/user 全区间。'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:
diff --git a/packages/app/content/blog/zh/gb200-nvl72-vs-b200-disagg-deepseek-r1-fp4-dynamo-trt.mdx b/packages/app/content/blog/zh/gb200-nvl72-vs-b200-disagg-deepseek-r1-fp4-dynamo-trt.mdx
index c1cd0c43..e70a6620 100644
--- a/packages/app/content/blog/zh/gb200-nvl72-vs-b200-disagg-deepseek-r1-fp4-dynamo-trt.mdx
+++ b/packages/app/content/blog/zh/gb200-nvl72-vs-b200-disagg-deepseek-r1-fp4-dynamo-trt.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'GB200 NVL72 对比 B200 运行 DeepSeek R1 670B：在 125 tok/s/user 下每 GPU 吞吐量最高达 4.4 倍'
 subtitle: 'DeepSeek R1 FP4 1k/1k。NVL72 的 72-GPU NVLink 扩展域允许解码使用最高 EP=32 的宽专家并行，而 B200 的 8-GPU NVLink 岛通过 RoCEv2 上限为 EP=8'
+seoDescription: 'GB200 NVL72 在 125 tok/s/user 下运行 DeepSeek R1 670B，每 GPU 吞吐量比 B200 最高提升 4.4 倍——72-GPU NVLink 支持 EP=32 宽专家并行。'
 date: '2026-05-23'
 publishDate: '2026-05-23'
 tags:
diff --git a/packages/app/content/blog/zh/gb300-nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4.mdx b/packages/app/content/blog/zh/gb300-nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4.mdx
index 38fb02d1..9970710b 100644
--- a/packages/app/content/blog/zh/gb300-nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4.mdx
+++ b/packages/app/content/blog/zh/gb300-nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'GB300 NVL72 vs GB200 NVL72 推理性能与性价比对比 — DeepSeek-V4-Pro 1.6T：吞吐量最高提升 2.83 倍'
 subtitle: 'DSv4-Pro FP4 8K/1K，Dynamo+vLLM，两套机架均采用分离式部署。GB300 多出 50% 的 HBM（每 GPU 288 GB vs 192 GB）解锁了 GB200 无法容纳的更宽预填充+解码配方——尽管单 GPU TCO 溢价 20%，曲线中段性价比仍提升 2.31 倍。'
+seoDescription: 'GB300 NVL72 多出的 50% HBM 使 DeepSeek-V4-Pro 性价比比 GB200 NVL72 最高提升 2.31 倍（vLLM FP4 8K/1K），每 GPU 吞吐量最高提升 2.83 倍。'
 date: '2026-05-27'
 publishDate: '2026-05-27'
 tags:
diff --git a/packages/app/content/blog/zh/inferencemax-open-source-inference-benchmarking.mdx b/packages/app/content/blog/zh/inferencemax-open-source-inference-benchmarking.mdx
index ecad0f7f..4958133b 100644
--- a/packages/app/content/blog/zh/inferencemax-open-source-inference-benchmarking.mdx
+++ b/packages/app/content/blog/zh/inferencemax-open-source-inference-benchmarking.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'InferenceMAX：开源推理基准测试'
 subtitle: 'NVIDIA GB200 NVL72、AMD MI355X、每 GPU 吞吐量 Token、延迟 Tok/s/user、性价比、每百万 Token 成本、每配置兆瓦 Token 数、DeepSeek R1 670B、GPTOSS 120B、Llama3 70B'
+seoDescription: 'InferenceMAX：面向 NVIDIA GB200 NVL72 与 AMD MI355X 的开源推理基准测试——吞吐量、延迟、每 token 成本与每瓦性能。'
 date: '2025-10-09'
 publishDate: '2025-10-09'
 tags:
diff --git a/packages/app/content/blog/zh/mi355x-deepseek-v4-pro-sglang-110x-in-26-days.mdx b/packages/app/content/blog/zh/mi355x-deepseek-v4-pro-sglang-110x-in-26-days.mdx
index 8a6134f2..a95695ad 100644
--- a/packages/app/content/blog/zh/mi355x-deepseek-v4-pro-sglang-110x-in-26-days.mdx
+++ b/packages/app/content/blog/zh/mi355x-deepseek-v4-pro-sglang-110x-in-26-days.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'MI355X 上 DeepSeek-V4-Pro 搭配 SGLang：26 天内每 GPU 吞吐量提升 110.5 倍'
 subtitle: 'amd/deepseek_v4 分支合入了 TileLang 注意力索引器、Triton 稀疏 MLA、融合 RoPE/Hadamard、FlyDSL MoE 以及 FP4 权重，历经 31 个性能优化 PR——将首次点亮时 20 tok/s/GPU、2.4 tok/s/user 的水平提升至 8K/1K 负载下 2,256 tok/s/GPU、9.4 tok/s/user，吞吐量与交互性同步攀升'
+seoDescription: 'MI355X 在 SGLang 上 26 天内将 DeepSeek-V4-Pro 的每 GPU 吞吐量提升 110.5 倍——8K/1K 下经 TileLang、稀疏 MLA 与 FP4 从 20 提升至 2,256 tok/s/GPU。'
 date: '2026-05-26'
 publishDate: '2026-05-26'
 tags:
diff --git a/packages/app/content/blog/zh/mi355x-glm5-fp8-sglang-40-cheaper-than-b200.mdx b/packages/app/content/blog/zh/mi355x-glm5-fp8-sglang-40-cheaper-than-b200.mdx
index a4817108..54067049 100644
--- a/packages/app/content/blog/zh/mi355x-glm5-fp8-sglang-40-cheaper-than-b200.mdx
+++ b/packages/app/content/blog/zh/mi355x-glm5-fp8-sglang-40-cheaper-than-b200.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'AMD MI355X GLM-5 推理：SGLang FP8 单节点每百万 token 成本比 B200 最高低 40%'
 subtitle: 'GLM-5 发布 14 周后，AMD 在 MI355X 上同时实现了 SGLang FP8 的 MTP 和非 MTP 方案 — 通过 TileLang 实现的融合 MLA + FP8 KV 缓存在大部分性能 Pareto 前沿上将单节点 FP8 成本曲线翻转为 AMD 占优'
+seoDescription: 'AMD MI355X 在 SGLang FP8 上运行 GLM-5，每百万 token 成本比 B200 最高低 40%——TileLang 融合 MLA 与 FP8 KV 缓存。'
 date: '2026-05-25'
 publishDate: '2026-05-25'
 tags:
diff --git a/packages/app/content/blog/zh/mi355x-qwen3-5-sglang-v0-5-12-up-to-17x.mdx b/packages/app/content/blog/zh/mi355x-qwen3-5-sglang-v0-5-12-up-to-17x.mdx
index ac7908b2..5e2e6738 100644
--- a/packages/app/content/blog/zh/mi355x-qwen3-5-sglang-v0-5-12-up-to-17x.mdx
+++ b/packages/app/content/blog/zh/mi355x-qwen3-5-sglang-v0-5-12-up-to-17x.mdx
@@ -1,6 +1,7 @@
 ---
 title: 'AMD MI355X Qwen3.5 397B-A17B 推理：SGLang FP8 三个月内每 GPU 吞吐量提升最高 19 倍'
 subtitle: '从 v0.5.8（2 月）→ v0.5.10rc0（4 月）→ v0.5.12（5 月），三次 AITER 内核合入 MI355X 加上从 TP=8 到 TP=2/TP=4 的重新调优，将 Qwen3.5 8k/1k 峰值从 1.3k 推高至 6.4k tok/s/GPU，并将曲线延伸至 75 tok/s/user'
+seoDescription: 'AMD MI355X 在 SGLang FP8 上 3 个月内将 Qwen3.5 397B-A17B 每 GPU 吞吐量最高提升 19 倍——经 AITER 内核从 1.3k 提升至 6.4k tok/s/GPU。'
 date: '2026-05-25'
 publishDate: '2026-05-25'
 tags:
diff --git a/packages/app/src/app/blog/[slug]/page.tsx b/packages/app/src/app/blog/[slug]/page.tsx
index 05bda7e6..e21b5b12 100644
--- a/packages/app/src/app/blog/[slug]/page.tsx
+++ b/packages/app/src/app/blog/[slug]/page.tsx
@@ -16,6 +16,7 @@ import { ShareTwitterButton, ShareLinkedInButton } from '@/components/share-butt
 import { Card } from '@/components/ui/card';
 import { JsonLd } from '@/components/json-ld';
 import {
+  blogDescription,
   getAllPosts,
   getAdjacentPosts,
   extractHeadings,
@@ -43,10 +44,14 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const result = getPostBySlug(slug);
   if (!result) return {};
   const { meta } = result;
+  const description = blogDescription(meta);
 
   return {
-    title: meta.title,
-    description: meta.subtitle,
+    // `absolute` keeps a short " | InferenceX" suffix instead of the long
+    // "%s | InferenceX by SemiAnalysis" root template, leaving more room for
+    // the headline before Google truncates the SERP title.
+    title: { absolute: `${meta.title} | ${SITE_NAME}` },
+    description,
     keywords: meta.tags,
     authors: [{ name: AUTHOR_NAME }],
     alternates: {
@@ -56,7 +61,7 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     },
     openGraph: {
       title: `${meta.title} | ${SITE_NAME}`,
-      description: meta.subtitle,
+      description,
       url: `${SITE_URL}/blog/${slug}`,
       type: 'article',
       publishedTime: `${meta.date}T00:00:00Z`,
@@ -67,7 +72,7 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     twitter: {
       card: 'summary_large_image',
       title: meta.title,
-      description: meta.subtitle,
+      description,
       site: AUTHOR_HANDLE,
       creator: AUTHOR_HANDLE,
     },
diff --git a/packages/app/src/app/compare/[slug]/page.tsx b/packages/app/src/app/compare/[slug]/page.tsx
index 763b000f..b11011d5 100644
--- a/packages/app/src/app/compare/[slug]/page.tsx
+++ b/packages/app/src/app/compare/[slug]/page.tsx
@@ -1,12 +1,7 @@
 import type { Metadata } from 'next';
 import { notFound, permanentRedirect } from 'next/navigation';
 
-import {
-  HW_REGISTRY,
-  SITE_NAME,
-  SITE_URL,
-  SUPPORTERS_LINE,
-} from '@semianalysisai/inferencex-constants';
+import { HW_REGISTRY, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
 
 import { JsonLd } from '@/components/json-ld';
 import { languageAlternates } from '@/lib/i18n';
@@ -15,11 +10,14 @@ import {
   canonicalCompareSlug,
   compareDisplayLabel,
   compareModelDisplayLabel,
+  compareModelSeoName,
+  compareSeoTitle,
   parseCompareSlug,
 } from '@/lib/compare-slug';
 import {
   buildBreadcrumbJsonLd,
   buildJsonLd,
+  compareMetaDescription,
   compareTableNarrative,
   computeCompareTableData,
   dateRangeForPair,
@@ -46,16 +44,30 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
   if (!parsed) return {};
   const fullLabel = compareModelDisplayLabel(parsed.model, parsed.a, parsed.b);
   const gpuLabel = compareDisplayLabel(parsed.a, parsed.b);
-  const url = `${SITE_URL}/compare/${canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b)}`;
-  const description = `${gpuLabel} inference benchmark on ${parsed.model.label}: verified, reproducible head-to-head results from InferenceX, the independent open-source GPU benchmark by SemiAnalysis. ${SUPPORTERS_LINE} Compare latency, throughput & cost.`;
+  const modelSeoName = compareModelSeoName(parsed.model);
+  const canonical = canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b);
+  const url = `${SITE_URL}/compare/${canonical}`;
+
+  // Lead the SEO title with the GPU pair — that's the phrase people search
+  // ("b200 vs b300") and must survive Google's ~60-char SERP truncation. The
+  // `absolute` form bypasses the long "%s | InferenceX by SemiAnalysis" root
+  // template so the query isn't pushed off the end.
+  const title = `${compareSeoTitle(gpuLabel, modelSeoName)} | ${SITE_NAME}`;
+
+  // Stat-led meta description built from the interpolated head-to-head numbers
+  // at the slug's default operating point (falls back to boilerplate for
+  // sparse-data pairs). Fetch is blob-cached and shared with the page render.
+  const rows = await getCachedBenchmarks(parsed.model.dbKeys);
+  const { sequence, precision } = pickPairDefaults(rows, parsed.a, parsed.b);
+  const { ssrRows } = computeCompareTableData(rows, parsed.a, parsed.b, sequence, precision);
+  const description = compareMetaDescription(parsed.model, parsed.a, parsed.b, ssrRows);
+
   return {
-    title: `${fullLabel} Inference Benchmark`,
+    title: { absolute: title },
     description,
     alternates: {
       canonical: url,
-      languages: languageAlternates(
-        `/compare/${canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b)}`,
-      ),
+      languages: languageAlternates(`/compare/${canonical}`),
     },
     openGraph: {
       title: `${fullLabel} | ${SITE_NAME}`,
diff --git a/packages/app/src/app/zh/blog/[slug]/page.tsx b/packages/app/src/app/zh/blog/[slug]/page.tsx
index 85b93660..391b90a3 100644
--- a/packages/app/src/app/zh/blog/[slug]/page.tsx
+++ b/packages/app/src/app/zh/blog/[slug]/page.tsx
@@ -16,7 +16,13 @@ import { ReadingProgressBar } from '@/components/blog/reading-progress-bar';
 import { ShareTwitterButton, ShareLinkedInButton } from '@/components/share-buttons';
 import { Card } from '@/components/ui/card';
 import { JsonLd } from '@/components/json-ld';
-import { getAllPosts, getAdjacentPosts, extractHeadings, getPostBySlug } from '@/lib/blog';
+import {
+  blogDescription,
+  getAllPosts,
+  getAdjacentPosts,
+  extractHeadings,
+  getPostBySlug,
+} from '@/lib/blog';
 import { ZH_LANG_TAG, ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
 import {
   AUTHOR_HANDLE,
@@ -38,16 +44,19 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const result = getPostBySlug(slug, 'zh');
   if (!result) return {};
   const { meta } = result;
+  const description = blogDescription(meta);
 
   return {
-    title: meta.title,
-    description: meta.subtitle,
+    // Short " | InferenceX" suffix via `absolute` (mirrors the English page)
+    // so the headline keeps more of the SERP title before truncation.
+    title: { absolute: `${meta.title} | ${SITE_NAME}` },
+    description,
     keywords: meta.tags,
     authors: [{ name: AUTHOR_NAME }],
     alternates: zhAlternates(`/blog/${slug}`),
     openGraph: {
       title: `${meta.title} | ${SITE_NAME}`,
-      description: meta.subtitle,
+      description,
       url: `${SITE_URL}/zh/blog/${slug}`,
       type: 'article',
       locale: ZH_OG_LOCALE,
@@ -59,7 +68,7 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     twitter: {
       card: 'summary_large_image',
       title: meta.title,
-      description: meta.subtitle,
+      description,
       site: AUTHOR_HANDLE,
       creator: AUTHOR_HANDLE,
     },
diff --git a/packages/app/src/app/zh/compare/[slug]/page.tsx b/packages/app/src/app/zh/compare/[slug]/page.tsx
index 4cd11d03..2dae1257 100644
--- a/packages/app/src/app/zh/compare/[slug]/page.tsx
+++ b/packages/app/src/app/zh/compare/[slug]/page.tsx
@@ -1,12 +1,7 @@
 import type { Metadata } from 'next';
 import { notFound, permanentRedirect } from 'next/navigation';
 
-import {
-  HW_REGISTRY,
-  SITE_NAME,
-  SITE_URL,
-  SUPPORTERS_LINE_ZH,
-} from '@semianalysisai/inferencex-constants';
+import { HW_REGISTRY, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
 
 import { JsonLd } from '@/components/json-ld';
 import { pickPairDefaults } from '@/lib/compare-pair-defaults';
@@ -14,6 +9,7 @@ import {
   canonicalCompareSlug,
   compareDisplayLabel,
   compareModelDisplayLabel,
+  compareModelSeoName,
   parseCompareSlug,
 } from '@/lib/compare-slug';
 import {
@@ -29,6 +25,7 @@ import {
 import {
   buildBreadcrumbJsonLdZh,
   buildJsonLdZh,
+  compareMetaDescriptionZh,
   compareTableNarrativeZh,
 } from '@/lib/compare-ssr-zh';
 import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
@@ -48,11 +45,23 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
   if (!parsed) return {};
   const fullLabel = compareModelDisplayLabel(parsed.model, parsed.a, parsed.b);
   const gpuLabel = compareDisplayLabel(parsed.a, parsed.b);
+  const modelSeoName = compareModelSeoName(parsed.model);
   const canonical = canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b);
   const url = `${SITE_URL}/zh/compare/${canonical}`;
-  const description = `${gpuLabel} 在 ${parsed.model.label} 上的推理基准测试：来自 InferenceX（SemiAnalysis 推出的独立开源 GPU 基准测试平台）的经验证、可复现的正面对比结果。${SUPPORTERS_LINE_ZH}对比延迟、吞吐量与成本。`;
+
+  // GPU-pair-first title (mirrors the English page) via `absolute` so the long
+  // root template doesn't bury the search phrase. Model name / SKUs stay English.
+  const title = `${gpuLabel}：${modelSeoName} 推理基准测试 | ${SITE_NAME}`;
+
+  // Stat-led Chinese meta description from the interpolated head-to-head numbers
+  // at the slug's default operating point (falls back to boilerplate when thin).
+  const rows = await getCachedBenchmarks(parsed.model.dbKeys);
+  const { sequence, precision } = pickPairDefaults(rows, parsed.a, parsed.b);
+  const { ssrRows } = computeCompareTableData(rows, parsed.a, parsed.b, sequence, precision);
+  const description = compareMetaDescriptionZh(parsed.model, parsed.a, parsed.b, ssrRows);
+
   return {
-    title: `${fullLabel} 推理基准测试`,
+    title: { absolute: title },
     description,
     alternates: zhAlternates(`/compare/${canonical}`),
     openGraph: {
diff --git a/packages/app/src/lib/blog.test.ts b/packages/app/src/lib/blog.test.ts
index 67633e87..e74fb1ea 100644
--- a/packages/app/src/lib/blog.test.ts
+++ b/packages/app/src/lib/blog.test.ts
@@ -2,6 +2,7 @@ import { describe, expect, it, vi, beforeEach, afterEach } from 'vitest';
 import fs from 'node:fs';
 
 import {
+  blogDescription,
   extractHeadings,
   getAdjacentPosts,
   getAllPosts,
@@ -9,6 +10,7 @@ import {
   getReadingTime,
   hasZhTranslation,
   slugify,
+  smartTruncate,
 } from './blog';
 
 const FAKE_MDX = `---
@@ -179,6 +181,60 @@ describe('getReadingTime', () => {
   });
 });
 
+describe('smartTruncate', () => {
+  it('returns the text unchanged (no ellipsis) when at or under the limit', () => {
+    expect(smartTruncate('Short and sweet.', 155)).toBe('Short and sweet.');
+    const exact = 'x'.repeat(20);
+    expect(smartTruncate(exact, 20)).toBe(exact);
+    expect(smartTruncate('  trimmed  ', 155)).toBe('trimmed');
+  });
+
+  it('cuts at a word boundary (never mid-word) and appends an ellipsis', () => {
+    const text = Array.from({ length: 60 }, () => 'word').join(' '); // 60×"word " → 299 chars
+    const out = smartTruncate(text, 155);
+    expect(out.length).toBeLessThanOrEqual(155);
+    expect(out.endsWith('…')).toBe(true);
+    const body = out.slice(0, -1);
+    // Every retained token is a complete "word" — no "wor"/"rd" fragments.
+    expect(body.split(' ').every((t) => t === 'word')).toBe(true);
+  });
+
+  it('strips trailing punctuation before the ellipsis', () => {
+    const text = `${'alpha, '.repeat(40)}beta`; // lands the cut right after a comma
+    const out = smartTruncate(text, 50);
+    const body = out.slice(0, -1);
+    expect(out.endsWith('…')).toBe(true);
+    expect(/[\s,]$/u.test(body)).toBe(false);
+    expect(out.length).toBeLessThanOrEqual(50);
+  });
+
+  it('hard-cuts CJK prose (no spaces) and stays within the limit', () => {
+    const cjk = '推'.repeat(200);
+    const out = smartTruncate(cjk, 155);
+    expect(out.length).toBeLessThanOrEqual(155);
+    expect(out.endsWith('…')).toBe(true);
+  });
+});
+
+describe('blogDescription', () => {
+  it('returns the explicit seoDescription verbatim when present', () => {
+    const meta = { seoDescription: 'Hand-written, punchy, ≤155 chars.', subtitle: 'x'.repeat(300) };
+    expect(blogDescription(meta)).toBe('Hand-written, punchy, ≤155 chars.');
+  });
+
+  it('smart-truncates the subtitle to ≤155 chars when no seoDescription is set', () => {
+    const subtitle = Array.from({ length: 60 }, () => 'word').join(' ');
+    const out = blogDescription({ subtitle });
+    expect(out).toBe(smartTruncate(subtitle, 155));
+    expect(out.length).toBeLessThanOrEqual(155);
+    expect(out.endsWith('…')).toBe(true);
+  });
+
+  it('passes a short subtitle through without an ellipsis', () => {
+    expect(blogDescription({ subtitle: 'A concise subtitle.' })).toBe('A concise subtitle.');
+  });
+});
+
 describe('getAllPosts', () => {
   it('returns an array of posts sorted by date descending', () => {
     vi.spyOn(fs, 'existsSync').mockReturnValue(true);
diff --git a/packages/app/src/lib/blog.ts b/packages/app/src/lib/blog.ts
index b69885ff..8fee5184 100644
--- a/packages/app/src/lib/blog.ts
+++ b/packages/app/src/lib/blog.ts
@@ -9,6 +9,11 @@ export interface BlogFrontmatter {
   subtitle: string;
   modifiedDate?: string;
   publishDate?: string;
+  /** Optional hand-written SEO/social meta description. When set it overrides
+   *  the auto-truncated `subtitle` for the SERP snippet — use it on posts whose
+   *  `subtitle` runs past ~155 chars and would otherwise truncate mid-sentence.
+   *  Keep it ≤155 chars (English) / ≤~78 CJK chars and compelling. */
+  seoDescription?: string;
   tags?: string[];
 }
 
@@ -49,6 +54,37 @@ export function getReadingTime(content: string): number {
   return Math.max(1, Math.ceil(words / WORDS_PER_MINUTE + cjkChars / CJK_CHARS_PER_MINUTE));
 }
 
+/** Trailing whitespace + punctuation (any script), so a truncated snippet
+ *  doesn't end in a dangling comma / dash / open bracket before the ellipsis. */
+const TRAILING_PUNCT_REGEX = /[\s\p{P}]+$/u;
+
+/**
+ * Truncate to at most `max` characters (ellipsis included) without cutting a
+ * word in half. Cuts at the last space that fits, strips trailing punctuation,
+ * and appends "…" only when the text was actually shortened. CJK prose has no
+ * spaces, so it falls back to a hard character cut (which is correct — there
+ * are no word boundaries to preserve).
+ */
+export function smartTruncate(text: string, max: number): string {
+  const clean = text.trim();
+  if (clean.length <= max) return clean;
+  // Reserve one char for the ellipsis so the result is guaranteed ≤ max.
+  const slice = clean.slice(0, max - 1);
+  const lastSpace = slice.lastIndexOf(' ');
+  const boundary = lastSpace > 0 ? slice.slice(0, lastSpace) : slice;
+  return `${boundary.replace(TRAILING_PUNCT_REGEX, '')}…`;
+}
+
+/**
+ * SERP / social meta description for a post: the explicit `seoDescription`
+ * frontmatter when present, otherwise the `subtitle` smart-truncated to a
+ * SERP-safe length. Used for the `<meta name="description">`, OpenGraph, and
+ * Twitter descriptions.
+ */
+export function blogDescription(meta: Pick<BlogPostMeta, 'seoDescription' | 'subtitle'>): string {
+  return meta.seoDescription ?? smartTruncate(meta.subtitle, 155);
+}
+
 export function getAllPosts(locale: BlogLocale = 'en'): BlogPostMeta[] {
   if (!fs.existsSync(CONTENT_DIR)) return [];
 
diff --git a/packages/app/src/lib/compare-slug.test.ts b/packages/app/src/lib/compare-slug.test.ts
index f85aac9e..c0c28cf1 100644
--- a/packages/app/src/lib/compare-slug.test.ts
+++ b/packages/app/src/lib/compare-slug.test.ts
@@ -6,6 +6,8 @@ import {
   canonicalCompareSlug,
   compareDisplayLabel,
   compareModelDisplayLabel,
+  compareModelSeoName,
+  compareSeoTitle,
   COMPARE_MODEL_ALIASES,
   COMPARE_MODEL_SLUGS,
   getCompareModelBySlug,
@@ -258,6 +260,69 @@ describe('compareModelDisplayLabel', () => {
   });
 });
 
+describe('compareModelSeoName', () => {
+  // Single, natural, current-version search name per model — no version-group
+  // slashes, no param-count suffix, no fully-qualified checkpoint id.
+  const EXPECTED: Record<string, string> = {
+    'deepseek-v4': 'DeepSeek V4 Pro',
+    'deepseek-r1': 'DeepSeek R1',
+    'kimi-k26': 'Kimi K2.6',
+    'glm-5-1': 'GLM-5',
+    'glm-5-2': 'GLM-5.2',
+    'minimax-m3': 'MiniMax M3',
+    'minimax-m27': 'MiniMax M2.7',
+    'qwen-3-5': 'Qwen3.5',
+    'gptoss-120b': 'gpt-oss-120b',
+    'llama-3-3-70b': 'Llama 3.3 70B',
+  };
+
+  it('returns the expected single-version name for every model', () => {
+    for (const model of COMPARE_MODEL_SLUGS) {
+      expect(compareModelSeoName(model), `seoName for ${model.slug}`).toBe(EXPECTED[model.slug]);
+    }
+  });
+
+  it('covers every model in the registry (no missing entries)', () => {
+    expect(Object.keys(EXPECTED).toSorted()).toEqual(
+      COMPARE_MODEL_SLUGS.map((m) => m.slug).toSorted(),
+    );
+  });
+
+  it('never returns a version-grouped label (no slashes) and is non-empty', () => {
+    for (const model of COMPARE_MODEL_SLUGS) {
+      const name = compareModelSeoName(model);
+      expect(name.length, `non-empty for ${model.slug}`).toBeGreaterThan(0);
+      expect(name.includes('/'), `no slash in ${model.slug}`).toBe(false);
+    }
+  });
+});
+
+describe('compareSeoTitle', () => {
+  it('leads with the GPU pair and keeps "Inference Benchmark" when it fits', () => {
+    const title = compareSeoTitle('B200 vs B300', 'GLM-5');
+    expect(title).toBe('B200 vs B300: GLM-5 Inference Benchmark');
+    expect(title.startsWith('B200 vs B300:')).toBe(true);
+    expect(title.length).toBeLessThanOrEqual(60);
+  });
+
+  it('drops "Inference" → "Benchmark" for a long GPU pair + model name', () => {
+    const title = compareSeoTitle('GB200 NVL72 vs GB300 NVL72', 'DeepSeek V4 Pro');
+    // "…Inference Benchmark" would exceed ~60 chars, so it shortens.
+    expect(title).toBe('GB200 NVL72 vs GB300 NVL72: DeepSeek V4 Pro Benchmark');
+    expect(title.includes('Inference Benchmark')).toBe(false);
+    expect(title.startsWith('GB200 NVL72 vs GB300 NVL72:')).toBe(true);
+  });
+
+  it('always starts with the GPU pair for every real (model × pair)', () => {
+    for (const model of COMPARE_MODEL_SLUGS) {
+      const gpuLabel = compareDisplayLabel('gb200', 'gb300');
+      const title = compareSeoTitle(gpuLabel, compareModelSeoName(model));
+      expect(title.startsWith(`${gpuLabel}:`), `pair-led for ${model.slug}`).toBe(true);
+      expect(/(?<suffix>Inference Benchmark|Benchmark)$/u.test(title)).toBe(true);
+    }
+  });
+});
+
 describe('getCompareModelBySlug', () => {
   it('returns canonical models for canonical slugs', () => {
     expect(getCompareModelBySlug('deepseek-r1')).toBe(DEEPSEEK_R1);
diff --git a/packages/app/src/lib/compare-slug.ts b/packages/app/src/lib/compare-slug.ts
index 10e9ed4f..fb636c40 100644
--- a/packages/app/src/lib/compare-slug.ts
+++ b/packages/app/src/lib/compare-slug.ts
@@ -28,6 +28,13 @@ export interface CompareModelSlug {
   dbKeys: string[];
   /** Human label for OG image, metadata, and page header. */
   label: string;
+  /** Single, natural, current-version model name for SEO `<title>` tags and
+   *  meta descriptions — e.g. `GLM-5` (not the version-grouped `GLM 5/5.1`)
+   *  or `Kimi K2.6` (the primary version the slug represents). Kept short so
+   *  the GPU-pair-led title survives Google's ~60-char SERP truncation, and
+   *  written the way people actually search rather than the fully-qualified
+   *  HF checkpoint id. See `compareModelSeoName`. */
+  seoName: string;
 }
 
 // Order matches the master /compare and /compare-per-dollar index display:
@@ -43,12 +50,14 @@ export const COMPARE_MODEL_SLUGS: CompareModelSlug[] = [
     displayName: 'DeepSeek-V4-Pro',
     dbKeys: ['dsv4'],
     label: 'DeepSeek V4 Pro 1.6T',
+    seoName: 'DeepSeek V4 Pro',
   },
   {
     slug: 'deepseek-r1',
     displayName: 'DeepSeek-R1-0528',
     dbKeys: ['dsr1'],
     label: 'DeepSeek R1',
+    seoName: 'DeepSeek R1',
   },
   {
     slug: 'kimi-k26',
@@ -62,6 +71,8 @@ export const COMPARE_MODEL_SLUGS: CompareModelSlug[] = [
     // surfaces every version so the URL doesn't read as "only K2.6". Param
     // count appended so the label conveys model scale alongside the version.
     label: 'Kimi K2.5/K2.6/K2.7-Code 1T',
+    // SEO name uses the primary version the slug canonicalizes to (K2.6).
+    seoName: 'Kimi K2.6',
   },
   {
     slug: 'glm-5-1',
@@ -69,12 +80,14 @@ export const COMPARE_MODEL_SLUGS: CompareModelSlug[] = [
     // GLM-5 and GLM-5.1 remain grouped under the stable canonical slug.
     dbKeys: ['glm5.1', 'glm5'],
     label: 'GLM 5/5.1',
+    seoName: 'GLM-5',
   },
   {
     slug: 'glm-5-2',
     displayName: 'GLM-5.2',
     dbKeys: ['glm5.2'],
     label: 'GLM 5.2',
+    seoName: 'GLM-5.2',
   },
   {
     slug: 'minimax-m3',
@@ -84,6 +97,7 @@ export const COMPARE_MODEL_SLUGS: CompareModelSlug[] = [
     // joining the minimax-m27 group.
     dbKeys: ['minimaxm3'],
     label: 'MiniMax M3 428B',
+    seoName: 'MiniMax M3',
   },
   {
     slug: 'minimax-m27',
@@ -91,6 +105,8 @@ export const COMPARE_MODEL_SLUGS: CompareModelSlug[] = [
     // Same point-release grouping pattern as Kimi and GLM.
     dbKeys: ['minimaxm2.7', 'minimaxm2.5'],
     label: 'MiniMax M2.5/M2.7',
+    // Primary version the slug canonicalizes to (M2.7).
+    seoName: 'MiniMax M2.7',
   },
   {
     slug: 'qwen-3-5',
@@ -98,18 +114,21 @@ export const COMPARE_MODEL_SLUGS: CompareModelSlug[] = [
     dbKeys: ['qwen3.5'],
     // 397B total parameters, 17B active per forward pass (MoE).
     label: 'Qwen 3.5 397B-A17B',
+    seoName: 'Qwen3.5',
   },
   {
     slug: 'gptoss-120b',
     displayName: 'gpt-oss-120b',
     dbKeys: ['gptoss120b'],
     label: 'gpt-oss 120B',
+    seoName: 'gpt-oss-120b',
   },
   {
     slug: 'llama-3-3-70b',
     displayName: 'Llama-3.3-70B-Instruct-FP8',
     dbKeys: ['llama70b'],
     label: 'Llama 3.3 70B',
+    seoName: 'Llama 3.3 70B',
   },
 ];
 
@@ -272,3 +291,25 @@ export function compareDisplayLabel(a: string, b: string): string {
 export function compareModelDisplayLabel(model: CompareModelSlug, a: string, b: string): string {
   return `${model.label} — ${compareDisplayLabel(a, b)}`;
 }
+
+/** Single natural, current-version model name for SEO titles / descriptions —
+ *  e.g. `GLM-5`, `Kimi K2.6`, `DeepSeek R1`. Unlike `label`, it never contains
+ *  version-group slashes ("GLM 5/5.1"), param-count suffixes ("1.6T"), or the
+ *  fully-qualified checkpoint id — it reads the way people search. */
+export function compareModelSeoName(model: CompareModelSlug): string {
+  return model.seoName;
+}
+
+/** Pre-pipe portion of the SEO `<title>`, GPU-pair first so the actual search
+ *  query ("B200 vs B300") survives Google's ~60-char truncation. Falls back
+ *  from "Inference Benchmark" to "Benchmark" when the longer suffix would push
+ *  the pair-led phrase past ~60 chars for a long GPU pair / model name.
+ *
+ *  Returns the text before the " | InferenceX" site-name suffix; callers add
+ *  the suffix via `title: { absolute: ... }` so the long root template
+ *  ("%s | InferenceX by SemiAnalysis") does not apply and truncate the query. */
+export function compareSeoTitle(gpuLabel: string, modelSeoName: string): string {
+  const full = `${gpuLabel}: ${modelSeoName} Inference Benchmark`;
+  if (full.length <= 60) return full;
+  return `${gpuLabel}: ${modelSeoName} Benchmark`;
+}
diff --git a/packages/app/src/lib/compare-ssr-zh.ts b/packages/app/src/lib/compare-ssr-zh.ts
index cd48611a..149f9fac 100644
--- a/packages/app/src/lib/compare-ssr-zh.ts
+++ b/packages/app/src/lib/compare-ssr-zh.ts
@@ -9,17 +9,25 @@ import {
   AUTHOR_NAME,
   AUTHOR_URL,
   HW_REGISTRY,
+  SITE_NAME,
   SITE_URL,
 } from '@semianalysisai/inferencex-constants';
 
-import { type CompareModelSlug, compareModelDisplayLabel } from '@/lib/compare-slug';
+import {
+  type CompareModelSlug,
+  compareDisplayLabel,
+  compareModelDisplayLabel,
+  compareModelSeoName,
+} from '@/lib/compare-slug';
 import {
   bandFor,
   type CompareJsonLdVariant,
+  computeCompareStat,
   fmtCost,
   fmtPctDelta,
   type FullBoth,
   jsonLdEntryFor,
+  META_DESCRIPTION_MAX,
   type PairSummary,
   type PerDollarBoth,
   pickRotated,
@@ -274,6 +282,67 @@ export function compareTableNarrativeZh(
   return paragraphs;
 }
 
+// ---------------------------------------------------------------------------
+// SEO meta description — Chinese port of compareMetaDescription
+// ---------------------------------------------------------------------------
+
+/** First candidate ≤ max, or undefined. Local mirror of the private helper in
+ *  compare-ssr.ts so the ladder logic stays identical between the two files. */
+function firstUnderZh(candidates: string[], max: number): string | undefined {
+  return candidates.find((c) => c.length <= max);
+}
+
+/** Simplified Chinese, stat-led, ≤155-char meta description for a
+ *  `/zh/compare/<slug>` page. 1:1 port of `compareMetaDescription`: same
+ *  representative-row stat (`computeCompareStat`), same fallback-to-boilerplate
+ *  and brand-clause-drop ladders. Model name, GPU SKUs and units stay English
+ *  per the translation rules; the connective prose is Chinese. */
+export function compareMetaDescriptionZh(
+  model: CompareModelSlug,
+  a: string,
+  b: string,
+  ssrRows: SsrInterpolatedRow[],
+): string {
+  const modelName = compareModelSeoName(model);
+  const gpuLabel = compareDisplayLabel(a, b);
+
+  const fallback =
+    firstUnderZh(
+      [
+        `${gpuLabel} 在 ${modelName} 上的推理基准测试：来自 ${SITE_NAME}（${AUTHOR_NAME} 出品）的经验证、可复现开源结果。对比延迟、吞吐量与成本。`,
+        `${gpuLabel} 在 ${modelName} 上的推理基准测试：来自 ${SITE_NAME}（${AUTHOR_NAME} 出品）的开源结果。`,
+        `${gpuLabel} 在 ${modelName} 上的推理基准测试，来自 ${SITE_NAME}。`,
+        `${gpuLabel} 在 ${modelName} 上的推理基准测试。`,
+      ],
+      META_DESCRIPTION_MAX,
+    ) ?? `${gpuLabel} 推理基准测试`.slice(0, META_DESCRIPTION_MAX);
+
+  const stat = computeCompareStat(a, b, ssrRows);
+  if (!stat) return fallback;
+
+  const tputClause =
+    stat.tputPct > 0 ? `${stat.faster} 每 GPU 吞吐量比 ${stat.slower} 高 ${stat.tputPct}%` : null;
+  const costClause = stat.costPct > 0 ? `${stat.cheaper} 每 token 成本低 ${stat.costPct}%` : null;
+
+  let core: string;
+  if (tputClause && costClause) core = `在 ${modelName} 上，${tputClause}；${costClause}。`;
+  else if (tputClause) core = `在 ${modelName} 上，${tputClause}。`;
+  else if (costClause)
+    core = `在 ${modelName} 上，${stat.cheaper} 每 token 成本比 ${stat.pricier} 低 ${stat.costPct}%。`;
+  else return fallback;
+
+  return (
+    firstUnderZh(
+      [
+        `${core}来自 ${SITE_NAME}（${AUTHOR_NAME} 出品）的可验证开源基准测试。`,
+        `${core}来自 ${SITE_NAME} 的开源基准测试。`,
+        core,
+      ],
+      META_DESCRIPTION_MAX,
+    ) ?? fallback
+  );
+}
+
 // ---------------------------------------------------------------------------
 // JSON-LD — Chinese
 // ---------------------------------------------------------------------------
diff --git a/packages/app/src/lib/compare-ssr.test.ts b/packages/app/src/lib/compare-ssr.test.ts
index 356fc869..be02369c 100644
--- a/packages/app/src/lib/compare-ssr.test.ts
+++ b/packages/app/src/lib/compare-ssr.test.ts
@@ -1,8 +1,17 @@
 import { describe, expect, it } from 'vitest';
 
 import type { BenchmarkRow } from '@/lib/api';
+import type { InterpolatedResult } from '@/components/calculator/types';
 
-import { computeCompareImageRows, KNOWN_MODELS } from './compare-ssr';
+import { COMPARE_MODEL_SLUGS } from './compare-slug';
+import {
+  compareMetaDescription,
+  computeCompareImageRows,
+  KNOWN_MODELS,
+  META_DESCRIPTION_MAX,
+  type SsrInterpolatedRow,
+} from './compare-ssr';
+import { compareMetaDescriptionZh } from './compare-ssr-zh';
 
 // BenchmarkRow.id is required (stable per-point id from benchmark_results);
 // hand out a fresh one per stub so id-keyed logic can't collide across rows.
@@ -141,3 +150,162 @@ describe('computeCompareImageRows', () => {
     ).toEqual([]);
   });
 });
+
+// ---------------------------------------------------------------------------
+// compareMetaDescription (+ zh port)
+// ---------------------------------------------------------------------------
+
+const GLM = COMPARE_MODEL_SLUGS.find((m) => m.slug === 'glm-5-1')!; // seoName 'GLM-5'
+const DSV4 = COMPARE_MODEL_SLUGS.find((m) => m.slug === 'deepseek-v4')!; // long seoName
+
+/** Minimal InterpolatedResult stub — compareMetaDescription only reads
+ *  `value` (tok/s/GPU) and `cost` ($/M tok); the rest are inert zeros. */
+function ir(value: number, cost: number): InterpolatedResult {
+  return {
+    hwKey: 'x',
+    resultKey: 'x',
+    value,
+    outputTputValue: value,
+    inputTputValue: 0,
+    cost,
+    costInput: 0,
+    costOutput: cost,
+    tpPerMw: 0,
+    inputTpPerMw: 0,
+    outputTpPerMw: 0,
+    concurrency: 0,
+    nearestPoints: [],
+  };
+}
+
+function makeSsrRows(
+  triples: [number, InterpolatedResult | null, InterpolatedResult | null][],
+): SsrInterpolatedRow[] {
+  return triples.map(([target, a, b]) => ({ target, a, b }));
+}
+
+describe('compareMetaDescription', () => {
+  it('leads with the throughput + cost stat when both dimensions differ', () => {
+    // a=B200 34% faster, b=B300 12% cheaper, at the middle (default) target.
+    const ssr = makeSsrRows([
+      [20, ir(60, 2), ir(50, 1.7)],
+      [40, ir(134, 1.12), ir(100, 1)], // 134/100 = +34%, 1.12/1.0 = +12%
+      [60, ir(200, 0.9), ir(160, 0.85)],
+    ]);
+    const desc = compareMetaDescription(GLM, 'b200', 'b300', ssr);
+    expect(desc).toBe(
+      'B200 delivers 34% more tok/s/GPU than B300 on GLM-5; B300 is 12% cheaper per token. Verified open-source benchmarks from InferenceX by SemiAnalysis.',
+    );
+    expect(desc.length).toBeLessThanOrEqual(META_DESCRIPTION_MAX);
+    expect(desc.startsWith('B200 delivers 34% more tok/s/GPU than B300 on GLM-5')).toBe(true);
+  });
+
+  it('emits only the throughput clause when cost is within 1% (tied)', () => {
+    const ssr = makeSsrRows([[40, ir(134, 1), ir(100, 1)]]);
+    const desc = compareMetaDescription(GLM, 'b200', 'b300', ssr);
+    expect(desc.includes('34% more tok/s/GPU')).toBe(true);
+    expect(desc.includes('cheaper per token')).toBe(false);
+    expect(desc.length).toBeLessThanOrEqual(META_DESCRIPTION_MAX);
+  });
+
+  it('emits only the cost clause (with "than") when throughput is tied', () => {
+    const ssr = makeSsrRows([[40, ir(100, 1.5), ir(100, 1)]]);
+    const desc = compareMetaDescription(GLM, 'b200', 'b300', ssr);
+    expect(desc).toContain('B300 is 50% cheaper per token than B200 on GLM-5');
+    expect(desc.includes('more tok/s/GPU')).toBe(false);
+    expect(desc.length).toBeLessThanOrEqual(META_DESCRIPTION_MAX);
+  });
+
+  it('falls back to boilerplate when there is no comparable data', () => {
+    const empty = compareMetaDescription(GLM, 'b200', 'b300', []);
+    expect(empty.startsWith('B200 vs B300 inference benchmark on GLM-5')).toBe(true);
+    expect(empty.includes('tok/s/GPU')).toBe(false);
+    expect(empty.length).toBeLessThanOrEqual(META_DESCRIPTION_MAX);
+
+    // One side missing at every target → still boilerplate.
+    const oneSided = compareMetaDescription(
+      GLM,
+      'b200',
+      'b300',
+      makeSsrRows([[40, ir(100, 1), null]]),
+    );
+    expect(oneSided.startsWith('B200 vs B300 inference benchmark on GLM-5')).toBe(true);
+  });
+
+  it('falls back when both dimensions are within 1%', () => {
+    const ssr = makeSsrRows([[40, ir(100, 1), ir(100, 1)]]);
+    const desc = compareMetaDescription(GLM, 'b200', 'b300', ssr);
+    expect(desc.startsWith('B200 vs B300 inference benchmark on GLM-5')).toBe(true);
+  });
+
+  it('prefers the middle (default) target row', () => {
+    // Middle row has the data; outer rows are degenerate.
+    const ssr = makeSsrRows([
+      [20, null, null],
+      [40, ir(134, 1.12), ir(100, 1)],
+      [60, null, null],
+    ]);
+    const desc = compareMetaDescription(GLM, 'b200', 'b300', ssr);
+    expect(desc.includes('34% more tok/s/GPU')).toBe(true);
+  });
+
+  it('falls back to an outer usable row when the middle target is degenerate', () => {
+    const ssr = makeSsrRows([
+      [20, ir(134, 1.12), ir(100, 1)],
+      [40, null, null],
+      [60, null, null],
+    ]);
+    const desc = compareMetaDescription(GLM, 'b200', 'b300', ssr);
+    expect(desc.includes('34% more tok/s/GPU')).toBe(true);
+  });
+
+  it('stays ≤155 chars even with long GPU labels, model name, and huge ratios', () => {
+    const ssr = makeSsrRows([[40, ir(9999, 99.9), ir(10, 0.5)]]);
+    const desc = compareMetaDescription(DSV4, 'gb200', 'gb300', ssr);
+    expect(desc.length).toBeLessThanOrEqual(META_DESCRIPTION_MAX);
+    // A differentiating stat is still present (not the boilerplate).
+    expect(desc.includes('tok/s/GPU') || desc.includes('cheaper per token')).toBe(true);
+  });
+});
+
+describe('compareMetaDescriptionZh — structural 1:1 port', () => {
+  const ssr = makeSsrRows([
+    [20, ir(60, 2), ir(50, 1.7)],
+    [40, ir(134, 1.12), ir(100, 1)],
+    [60, ir(200, 0.9), ir(160, 0.85)],
+  ]);
+
+  it('surfaces the same numbers, model name, and GPU labels as the English version', () => {
+    const zh = compareMetaDescriptionZh(GLM, 'b200', 'b300', ssr);
+    expect(zh).toContain('GLM-5');
+    expect(zh).toContain('B200');
+    expect(zh).toContain('B300');
+    expect(zh).toContain('34%');
+    expect(zh).toContain('12%');
+    expect(zh).toContain('每 GPU 吞吐量');
+    expect(zh).toContain('每 token 成本');
+    expect(zh.length).toBeLessThanOrEqual(META_DESCRIPTION_MAX);
+  });
+
+  it('is stat-led exactly when the English version is (and boilerplate otherwise)', () => {
+    // With data: both stat-led.
+    const en = compareMetaDescription(GLM, 'b200', 'b300', ssr);
+    const zh = compareMetaDescriptionZh(GLM, 'b200', 'b300', ssr);
+    expect(en.includes('34%')).toBe(true);
+    expect(zh.includes('34%')).toBe(true);
+
+    // Without data: both boilerplate.
+    const enFb = compareMetaDescription(GLM, 'b200', 'b300', []);
+    const zhFb = compareMetaDescriptionZh(GLM, 'b200', 'b300', []);
+    expect(enFb.includes('inference benchmark')).toBe(true);
+    expect(zhFb.includes('推理基准测试')).toBe(true);
+    expect(zhFb.length).toBeLessThanOrEqual(META_DESCRIPTION_MAX);
+  });
+
+  it('stays ≤155 chars with long labels + huge ratios', () => {
+    const big = makeSsrRows([[40, ir(9999, 99.9), ir(10, 0.5)]]);
+    expect(compareMetaDescriptionZh(DSV4, 'gb200', 'gb300', big).length).toBeLessThanOrEqual(
+      META_DESCRIPTION_MAX,
+    );
+  });
+});
diff --git a/packages/app/src/lib/compare-ssr.ts b/packages/app/src/lib/compare-ssr.ts
index 44e80939..592a8238 100644
--- a/packages/app/src/lib/compare-ssr.ts
+++ b/packages/app/src/lib/compare-ssr.ts
@@ -14,6 +14,7 @@ import {
   AUTHOR_NAME,
   AUTHOR_URL,
   HW_REGISTRY,
+  SITE_NAME,
   SITE_URL,
   sequenceToIslOsl,
 } from '@semianalysisai/inferencex-constants';
@@ -35,6 +36,7 @@ import {
   type ComparePair,
   type CompareModelSlug,
   compareModelDisplayLabel,
+  compareModelSeoName,
 } from '@/lib/compare-slug';
 import { getHardwareConfig, getGpuSpecs } from '@/lib/constants';
 import { loadFixture } from '@/lib/test-fixtures';
@@ -726,6 +728,150 @@ export function compareTableNarrative(
   return paragraphs;
 }
 
+// ---------------------------------------------------------------------------
+// SEO meta description — stat-led, ≤155 chars, per (model, GPU pair)
+// ---------------------------------------------------------------------------
+
+/** The single interpolated head-to-head stat surfaced in the meta description.
+ *  `tputPct` / `costPct` are 0 when that dimension is within ~1% (a tie). */
+export interface CompareStat {
+  aLabel: string;
+  bLabel: string;
+  gpuLabel: string;
+  faster: string;
+  slower: string;
+  /** Throughput-per-GPU advantage of `faster` over `slower`, as a whole
+   *  percent (0 when the two are within 1%). */
+  tputPct: number;
+  cheaper: string;
+  pricier: string;
+  /** Cost-per-token advantage of `cheaper` over `pricier`, as a whole percent
+   *  (0 when the two are within 1%). */
+  costPct: number;
+}
+
+/** Does a row carry a real, comparable data point for both GPUs? */
+function statRowUsable(row: SsrInterpolatedRow): boolean {
+  return Boolean(
+    row.a && row.b && row.a.value > 0 && row.b.value > 0 && row.a.cost > 0 && row.b.cost > 0,
+  );
+}
+
+/** Pick the representative interpolated row for the meta description — the
+ *  default (middle) interactivity target if it has data for both GPUs, else
+ *  the first target that does. Returns null when no target has a comparable
+ *  pair (sparse-data pairs fall back to boilerplate). */
+function pickCompareStatRow(rows: SsrInterpolatedRow[]): SsrInterpolatedRow | null {
+  const mid = rows[Math.floor(rows.length / 2)];
+  if (mid && statRowUsable(mid)) return mid;
+  return rows.find(statRowUsable) ?? null;
+}
+
+/** Language-agnostic head-to-head stat for the meta description, computed at
+ *  the default interactivity target. Shared by the English `compareMetaDescription`
+ *  and its Chinese port so both surface the same numbers. Returns null when the
+ *  pair has no comparable interpolated data. */
+export function computeCompareStat(
+  a: string,
+  b: string,
+  ssrRows: SsrInterpolatedRow[],
+): CompareStat | null {
+  const row = pickCompareStatRow(ssrRows);
+  if (!row || !row.a || !row.b) return null;
+
+  const aLabel = HW_REGISTRY[a]?.label ?? a.toUpperCase();
+  const bLabel = HW_REGISTRY[b]?.label ?? b.toUpperCase();
+  const rA = row.a;
+  const rB = row.b;
+
+  const aFaster = rA.value > rB.value;
+  const tputRatio = aFaster ? rA.value / rB.value : rB.value / rA.value;
+  const tputPct = Math.round((tputRatio - 1) * 100);
+
+  const aCheaper = rA.cost < rB.cost;
+  const costRatio = aCheaper ? rB.cost / rA.cost : rA.cost / rB.cost;
+  const costPct = Math.round((costRatio - 1) * 100);
+
+  return {
+    aLabel,
+    bLabel,
+    gpuLabel: compareDisplayLabel(a, b),
+    faster: aFaster ? aLabel : bLabel,
+    slower: aFaster ? bLabel : aLabel,
+    tputPct: tputPct >= 1 ? tputPct : 0,
+    cheaper: aCheaper ? aLabel : bLabel,
+    pricier: aCheaper ? bLabel : aLabel,
+    costPct: costPct >= 1 ? costPct : 0,
+  };
+}
+
+/** Length cap for meta descriptions — Google truncates SERP snippets past
+ *  ~155–160 chars, so every branch below must stay at or under this. */
+export const META_DESCRIPTION_MAX = 155;
+
+/** First candidate that fits within `max`, or undefined if none do. */
+function firstUnder(candidates: string[], max: number): string | undefined {
+  return candidates.find((c) => c.length <= max);
+}
+
+/** Stat-led, ≤155-char meta description for a `/compare/<slug>` page. Leads
+ *  with the differentiating throughput / cost delta at the default interactivity
+ *  target (the actual reason a searcher clicks) instead of the boilerplate that
+ *  was identical across ~360 pages. Falls back to a compact, still-≤155-char
+ *  boilerplate sentence when the pair has thin / degenerate data.
+ *
+ *  A trailing brand clause is dropped down a ladder (full → short → none) so a
+ *  long GPU-pair / model name never pushes the description over the cap. */
+export function compareMetaDescription(
+  model: CompareModelSlug,
+  a: string,
+  b: string,
+  ssrRows: SsrInterpolatedRow[],
+): string {
+  const modelName = compareModelSeoName(model);
+  const gpuLabel = compareDisplayLabel(a, b);
+
+  // Boilerplate fallback ladder — progressively shorter so one always fits.
+  const fallback =
+    firstUnder(
+      [
+        `${gpuLabel} inference benchmark on ${modelName}: verified, reproducible open-source results from ${SITE_NAME} by ${AUTHOR_NAME}. Latency, throughput & cost.`,
+        `${gpuLabel} inference benchmark on ${modelName}: verified open-source results from ${SITE_NAME} by ${AUTHOR_NAME}.`,
+        `${gpuLabel} inference benchmark on ${modelName} from ${SITE_NAME} by ${AUTHOR_NAME}.`,
+        `${gpuLabel} inference benchmark on ${modelName}.`,
+      ],
+      META_DESCRIPTION_MAX,
+    ) ?? `${gpuLabel} inference benchmark`.slice(0, META_DESCRIPTION_MAX);
+
+  const stat = computeCompareStat(a, b, ssrRows);
+  if (!stat) return fallback;
+
+  const tputClause =
+    stat.tputPct > 0
+      ? `${stat.faster} delivers ${stat.tputPct}% more tok/s/GPU than ${stat.slower} on ${modelName}`
+      : null;
+  const costClause =
+    stat.costPct > 0 ? `${stat.cheaper} is ${stat.costPct}% cheaper per token` : null;
+
+  let core: string;
+  if (tputClause && costClause) core = `${tputClause}; ${costClause}.`;
+  else if (tputClause) core = `${tputClause}.`;
+  else if (costClause)
+    core = `${stat.cheaper} is ${stat.costPct}% cheaper per token than ${stat.pricier} on ${modelName}.`;
+  else return fallback; // both dimensions within 1% — nothing differentiating to lead with
+
+  return (
+    firstUnder(
+      [
+        `${core} Verified open-source benchmarks from ${SITE_NAME} by ${AUTHOR_NAME}.`,
+        `${core} Verified open-source ${SITE_NAME} benchmarks.`,
+        core,
+      ],
+      META_DESCRIPTION_MAX,
+    ) ?? fallback
+  );
+}
+
 // ---------------------------------------------------------------------------
 // Master-index helpers (shared by /compare and /compare-per-dollar)
 // ---------------------------------------------------------------------------
diff --git a/packages/app/src/lib/compare-variant-ssr.test.ts b/packages/app/src/lib/compare-variant-ssr.test.ts
index 0e85c225..857e8ef0 100644
--- a/packages/app/src/lib/compare-variant-ssr.test.ts
+++ b/packages/app/src/lib/compare-variant-ssr.test.ts
@@ -159,6 +159,7 @@ const MODEL_SLUG = {
   displayName: 'DeepSeek-R1-0528',
   dbKeys: ['dsr1'],
   label: 'DeepSeek R1',
+  seoName: 'DeepSeek R1',
 };
 
 // ---------------------------------------------------------------------------

From d3116d73233857e3f2d8187b050af4f87815999f Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Mon, 20 Jul 2026 17:04:39 +0800
Subject: [PATCH 2/2] fix(seo): sync blog JSON-LD description with meta/OG
 description
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Cursor Bugbot flagged that the BlogPosting JSON-LD description still used
the full meta.subtitle while the HTML meta, OpenGraph, and Twitter
descriptions were routed through blogDescription — so posts with a
seoDescription or long subtitle exposed different text in structured data
than in the SERP/social tags. Route the JSON-LD description through the
same blogDescription helper (EN + zh) so they stay in sync. Logic is
already covered by the blogDescription unit tests.

中文：修复博客 JSON-LD 描述与 meta/OG 描述不一致的问题。Cursor Bugbot 指出
BlogPosting JSON-LD 的 description 仍使用完整的 meta.subtitle，而 HTML meta、
OpenGraph 和 Twitter 描述已改用 blogDescription，导致设置了 seoDescription 或
副标题较长的文章在结构化数据与搜索/社交标签中显示不同文本。这里让 JSON-LD 的
description 同样走 blogDescription（EN 与 zh 两侧），保持一致；相关逻辑已由
blogDescription 单元测试覆盖。
---
 packages/app/src/app/blog/[slug]/page.tsx    | 5 ++++-
 packages/app/src/app/zh/blog/[slug]/page.tsx | 5 ++++-
 2 files changed, 8 insertions(+), 2 deletions(-)

diff --git a/packages/app/src/app/blog/[slug]/page.tsx b/packages/app/src/app/blog/[slug]/page.tsx
index e21b5b12..eaad761c 100644
--- a/packages/app/src/app/blog/[slug]/page.tsx
+++ b/packages/app/src/app/blog/[slug]/page.tsx
@@ -144,7 +144,10 @@ export default async function BlogPostPage({ params }: Props) {
     publisher: { '@type': 'Organization', name: AUTHOR_NAME },
     datePublished: `${meta.date}T00:00:00Z`,
     ...(meta.modifiedDate && { dateModified: `${meta.modifiedDate}T00:00:00Z` }),
-    description: meta.subtitle,
+    // Keep structured-data description in sync with the SERP/OG/Twitter meta
+    // (both go through blogDescription) so they never diverge for posts with a
+    // seoDescription or a long subtitle.
+    description: blogDescription(meta),
     url: `${SITE_URL}/blog/${slug}`,
     wordCount: raw.trim().split(/\s+/u).length,
     timeRequired: `PT${meta.readingTime}M`,
diff --git a/packages/app/src/app/zh/blog/[slug]/page.tsx b/packages/app/src/app/zh/blog/[slug]/page.tsx
index 391b90a3..4d251d8b 100644
--- a/packages/app/src/app/zh/blog/[slug]/page.tsx
+++ b/packages/app/src/app/zh/blog/[slug]/page.tsx
@@ -140,7 +140,10 @@ export default async function ZhBlogPostPage({ params }: Props) {
     publisher: { '@type': 'Organization', name: AUTHOR_NAME },
     datePublished: `${meta.date}T00:00:00Z`,
     ...(meta.modifiedDate && { dateModified: `${meta.modifiedDate}T00:00:00Z` }),
-    description: meta.subtitle,
+    // Keep structured-data description in sync with the SERP/OG/Twitter meta
+    // (both go through blogDescription) so they never diverge for posts with a
+    // seoDescription or a long subtitle.
+    description: blogDescription(meta),
     url: `${SITE_URL}/zh/blog/${slug}`,
     inLanguage: ZH_LANG_TAG,
     wordCount: raw.trim().split(/\s+/u).length,