These are my working notes on getting indexed — and cited — by both classic search engines and the newer AI answer engines. I write a lot of bilingual, fairly niche technical content (the TokenChef series, wtcraft), and the honest reality is that nobody types my project names into Google yet. So the question I actually care about isn’t “how do I rank,” it’s “when someone asks Perplexity or ChatGPT about git-native multi-agent tooling, does my work show up in the answer — with a link back to me?”
That framing — optimizing for the citation, not the blue link — is what GEO (Generative Engine Optimization) is about, and it’s where I spend most of my effort. SEO still matters, but it’s largely a solved, commodity checklist. GEO is the part that’s still soft and worth thinking about.
摘要：这一节是总览——为什么做、怎么做、什么时候做，以及最关键的:GEO 怎么验收。下面三篇是分场景的实战(开源包 / 个人博客 / 商业站重构)。SEO 是老生常谈,一笔带过;重点在 GEO。
Why — SEO is table stakes, GEO is the new surface
Two different machines read your site, and they fail in different ways.
A search crawler wants clean HTML, a sitemap, fast paint, and unique titles. If you give it that, you rank roughly as well as your backlinks deserve. This is well-trodden ground; there’s nothing clever left to do here, only things you can forget to do.
An answer engine (Perplexity, ChatGPT Search, Google AI Overviews, Copilot) does something else entirely. It runs a RAG (检索增强生成) loop: retrieve a handful of pages, slice them into chunks, and synthesize an answer that quotes a few sources. You’re no longer competing for position #1 — you’re competing to be one of the 3–5 chunks the model decides are worth quoting. That’s a content-shape problem, not a backlink problem, and almost nobody optimizes for it yet. That gap is the opportunity.
为什么重点在 GEO:传统 SEO(关键词 / Domain Authority / 外链)是红海,人人都在做。GEO 还很「软」,内容结构稍微为 LLM 提炼习惯调整一下,就能被引用——这是性价比最高的地方。
How — the whole playbook on one screen
Everything in the three practical docs reduces to the same handful of moves:
- Let the AI crawlers in. A
robots.txtthat explicitly allowsGPTBot,OAI-SearchBot,Claude-Web,Google-Extended. Most people reflexively block these; if you want to be cited, do the opposite. - Lead with the answer (BLUF — Bottom Line Up Front). The first 1–2 sentences under any heading should state the conclusion in plain language. LLMs lift that line almost verbatim.
- Make facts liftable. Comparison tables, FAQ sections with real question-shaped
H3s, concrete numbers with authoritative outbound links. Models love a table they can render straight into an answer card. - Be a clean knowledge graph. JSON-LD (
Person,SoftwareApplication,Article) so the model can resolve who/what without guessing. - Close the backlink loop. Point your high-authority surfaces (GitHub, NPM, PyPI, Dev.to) back at the canonical page, with
rel=canonicalset correctly so the equity settles on your domain.
The mechanics differ per surface — a README is not a blog post is not a marketing landing page — which is exactly why the three docs below exist.
When — sequencing matters more than you’d think
摘要:别一上来就追 GEO。先让爬虫拿到干净 HTML,再谈被 AI 引用。
A rough timeline, because doing these out of order wastes weeks:
- At launch: ship static HTML, valid sitemap, per-page
<title>/description/canonical, JSON-LD, and the permissiverobots.txt. This is the SEO baseline — without it, there’s nothing for an answer engine to retrieve. - Day 0–7: seed the high-authority backlinks (NPM/PyPI
homepage, README footer link, a Dev.to cross-post with canonical pointing home). This is what gets you discovered fast. - After ~1 week: start the GEO acceptance loop below. Indexing isn’t instant; testing on day one just tells you the crawler hasn’t been around yet.
- Ongoing: reshape content based on what the models actually quote (or get wrong).
How to verify GEO — the acceptance test (验收)
This is the part people skip, and it’s the only part that tells you whether any of the above worked. SEO has dashboards; GEO mostly doesn’t, so you verify it by interrogating the models directly and scoring the answers.
摘要:GEO 没有现成后台,验收靠「向 AI 提问 + 给回答打分」。三个核心指标:引用率、域名提及、信息准确度。优先在 Perplexity 上测。
Run a fixed prompt set across Perplexity, ChatGPT Search, Gemini, and Copilot — the same prompts each week so the trend is comparable:
| Intent | Example prompt | What “pass” looks like |
|---|---|---|
| Identity / 履历 | ”Who is Victor Zhang (zywkloo) and what does he work on?” | Accurate role + projects, with a link to my site |
| Project lookup | ”What is wtcraft and how does it do bounded multi-agent coding?” | Correct mechanism, cites a zywkloo.github.io/blog/... page |
| Category query | ”Lightweight, git-native tooling for AI multi-agent coding?” | wtcraft appears in the list, ideally with the comparison table lifted |
Score three things:
- Citation Rate（引用率） — out of ~10 strongly-relevant prompts, how often does my page show up as a cited source? This is the headline metric.
- Domain Mentions（域名提及） — how often my domain surfaces across models, cited or not.
- Information Accuracy（信息准确度） — is what the model says about me/the project correct, or is it hallucinating a背景 I don’t have? A wrong citation is worse than no citation.
Why Perplexity first: it re-indexes tech blogs and GitHub Pages aggressively (often within hours), it surfaces a “Sources” panel that makes pass/fail obvious at a glance, and its audience skews toward exactly the developers who’d search for a tool like wtcraft. It’s the fastest, clearest feedback loop — I treat it as the canary before checking the slower engines.
SEO, briefly (the commodity part)
I won’t belabor this — it’s a checklist, not an art:
- Static render (SSG), never ship a CSR shell to a crawler.
- Unique
<title>(50–60 chars) anddescription(120–150 chars) per page; titles must be plain text (an<em>tag leaking into<title>is a real, self-inflicted ranking penalty — see the blog doc). rel=canonicaleverywhere to neutralize mirror domains.- Submit the sitemap to Google Search Console; keep Lighthouse SEO/Performance ≥ 95, LCP < 2.5s, CLS < 0.1.
If you’re on Astro this is mostly free out of the box. Get it right once and forget it.
GEO, in depth (the part worth your time)
Everything that makes a page quotable rather than merely rankable:
- RAG-shaped content — short, self-contained chunks that answer one question each, so a retriever can pull a clean slice without surrounding context.
- BLUF under every
H2— 50–80 字 conclusion first, elaboration after. - Question-shaped headings —
### Why does wtcraft use git worktrees?matches how people actually phrase prompts; academic headings (“引子:从一次源码泄露说起”) don’t. - Tables and FAQs — the two formats models reach for when answering “compare X vs Y” or “how do I…”.
- Authority signals — exact numbers, benchmarks, and outbound links to canonical specs (Git docs, Apple Technical Notes). Models prefer sources that look sourced.
- Structured data — JSON-LD so identity and relationships are explicit, not inferred.
The three docs below apply all of this to a specific surface, because the details genuinely differ:
- 开源包 (NPM / PyPI / GitHub) marketing — turning registry domain authority into citations.
- 个人 GitHub Pages + Astro 博客 marketing — the bilingual-blog playbook, plus the technical SEO gotchas specific to Astro.
- 两种 legacy 架构商业站 marketing — a forward plan for two common archetypes: a client-rendered product landing page and a legacy-WordPress B2B site.