feat(seo): lead compare/blog titles with search phrase + stat-led meta descriptions / compare 与博客标题以搜索短语开头并优化 meta 描述

[Oseltamivir]

Summary

SEO click-through improvements for pages that already rank on Google page 1 but lose the click at the SERP. Grounded in Google Search Console data:

• /compare/deepseek-v4-b200-vs-b300 ranks position 7–9 for "b200 vs b300" with 0 clicks — the title buries the query behind a version-grouped model label and em-dash, and truncates after the long site-name suffix.
• Blog meta descriptions come straight from subtitle and run 200–336 chars, so Google truncates them mid-sentence.

Changes

Compare <title> leads with the GPU pair (the actual query).

• DeepSeek V4 Pro 1.6T — B200 vs B300 Inference Benchmark | InferenceX by SemiAnalysis (82c) → B200 vs B300: DeepSeek V4 Pro Inference Benchmark | InferenceX (61c).
• New compareModelSeoName (single natural version name, e.g. GLM-5 not GLM 5/5.1) + compareSeoTitle in compare-slug.ts. Uses title.absolute to bypass the long root template; shortens Inference Benchmark→Benchmark for long pairs.

Compare meta description is stat-led, not identical boilerplate across ~360 pages.

• Built from the interpolated head-to-head numbers at the slug's default operating point (tok/s/GPU + $/M-tok), ≤155 chars, e.g. "B200 delivers 34% more tok/s/GPU than B300 on GLM-5; B300 is 12% cheaper per token…". Falls back to compact boilerplate for sparse-data pairs. New compareMetaDescription / computeCompareStat in compare-ssr.ts.
• ⚠️ Behavior note: compare generateMetadata now performs a benchmark fetch to build this description. It is blob-cached and shared with the page render, and uses the slug's canonical operating point (pickPairDefaults), so it stays deterministic and adds no separate data round-trip on the hot path.

Blog <title> uses the shorter | InferenceX suffix (title.absolute), leaving more room for the headline before SERP truncation.

Blog meta/OG/Twitter descriptions stop truncating.

• New optional seoDescription frontmatter + blogDescription = seoDescription ?? smartTruncate(subtitle, 155) (word-boundary cut, no mid-word). Added explicit seoDescription to 9 posts whose subtitle exceeded ~155 chars.

/zh mirrored — compareMetaDescriptionZh (1:1 port in compare-ssr-zh.ts), zh titles, and zh seoDescription on the 9 translated posts.

Testing

• pnpm typecheck ✅ · pnpm lint ✅ · pnpm fmt ✅
• pnpm test:unit: +24 new passing tests (compare-slug, compare-ssr, blog), 0 regressions. (13 pre-existing, unrelated failures on a clean tree in use-feature-gate/visit-tracking/chunk-load-recovery — a localStorage test-isolation issue, not touched here.)

Overlay support

Not applicable — this PR changes only page <title>/meta descriptions, not any inference/evaluation chart-data rendering path.

中文说明

针对已经排在 Google 搜索结果第一页、却在结果页流失点击的页面做 SEO 点击率优化，依据 Google Search Console 数据：/compare/deepseek-v4-b200-vs-b300 在 “b200 vs b300” 查询上排在第 7–9 位却零点击（标题把查询词埋在版本分组标签和破折号后面，且被过长的站点名后缀截断）；博客的 meta 描述直接取自 subtitle，长达 200–336 字符，被 Google 从句中截断。

• compare 的 <title> 改为以 GPU 对（真实搜索词）开头：例如上面的例子改为 B200 vs B300: DeepSeek V4 Pro Inference Benchmark | InferenceX（61 字符）。新增 compareModelSeoName（单一自然版本名，如 GLM-5 而非 GLM 5/5.1）与 compareSeoTitle，用 title.absolute 绕过冗长的根模板，长组合时把 Inference Benchmark 缩短为 Benchmark。
• compare 的 meta 描述改为数据驱动：在该 slug 默认工作点下由插值出的正面对比数据（tok/s/GPU 与 $/M tok）生成，控制在 155 字符内，数据稀疏时回退到精简样板文案。⚠️ 说明：compare 的 generateMetadata 现在会拉取一次基准测试数据来生成描述，该请求经 blob 缓存并与页面渲染共享，且使用 slug 的规范工作点（pickPairDefaults），保持确定性、不在热路径上新增数据往返。
• 博客 <title> 改用更短的 | InferenceX 后缀，为标题留出更多空间。
• 博客 meta/OG/Twitter 描述不再被截断：新增可选 seoDescription frontmatter，配合 blogDescription = seoDescription ?? smartTruncate(subtitle, 155)（按词边界截断），并为 9 篇超长副标题文章补充 seoDescription。
• /zh 侧同步：compareMetaDescriptionZh（compare-ssr-zh.ts 中 1:1 移植）、中文标题，以及 9 篇译文的中文 seoDescription。

测试：pnpm typecheck、pnpm lint、pnpm fmt 均通过；pnpm test:unit 新增 24 个通过用例，无回归（13 个既有失败与本 PR 无关）。叠加层（overlay）支持：不适用，本 PR 仅修改页面标题与 meta 描述，不涉及推理/评估图表数据渲染路径。

---

Note

Low Risk
Metadata and content frontmatter only; compare metadata adds a cached benchmark read in generateMetadata but reuses existing SSR fetch paths with no auth or data-model changes.

Overview
Improves SERP click-through by retuning blog and compare metadata: shorter titles, bounded descriptions, and compare pages that lead with the GPU query and real benchmark deltas.

Compare pages (/compare and /zh/compare) now build <title> as {GPU pair}: {model seoName} Inference Benchmark | InferenceX via compareSeoTitle and title.absolute, so queries like “b200 vs b300” aren’t buried behind long model labels or the root “by SemiAnalysis” template. Each model gains a seoName (e.g. GLM-5, Kimi K2.6) in compare-slug.ts. Meta descriptions are generated by compareMetaDescription / compareMetaDescriptionZh: tok/s/GPU and $/token percentages from interpolated data at the slug’s default operating point (≤155 chars), with a shorter boilerplate fallback when data is thin. generateMetadata on compare now awaits cached benchmark data (getCachedBenchmarks + pickPairDefaults)—same cache as the page body.

Blog posts (EN + ZH) use title.absolute with | InferenceX, and blogDescription (seoDescription frontmatter or smartTruncate(subtitle, 155)) for meta, OG, Twitter, and JSON-LD. Nine EN + nine ZH posts add explicit seoDescription where subtitles were too long.

Unit tests cover truncation, blog descriptions, SEO title helpers, and compare meta description logic.

^{Reviewed by Cursor Bugbot for commit d3116d7. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview, Comment	Jul 20, 2026 9:05am

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 601e448. Configure here.}