Official website GEO transformation checklist: six layers of information points and execution order
Chapter 6: Practical Advanced Section · Question 103
There is only one test to determine whether a piece of official website copy can be cited by AI:Extract it on its own, paste it into someone else's answer, and see whether it still holds.
If it holds, it is a citable unit. If it does not, it is just marketing copy. The model does not cite your page as a whole; it extracts sentences. So the transformation target is not the page, but the paragraphs and fields.
The following is explained in three parts: "what to do, in what order, and how to verify." All directly copyable tools are collected in the appendix at the end, which can be copied as a standalone checklist.
How exactly to do this test
Use this site as a direct demonstration. Open songbing.com.cn/100wen/, and compare while reading; no tools need to be prepared.
Step one: write three questions. Write them the way users would ask, and do not rewrite them into industry terminology.
- What does this site do?
- Who is it suitable for?
- Compared with other SEO tutorials, what is different?
Step two: copy sentences from the page. Find the sentences that already answer these three questions, copy them exactly as they are, and do not change a single word. Six can be copied from this homepage:
- Eyebrow and title: Chinese Organic Traffic Playbook · 100 Questions on Chinese Organic Marketing
- Subtitle: A practical guide to Chinese organic traffic for marketers, founders, and PR professionals
- Description line: 800-1500 words per question, each question explains one knowledge point thoroughly, readable independently
- Data line: 7 major sections, 103 Q&As, 800+ words each, continuously updated weekly
- Selling point: The first Chinese book to systematically cover GEO/AEO, teaching you how to get recommended by AI
- Selling point: Full-platform coverage, Baidu + Douyin + Xiaohongshu + Zhihu + WeChat Search
Step three: read the question and the sentence together, and see whether the reader would ask a follow-up question.
| User asks | The original sentence on the page | The reader's first reaction |
|---|---|---|
| What does this site do | Chinese Organic Traffic Playbook · 100 Questions on Chinese Organic Marketing | Half understood. Knows it's a playbook, but "organic marketing" still requires another layer of thought |
| What does this site do | A practical guide to organic traffic in Chinese for marketers, founders, and PR professionals | Understood. Pass |
| Who is it for | Each question runs 800-1500 words, fully explaining one concept per question and readable on its own | Knows how to read it, but not who it's written for. Doesn't answer the question |
| Is it worth reading | 7 major sections, 103 Q&As, 800+ words each, continuously updated weekly | Knows the scale. Pass |
| Compared with other tutorials | The first Chinese book to systematically cover GEO/AEO, teaching you how to get recommended by AI | Half-convinced. "The first book of its kind" can't be verified |
| Compared with other tutorials | Baidu + Douyin + Xiaohongshu + Zhihu + WeChat Search | Knows the coverage. Pass. |
Four of the six sentences are usable. This is the real state of the official website: not all correct, not all useless, but stuck on two or three points.
The two stuck sentences don't need the whole paragraph rewritten—just swap out one phrasing:
Original: The first Chinese book to systematically cover GEO/AEO
Problem: "The first of its kind" can't be verified, AI can't paraphrase it, and readers can't check it either
Rewrite: Chapter 3 devotes 15 Q&As to AI search, from citation mechanisms to llms.txt configuration, all with copyable templates"Organic marketing" is the same. The term makes sense within the industry, but for first-time readers it's a new word—one explanatory sentence in front is enough:
Original: 100 Questions on Chinese Organic Marketing
Rewrite: 100 Questions on Chinese Organic Marketing — no ad spend, just content and search for organic trafficThere's only one criterion:Can readers use this sentence to judge for themselves? After fixing these two spots, all six sentences stand on their own. The total change is two short sentences, not a full redesign.
One last result to state upfront: if all three questions are on the pagenot a single one can be copied, don't test further. This page lacks answer units—just fill them in according to the six-layer checklist below.
1. Six layers of information points
Answering a question requires four types of raw material: who you are, what you can do, why you should be trusted, and how you differ from others. If the official website lacks any one type, questions of that type have nothing to do with you.
| Tier | What it solves | What happens if missing |
|---|---|---|
| 1. Entity identity | Confirm which company you are | Confused with a brand of the same name, or simply not mentioned |
| 2 Capability boundaries | Determine whether you are the answer | Can only be mentioned generically, cannot enter recommendation slots |
| 3 Authoritative evidence | Determine whether they dare to use you | Content is trusted but marked as unreliable |
| 4 Differentiation comparison | Make selection comparisons | There is no place for you in selection questions |
| 5 Conversion path | Land on one action | Mentioned but generates no leads |
| 6 Machine-readable | Let the first five layers be read | Content is equivalent to nonexistent for AI |
The first three layers are the prerequisite; the last three layers are the amplifier.
Layer 1: Entity identity
Information to supplement: the correspondence among the company's full name, brand abbreviation, English name, and former name; founding time, place of registration, unified social credit code; a one-sentence definition of the main business; the affiliation between brand name and product name; the single version of name, address, and phone; external profile links pointing to business registration, encyclopedia, and official account.
Two requirements are easy to miss. First,All use plain text, don't make it into images or charts. Second, the wording in all four places must be identical word for word—if the business registration info, encyclopedia profile, WeChat official account bio, and official website footer are written differently, the entity's confidence score will be lowered.
This layer is the one most often skipped, on the grounds that "who even looks at this business registration info."AI does.
Second layer: capability boundaries
Write a standard answer paragraph for each business offering, containing six items: definition sentence, applicable audience, non-applicable situations, deliverables, timeline, and pricing method. Then add service areas and eligibility conditions, restrictions and prerequisites.
"Not applicable" is the most counterintuitive and also the most valuable. When answering selection questions, models usually need to give both reasons for recommending and reasons for excluding; if the whole site has no exclusion sentences, you are effectively excluding yourself from the candidates. Being willing to state clearly "under what circumstances we are not the optimal solution" actually increases the probability of being recommended.
The other two most commonly skipped items are pricing and restriction conditions. It's fine not to disclose specific prices, but you need to give AI something to say; otherwise it can only answer "needs consultation."
Third layer: authoritative evidence
What to add: author bylines and author pages, qualification certificate numbers and issuing agencies, patent numbers, test report numbers, publicly disclosable data from customer cases, data sources, page update times, editorial and fact-checking policies, and disclaimers.
This layer determines whether AI dares to use your content. All numbers must come with thesource, definition, and timetrio.
Unqualified: conversion rate improved significantly.
Qualified: trial conversion rate increased from 2.1% to 3.4% (definition: registration to payment, 14-day window, source: internal tracking, March 2026).As for disclaimers, almost no domestic sites do this. Doing it is differentiation.
Fourth layer: differentiated comparison
What to add: comparison tables, capability matrices, and a clear statement of "what we don't do."
Write the comparison targets as solution types rather than naming competitors, which both avoids legal risk and covers users' real comparison context. Comparisons must state the costs; comparisons that only talk about advantages will be judged as marketing content and won't make it into answers.
Fifth layer: conversion path
Converge, not expand: give only one primary CTA per page; keep contact information consistent across all channels and provide copyable plain text; state response commitments clearly (how long until reply, who replies); and complete after-sales and guarantee terms.
Making contact information only into icons or QR codes means AI gets no information it can relay.
Layer 6: Machine-readable
Execute by priority: server-side rendering or static generation; semantic HTML (one H1 per page, H2s as questions, body content within main and article); structured data; robots.txt distinguishing AI crawler purposes; llms.txt; sitemap partitioning; canonical and hreflang governance; 404 and redirect governance; image alt text and key data as text.
If you can only do one thing,do server-side rendering. If the page body can't be crawled, the other eight items all come to nothing.
Structured data is a label, not content. Sentences that don't exist on the page can't be conjured up by writing them into markup. Conversely, marking up information that doesn't exist on the page is cheating, and it backfires when models cross-verify.
II. Transformation Order
| Level | Content | Timeline |
|---|---|---|
| P0 | SSR or static generation; add an entity identity section to About Us; add a one-sentence definition to each page; robots and AI crawler strategy; Organization and Article markup | 2 weeks |
| P1 | Q&A-style section restructuring; author page, qualifications page, case study page; llms.txt; sitemap and hreflang governance | 1 month |
| P2 | Comparison matrix; pricing information; monitoring mechanism | Ongoing |
There's an order within P0 too. If the rendering method isn't changed, everything after is just decorating a page nobody can see.
III. Acceptance
GEO has no backend data, so you can only simulate questions and record results. Build 40 to 60 questions in four categories: category demand (who can do a certain thing), selection comparison (which of A and B is more suitable), scenario pain points (how to solve a specific problem), and direct brand lookup (what does this company do, is it reliable).
Run each question in mainstream models using a fresh, independent session, at least twice per question, and record the answer verbatim.
| Metric | Definition |
|---|---|
| Mention rate | Number of questions where the brand appears / total questions |
| First-mention rate | Number of questions where the brand ranks first / total questions |
| Top-recommendation rate | Number of questions where the brand is explicitly recommended or described as more suitable / total questions |
| Official website citation rate | Number of questions citing or linking to the official website / total questions |
| Factual accuracy rate | Number of questions where brand-related facts are stated correctly / number of questions where the brand is mentioned |
Among these, the official website citation rate is the most direct metric for official website optimization; the others are also affected by off-site content.
Must cover 3 to 5 models. The basis is an observational study from September 2026 covering ChatGPT, Copilot, Google, and Perplexity:96.4% of cited URLs appear in only a single engine, and the median cross-engine URL overlap is 0 (paper Scoring With the Engine, arXiv:2609.22655). Drawing conclusions from testing only one model is as good as not testing at all.
The most valuable column in the record table is "cited original sentence". It tells you which sentence on the official website was actually selected. Moving that sentence to a more prominent position and a more structured block is the most effective single-point optimization.
Cadence: keep a baseline before the overhaul, retest once at 1 month and once at 3 months after completion, then run it on a fixed day each month thereafter.
IV. Six common misjudgments
1. Adding structured data means it's done. Markup solves machine recognition, not whether content exists. Without quotable sentences, no amount of markup will get you cited.
Two: only optimizing the homepage. What gets cited is specific pages. Product pages, Q&A pages, and case study pages are the main body; the homepage is only responsible for making the entity identity clear.
Three: treating "industry-leading" as differentiation. These kinds of words can't be extracted and can't be verified—it's as good as saying nothing.
Four: assuming that opening up to AI crawlers means you'll get indexed. That's just a ticket in; whether you get cited is determined by content quality.
Five: treating robots as a confidentiality measure. It's not a permissions system. If you really need confidentiality, don't put it on a public web page.
Six: finishing after one round of changes. Models change, competitors change, and your information goes stale too. This is maintenance work, not a project deliverable.
FAQ
Q: How much does official website optimization actually help GEO?
A: The official website is the baseline source for AI cross-verification. Before changes, what AI gets is a page whose body text can't be read and whose entity can't be recognized, so the citation probability is extremely low. After changes, it at least meets the conditions for being recalled and extracted. But note: official website optimization only solves "having citable evidence"; whether you actually get cited also depends on the off-site content ecosystem, and no method can promise a specific improvement margin.
Q: Is optimizing only the official website enough?
A: Not enough. The official website is the main battleground, but AI makes judgments by synthesizing multiple sources. Improving the official website gives AI citable evidence; content on Baidu Baike, Zhihu, and industry media determines whether it is willing to cite. The statements on both sides must also be consistent, otherwise cross-verification will conclude that "the information is unreliable."
Q: How long after optimization until results are visible?
A: At the indexing level, crawling changes are usually visible in two to four weeks. Citations at the answer level generally take one to three months and require aggregated observation across multiple models. Model answers fluctuate greatly; differences in a single query cannot be treated as a conclusion—look at the trend across two consecutive cycles.
Q: Can template-built sites do GEO optimization?
A: If it can be changed, it can be done. The real blockers are two: whether the rendering method can be changed to server-side output, and whether the backend can customize the page head. If neither is possible, the sixth-layer pipeline is basically dead, and it's worth evaluating a switch of site-building solution. The other five layers are content issues and unrelated to how the site is built.
This article is part of the "100 Questions on Organic Marketing in Chinese" series.View the full table of contents →
Appendix: GEO Official Website Checklist (copy the whole section)
The following content can be copied in full and saved as geo-site-checklist.md for standalone use, or printed as a self-check form. In this document, {{}} are placeholders; replace them when using.
0. How to use
- First run the self-check commands in Section 1 to determine in five minutes whether there are any issues with the technical foundation.
- Go through the six-layer checklist in Section 2 item by item, tick each one off, and record the gaps.
- Schedule according to the priorities in Section 7; do not touch P1 until P0 is cleared.
- Use the templates directly: see Section 3 for structured data, Section 4 for robots, Section 5 for llms.txt, and Section 6 for copy sentence patterns.
- After launch, set up monitoring according to Section 8.
1. Self-check commands
Is the body content in the HTML?
curl -s https://{{DOMAIN}}/ | grep -E '<div id="(root|app|__next)"></div>'If there is output, it indicates pure client-side rendering. If there is no output, then test the body content length:
curl -s https://{{DOMAIN}}/ | sed 's/<[^>]*>//g' | tr -d ' \n' | wc -mIf it is below 500, it can basically be determined that the body content is not in the HTML.
H1 and structured data
curl -s https://{{DOMAIN}}/ | grep -c '<h1'
curl -s https://{{DOMAIN}}/ | grep -o '"@type" *: *"[^"]*"' | sort -uThe first item should be 1. If the second item is empty, it means the page has no structured data at all.
Are AI crawlers being mistakenly blocked?
curl -s -o /dev/null -w "%{http_code}\n" -A "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; GPTBot/1.2; +https://openai.com/gptbot" https://{{DOMAIN}}/
curl -s -o /dev/null -w "%{http_code}\n" -A "Mozilla/5.0 (Macintosh) AppleWebKit/537.36 Chrome/124.0 Safari/537.36" https://{{DOMAIN}}/If the two status codes are inconsistent, it means WAF or anti-crawling rules are blocking AI crawlers. This is more common than robots.txt issues.
Two fixed addresses
https://{{DOMAIN}}/robots.txt
https://{{DOMAIN}}/llms.txtCheck separately: whether robots declares the sitemap and whether AI crawler rules are explicitly written; whether llms.txt exists.
2. Six-layer checklist
[Layer 1 Entity Identity]
[ ] Full company name
[ ] Brand short name
[ ] English name (fill in if available)
[ ] Former name (fill in if available)
[ ] Founding date
[ ] Registered location
[ ] Unified Social Credit Code
[ ] One-sentence definition of the main business (within 40 characters)
[ ] Affiliation between brand name and product name; product pages indicate the publisher
[ ] Name, address, and phone number, with a single version sitewide
[ ] Links to external authoritative profiles (business registration, encyclopedia, WeChat Official Account, GitHub, LinkedIn)
[ ] All of the above are plain text, not images
[Layer 2 Capability Boundaries]
[ ] Definition sentence: what it is, in one sentence
[ ] Target users
[ ] Inapplicable situations (including suggested alternatives)
[ ] Deliverables list
[ ] Delivery cycle and prerequisites
[ ] Pricing method or pricing logic
[ ] Service area, language, and eligibility requirements
[ ] Limitations and prerequisites
[Layer 3 Authoritative Evidence]
[ ] Author byline + author page (bio, qualifications, external identity links)
[ ] Qualification certificate number and issuing body
[ ] Patent number
[ ] Test report number and testing institution
[ ] Customer cases: industry, background, publicly shareable data
[ ] Every number has a source, definition, and date
[ ] Page shows the last updated time
[ ] Key pages retain revision records
[ ] Editorial and fact-checking policy
[ ] Disclaimer
[Layer 4 Differentiated Comparison]
[ ] Comparison table (compare solution types, do not name competitors)
[ ] Capability matrix (by function / scenario / scale tier)
[ ] Explicitly state "what we do not do"
[ ] State costs and trade-offs in comparisons
[Layer 5 Conversion Path]
[ ] Only one primary CTA per page
[ ] Contact information is consistent across all channels
[ ] Contact information has copyable plain text
[ ] Response commitment: how long until reply, who replies
[ ] After-sales and guarantee terms
[Layer 6 Machine Readable]
[ ] Server-side rendering or static generation
[ ] One H1 per page
[ ] H2s use real questions
[ ] Body content is inside main / article
[ ] Organization structured data
[ ] Article structured data
[ ] BreadcrumbList structured data
[ ] Product / Service (by page type)
[ ] FAQPage (only mark visible Q&A)
[ ] robots.txt explicitly declares AI crawlers
[ ] llms.txt is live
[ ] Sitemap is partitioned and lastmod is real
[ ] Canonical is unified
[ ] hreflang is self-consistent and closed (for multilingual sites)
[ ] No soft 404s
[ ] Image alt text is complete
[ ] Key data has a text version3. Structured data templates
Organization entity (injected into every page sitewide)
{
"@context": "https://schema.org",
"@type": "Organization",
"@id": "https://{{DOMAIN}}/#organization",
"name": "{{Full brand name}}",
"alternateName": ["{{Short brand name}}", "{{English name}}"],
"legalName": "{{Full registered business name}}",
"url": "https://{{DOMAIN}}/",
"logo": {
"@type": "ImageObject",
"url": "https://{{DOMAIN}}/logo.png",
"width": 512,
"height": 512
},
"foundingDate": "{{2021-03-15}}",
"description": "{{One-sentence definition, identical to the first paragraph of About Us}}",
"address": {
"@type": "PostalAddress",
"streetAddress": "{{Street address}}",
"addressLocality": "{{City}}",
"addressRegion": "{{Province}}",
"postalCode": "{{Postal code}}",
"addressCountry": "CN"
},
"contactPoint": [{
"@type": "ContactPoint",
"telephone": "{{+86-xxx}}",
"contactType": "sales",
"areaServed": "CN",
"availableLanguage": ["zh-CN", "en"]
}],
"sameAs": [
"{{Business registration record link}}",
"{{Encyclopedia link}}",
"{{WeChat Official Account or social media link}}",
"{{GitHub / LinkedIn / careers page}}"
],
"knowsAbout": ["{{Core capability 1}}", "{{Core capability 2}}", "{{Core capability 3}}"]
}Article page
{
"@context": "https://schema.org",
"@type": "Article",
"@id": "https://{{DOMAIN}}/{{PATH}}#article",
"headline": "{{Title, within 110 characters}}",
"description": "{{80-140 character summary, including brand name and a one-sentence capability definition}}",
"image": ["https://{{DOMAIN}}/{{cover image}}"],
"datePublished": "{{2026-10-23T09:00:00+08:00}}",
"dateModified": "{{2026-10-23T09:00:00+08:00}}",
"inLanguage": "zh-CN",
"author": {
"@type": "Person",
"@id": "https://{{DOMAIN}}/author/{{slug}}#person",
"name": "{{author name}}",
"jobTitle": "{{job title}}",
"url": "https://{{DOMAIN}}/author/{{slug}}"
},
"publisher": { "@id": "https://{{DOMAIN}}/#organization" },
"isPartOf": { "@id": "https://{{DOMAIN}}/#website" },
"about": [{ "@type": "Thing", "name": "{{topic keyword}}" }],
"citation": ["{{URL of external source cited}}"]
}publisher.@id Must match the organization entity's @id exactly the same, otherwise the two nodes are disconnected in the graph.
Q&A block
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [{
"@type": "Question",
"name": "{{question users actually ask}}",
"acceptedAnswer": {
"@type": "Answer",
"text": "{{40-120 characters, conclusion first, no preamble}}"
}
}]
}4. robots.txt template
# Retrieval bots: must allow
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Claude-SearchBot
Allow: /
User-agent: Baiduspider
Allow: /
User-agent: YisouSpider
Allow: /
# Training bots: decide per brand policy
User-agent: GPTBot
Allow: /
User-agent: ClaudeBot
Allow: /
User-agent: Google-Extended
Allow: /
User-agent: Bytespider
Allow: /
# Sensitive paths blocked for all
User-agent: *
Disallow: /admin/
Disallow: /api/
Allow: /
Sitemap: https://{{DOMAIN}}/sitemap.xmlDo not Disallow CSS and JS resources; some engines need them to determine page structure.
5. llms.txt template
# {{Brand Name}}
> {{One-sentence definition: provides C services to B customers in industry A, solving problem D. Also state founding date and location.}}
## Core Pages
- [{{Product Name}}](https://{{DOMAIN}}/{{path}}): {{What question this page answers}}
- [{{Solution}}](https://{{DOMAIN}}/{{path}}): {{One-sentence description}}
- [Pricing](https://{{DOMAIN}}/pricing): {{One-sentence pricing method}}
## Authoritative Information
- [About Us](https://{{DOMAIN}}/about): Full company name, founding date, qualifications, team
- [Customer Cases](https://{{DOMAIN}}/cases): {{Industry and publicly shareable results}}
- [FAQ](https://{{DOMAIN}}/faq): {{Topics covered}}
## Content Sections
- [{{Section Name}}](https://{{DOMAIN}}/{{path}}/): {{Section positioning}}
## Boundaries and Disclaimers
- Service area: {{...}}
- Not applicable scenarios: {{...}}
- Content updates: {{Frequency}}
## Optional
- [{{Secondary Page}}](https://{{DOMAIN}}/{{path}}): Supplementary materialThe opening blockquote will be quoted in full, so it must stand on its own. The sentence after each link cannot be omitted.
6. Answer Unit Sentence Patterns
Definition sentence
{{Brand/Product}} is a {{category}} that provides {{capability}} to {{audience}}, solving {{problem}}.
Boundary sentence
Applicable to scenarios with {{condition A}} and {{condition B}}; if {{counter-condition}}, {{alternative}} is more suitable.
Comparison sentence
Compared with {{solution type A}}, {{this product}} {{difference}} in {{dimension}}, at the cost of {{trade-off}}.
Data sentence
{{Metric}} is {{value}} (source: {{organization or internal statistics}}, definition: {{definition}}, time: {{year-month}}).Self-check standard: delete all adjectives and see if the sentence still holds. If all that remains is "we are committed to providing quality service," it is marketing copy.
Paragraph guidelines: 2 to 4 sentences, conclusion in the first sentence; key information in tables; write absolute dates, not "recently"; numbers with units and definitions; key data must not exist only in images.
7. Priority and Timeline
| Level | Content | Timeline |
|---|---|---|
| P0 | SSR or static generation; entity identity section; one-sentence definition per page; robots and AI crawler policy; Organization and Article markup | 2 weeks |
| P1 | Q&A-style section restructuring; author page, qualifications page, case study page; llms.txt; sitemap and hreflang governance | 1 month |
| P2 | Comparison matrix; pricing information; monitoring mechanism | Ongoing |
P0 internal order: rendering → entity → definition sentence → crawler policy → structured data. If earlier steps are not done, later ones are ineffective.
8. Acceptance Records
Four types of question set sentence patterns
Category demand: Which service providers in China offer {{category}}
Selection comparison: Which is more suitable for {{object}} of {{scale}}, {{solution A}} or {{solution B}}
Scenario pain point: How to solve {{specific problem}}
Brand direct query: What does {{brand}} do, and is it reliableFive metrics
| Metric | Definition |
|---|---|
| Mention rate | Number of questions where the brand appears / total questions |
| First-mention rate | Number of questions where the brand ranks first / total questions |
| Top-recommendation rate | Number of questions where the brand is explicitly recommended or described as more suitable / total questions |
| Official website citation rate | Number of questions citing or linking to the official website / total questions |
| Factual accuracy rate | Number of questions where brand-related facts are stated correctly / number of questions where the brand is mentioned |
Record table fields
Date | Model | Question ID | Question text | Web-enabled | Mentioned | Position | Recommended | Sentiment | Official site cited | Cited page URL | Cited original sentence | Factual error"Cited original sentence" is the most valuable column: it tells you which sentence was actually selected. Moving that sentence to a more prominent, more structured block is the most effective single-point optimization.
Cadence: Keep a baseline before changes, re-test at 1 month and 3 months after completion, then run on a fixed day each month. Cover at least 3 to 5 models.
Appendix: Full SKILL Document (One-Click Copy)
The appendix above is a checklist for human reference; here is the complete SKILL that can be loaded directly into AI tools. It includes workflow definitions, the six-layer checklist, all structured data templates, AI crawler policy, answer unit sentence patterns, acceptance methods, and a health-check script that relies only on the standard library.
After copying, split it into the corresponding files according to the <!-- ===== File name ===== --> comments at the top of the file. The directory structure is as follows:
geo-site-revamp/
├── SKILL.md
├── scripts/audit_site.py
└── references/
├── audit-checklist.md
├── schema-templates.md
├── ai-crawler-policy.md
├── answer-unit-patterns.md
└── monitoring.md<!-- ===== SKILL.md ===== -->
---
name: geo-site-revamp
description: 审计并改造企业官网以满足生成式引擎优化（GEO）要求，让品牌在 AI 搜索与对话式回答中被提及、被引用。当用户要做 GEO 官网改造、AI 搜索优化、官网结构化数据、llms.txt、AI 爬虫策略、AEO、品牌在 AI 回答中的可见度，或需要对站点做 GEO 体检时使用。包含可直接执行的站点体检脚本、六层信息点清单、JSON-LD 模板、答案单元句式与验收指标。
---
# GEO 官网改造
## 核心判断
被大模型引用的最小单位不是页面，而是一个能被单独摘走的句子或字段。改造的一切动作都服务于：让官网存在足够多、足够准、足够结构化的「答案单元」。
两类信息点必须同时补：
- **事实型**：可结构化、可交叉校验的客观字段（实体信息、参数、价格、资质）。
- **叙述型**：可整段摘录的定义句、边界句、对比句、数据句。
只补 schema 不改正文，不会被引用；只改正文不加结构化数据，召回效率低。
## 工作流
### 1. 确定范围
确认目标域名、主要语言、业务类型（B2B 服务 / 电商 / 内容站 / 多语言站），以及是否已有 sitemap、robots、llms.txt。业务类型决定第 2、4 层的具体写法。
### 2. 跑体检
```bash
python3 scripts/audit_site.py https://example.com --max-pages 5 --json geo-audit.json
```
脚本检查：robots 与各 AI 爬虫策略、llms.txt、sitemap、每页 H1/H2 数量、JSON-LD 类型、canonical、hreflang、alt 缺失、正文长度、纯 CSR 迹象，并用 GPTBot UA 复抓首页对比是否被反爬误拦。输出按 P0/P1/P2 排序的改造项。
脚本只覆盖第 6 层（机器可读载体）和第 1 层（实体 schema）。第 2-5 层是内容问题，必须人工读页面判断。
### 3. 逐层诊断
按 `references/audit-checklist.md` 的六层顺序过一遍，产出缺口表：
1. 实体身份：名称、成立时间、注册地、一句话定义、NAP、sameAs。
2. 能力边界：定义句、适用与不适用、交付物、计价、限制条件。
3. 权威证据：作者页、资质编号、案例数据、更新时间、事实核查政策。
4. 差异化对比：对比表、能力矩阵、「我们不做什么」。
5. 转化路径：唯一 CTA、可复制联系方式、响应承诺。
6. 机器可读：SSR、语义化 HTML、schema、robots、llms.txt、sitemap、canonical/hreflang。
### 4. 改造与产出
按 P0 → P1 → P2 推进。写代码时遵循：
- schema 按 `references/schema-templates.md` 注入，全站用同一个 `@id` 串成图；只标记页面真实可见的内容。
- 页面文案按 `references/answer-unit-patterns.md` 重写，一页八问必须答到五项以上。
- robots 与 llms.txt 按 `references/ai-crawler-policy.md` 落地，配置后用真实 UA 实测状态码与正文长度。
- 纯 CSR 站点必须先改服务端渲染或静态生成，其余优化在此之前没有意义。
### 5. 验收
按 `references/monitoring.md` 建品牌问题集并留基线。验收看四项：体检 P0 归零、官网引用率相对基线提升、事实准确率 ≥90%、核心品类问题能被官网信息覆盖。
## 资源索引
| 文件 | 用途 |
|---|---|
| `scripts/audit_site.py` | 站点 GEO 体检，输出优先级化改造项 |
| `references/audit-checklist.md` | 六层信息点清单、优先级、常见误判 |
| `references/schema-templates.md` | Organization/WebSite/Article/Product/FAQPage/Breadcrumb JSON-LD 模板 |
| `references/ai-crawler-policy.md` | AI 爬虫清单、robots 模板、llms.txt 模板 |
| `references/answer-unit-patterns.md` | 五类可引用的句式模板与段落规范 |
| `references/monitoring.md` | 问题集设计、指标定义、记录表、归因纪律 |
## 输出规范
- 诊断结果按「层级 → 缺口 → 改造动作 → 优先级」四列呈现，不要只给问题清单。
- 涉及多语言站时，检查各语言的实体信息是否一致，hreflang 是否闭合。
- 不要建议伪造评分、虚构客户、编造数据，AI 交叉验证会反噬品牌。
- 不要承诺「做了就能被引用」，GEO 只能提高概率，答案由模型侧决定。
<!-- ===== references/audit-checklist.md ===== -->
# 官网 GEO 改造六层信息点清单
改造目标不是「让爬虫收录页面」，而是「让大模型在回答某类问题时，愿意用你的官网内容当依据，并在答案里点到你的名字」。
被引用的最小单位不是页面，是**一个可以被直接摘走的句子或字段**。所以清单按两类信息点组织：
- **事实型**：可结构化、可交叉校验的客观字段（名称、地址、参数、价格、资质）。
- **叙述型**：可被整段摘录的结论句、定义句、对比句（是什么、适合谁、和谁不同）。
大多数官网的现状是事实散落、叙述空洞，两类都缺。
---
## 第 1 层 实体身份：AI 要能确认「你是哪家公司」
优先级最高且最常被忽略。大模型把品牌当作知识图谱中的一个实体，找不到唯一指代就会混同或干脆不提。
| 信息点 | 常见问题 | 改造要求 |
|---|---|---|
| 公司全称 / 简称 / 英文名 / 曾用名 | 全站只有品牌简称 | 「关于我们」首段一句话讲清四者对应关系 |
| 成立时间 / 注册地 / 统一社会信用代码 | 完全没有 | 纯文本列出，不要做成图片 |
| 主营一句话定义 | 「致力于打造生态闭环」 | 换成「我们为 A 行业的 B 客户提供 C 服务，解决 D 问题」 |
| 品牌名与产品名的隶属关系 | 产品页不提出品方 | 每个产品页标注「由 X 出品」 |
| NAP（名称/地址/电话） | 页脚、联系页、招聘页三处不一致 | 收敛为全站唯一版本 |
| sameAs 外部权威档案 | 无 | 工商、百科、公众号、GitHub、领英、招聘页互指 |
硬要求：外部档案上的说法必须与官网**逐字一致**。AI 交叉验证时不一致会降低实体置信度。
## 第 2 层 能力边界：AI 要能判断「你是不是这个问题的答案」
用户问 AI 时最常见的形式是「谁能做某件事」。每项业务都要有一份「标准答案段」，而不是营销文案。
- **定义句**：是什么，一句话，不超过 40 字。
- **适用对象**：明确「适合谁」和「不适合谁」。有边界的表述才可能被引用，因为 AI 要回答「是否适合我」。
- **交付物与周期**：细到可以列清单。
- **计价方式**：区间或计价逻辑。可以不公开价格，但不能让 AI 无话可说。
- **服务区域 / 语言 / 准入条件**。
- **限制与前置条件**：反直觉但关键，它是可信度的来源。
## 第 3 层 权威证据：AI 要能回答「凭什么信你」
- 作者或专家署名 + 个人页（履历、资质、可验证的外部身份）。
- 资质证书、专利、检测报告：写编号与发证机构，不要只放扫描图。
- 客户案例：行业、项目背景、可公开的结果数据。
- 数据出处：数字标注来源、统计口径、时间。
- 页面显示「最后更新时间」，关键页有修订记录。
- 编辑与事实核查政策、免责声明（国内站点几乎都没做，做了就是差异化）。
## 第 4 层 差异化对比：AI 要能回答「和别家比怎么样」
AI 回答选型类问题必然做比较。官网没有对比信息，AI 就去第三方来源拼凑，而第三方评价你控制不了。
- 与同类方案的对比表（对比「品类/方案类型」，不点名对手）。
- 能力矩阵：按功能、场景、规模分档。
- 明确写下「我们不做什么」「什么场景下我们不是最优解」。
## 第 5 层 转化与信任路径：AI 的答案要能落到一个动作
- 唯一首选行动入口：咨询 / 试用 / 报价 / 预约，一页只给一个主 CTA。
- 联系方式全渠道一致，提供可复制的纯文本（不要只有图标或二维码）。
- 服务响应承诺：多久回、谁回。
- 售后与保障条款。
## 第 6 层 机器可读载体：把前五层送出去
按优先级：
1. **服务端渲染或静态生成**。AI 爬虫普遍不执行 JS，纯 CSR 的正文等于不存在。P0。
2. **语义化 HTML**：每页一个 H1，H2 用真实问题句式，正文在 `main`/`article` 内。
3. **结构化数据**：`Organization`（含 `@id`、`sameAs`）、`WebSite`、`Article`（`author`/`datePublished`/`dateModified`）、`Product`/`Offer`、`FAQPage`、`HowTo`、`BreadcrumbList`。用 `@id` 把作者、发布方、主站实体串成一张图，避免孤岛。
4. **robots.txt 区分 AI 爬虫用途**（检索 vs 训练），见 `ai-crawler-policy.md`。
5. **llms.txt / ai.txt**：站点摘要 + 关键页面索引。
6. **sitemap.xml**：分语言、分内容类型，`lastmod` 真实。
7. **canonical 与 hreflang 治理**：多语言站必须自洽闭合，否则各语言版本互相镜像。
8. **404 与重定向治理**：软 404 会拖低站点质量判定。
9. **alt 与表格文本化**：关键数据不能只存在于图片里，AI 读不到图里的数字。
---
## 优先级与工期
| 级别 | 内容 | 工期 |
|---|---|---|
| P0 | SSR/静态化、实体身份段、每页一句话定义、robots 与 AI 爬虫策略、Organization 与 Article schema | 2 周 |
| P1 | 问答式小节重构、权威证据页（作者/资质/案例）、llms.txt、sitemap 与 hreflang 治理 | 1 个月 |
| P2 | 对比矩阵、计价信息、监测机制 | 持续 |
## 验收
- 体检脚本 P0 归零，P1 逐项消解：`scripts/audit_site.py <站点域名>`
- 品牌问题集的官网覆盖率（见 `monitoring.md`）
- 各模型下的提及率、首位提及率、首选推荐率、正负面提及率的变化
## 常见误判
- 加了一堆 schema 但正文没有可引用的句子 → 仍不会被引用。schema 是标签，不是内容。
- 只优化首页 → AI 引用的是具体页面，产品页、问答页、案例页才是主战场。
- 把「行业领先」「赋能」当差异化 → 这类词无法被摘取，等于没说。
- 给 AI 爬虫开白就等于被收录 → 只是入场券，能否被引用由内容质量决定。
<!-- ===== references/schema-templates.md ===== -->
# JSON-LD 模板（GEO 用）
原则：**全站一张图**，不是一堆碎片。用 `@id` 让文章、作者、发布方、主站互相指向同一个实体节点。
把占位符（`{{...}}`）替换后直接内联到 `<head>` 或 `<body>` 末尾。
## 1. 组织实体（全站每页注入，P0）
```json
{
"@context": "https://schema.org",
"@type": "Organization",
"@id": "https://{{DOMAIN}}/#organization",
"name": "{{公司全称}}",
"alternateName": ["{{品牌简称}}", "{{英文名}}"],
"legalName": "{{工商登记全称}}",
"url": "https://{{DOMAIN}}/",
"logo": {
"@type": "ImageObject",
"url": "https://{{DOMAIN}}/logo.png",
"width": 512,
"height": 512
},
"foundingDate": "{{YYYY-MM-DD}}",
"description": "{{一句话定义：为 A 行业的 B 客户提供 C 服务，解决 D 问题}}",
"address": {
"@type": "PostalAddress",
"streetAddress": "{{详细地址}}",
"addressLocality": "{{城市}}",
"addressRegion": "{{省}}",
"postalCode": "{{邮编}}",
"addressCountry": "CN"
},
"contactPoint": [{
"@type": "ContactPoint",
"telephone": "{{+86-xxx}}",
"contactType": "sales",
"areaServed": "CN",
"availableLanguage": ["zh-CN", "en"]
}],
"sameAs": [
"{{工商档案链接}}",
"{{百科链接}}",
"{{公众号或社媒链接}}",
"{{GitHub / 领英 / 招聘主页}}"
],
"knowsAbout": ["{{核心能力1}}", "{{核心能力2}}", "{{核心能力3}}"]
}
```
要点：
- `legalName` 与 `name` 分开，工商全称和对外品牌名不同时尤其重要。
- `sameAs` 至少 3 条，且每个链接页面上必须能找到公司全称。
- `knowsAbout` 填能力词，这是引擎判断「你属于哪个领域」的直接信号。
## 2. 站点节点
```json
{
"@context": "https://schema.org",
"@type": "WebSite",
"@id": "https://{{DOMAIN}}/#website",
"url": "https://{{DOMAIN}}/",
"name": "{{站点名}}",
"inLanguage": "zh-CN",
"publisher": { "@id": "https://{{DOMAIN}}/#organization" }
}
```
## 3. 文章页
```json
{
"@context": "https://schema.org",
"@type": "Article",
"@id": "https://{{DOMAIN}}/{{PATH}}#article",
"headline": "{{标题，<=110 字符}}",
"description": "{{80-140 字摘要，含实体名与一句能力定义}}",
"image": ["https://{{DOMAIN}}/{{封面图}}"],
"datePublished": "{{2026-01-01T09:00:00+08:00}}",
"dateModified": "{{2026-01-01T09:00:00+08:00}}",
"inLanguage": "zh-CN",
"author": {
"@type": "Person",
"@id": "https://{{DOMAIN}}/author/{{slug}}#person",
"name": "{{作者名}}",
"jobTitle": "{{职位}}",
"url": "https://{{DOMAIN}}/author/{{slug}}",
"sameAs": ["{{作者可验证的外部身份链接}}"]
},
"publisher": { "@id": "https://{{DOMAIN}}/#organization" },
"isPartOf": { "@id": "https://{{DOMAIN}}/#website" },
"about": [{ "@type": "Thing", "name": "{{主题词}}" }],
"citation": ["{{引用的外部权威来源 URL}}"]
}
```
`dateModified` 必须真实更新，否则是自毁可信度。`citation` 是 GEO 里价值最高的字段之一。
## 4. 产品 / 服务
```json
{
"@context": "https://schema.org",
"@type": "Product",
"@id": "https://{{DOMAIN}}/{{PATH}}#product",
"name": "{{产品名}}",
"description": "{{一句话定义 + 适用对象}}",
"brand": { "@id": "https://{{DOMAIN}}/#organization" },
"category": "{{品类}}",
"offers": {
"@type": "Offer",
"priceCurrency": "CNY",
"price": "{{价格或起价}}",
"priceSpecification": {
"@type": "PriceSpecification",
"priceCurrency": "CNY",
"minPrice": "{{起}}",
"maxPrice": "{{止}}"
},
"availability": "https://schema.org/InStock",
"url": "https://{{DOMAIN}}/{{PATH}}"
}
}
```
纯服务型业务用 `Service` + `areaServed` + `provider`，不要硬套 `Product`。
## 5. 问答页 / FAQ 区块
```json
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [{
"@type": "Question",
"name": "{{用户真实会问的问题}}",
"acceptedAnswer": {
"@type": "Answer",
"text": "{{40-120 字直接回答，结论先行，不铺垫}}"
}
}]
}
```
只标记页面上**真实可见**的问答。隐藏内容标记 FAQ 会被判作弊。
## 6. 面包屑
```json
{
"@context": "https://schema.org",
"@type": "BreadcrumbList",
"itemListElement": [
{ "@type": "ListItem", "position": 1, "name": "首页", "item": "https://{{DOMAIN}}/" },
{ "@type": "ListItem", "position": 2, "name": "{{栏目}}", "item": "https://{{DOMAIN}}/{{栏目路径}}/" }
]
}
```
## 7. 多语言站
每个语言版本各自注入本地化的 Organization 与 Article，`inLanguage` 与站点 `hreflang` 严格对应；`@id` 用同一路径 + 语言后缀区分，不要复用同一个 `@id`。
## 校验
- Google Rich Results Test、Schema Markup Validator 过一遍语法。
- 用同一 `@id` 检查是否真的连成图：文章 `publisher.@id` 是否等于组织节点 `@id`。
- 禁止把不存在的信息写进 schema（虚构评分、虚构价格），AI 交叉验证会反噬。
<!-- ===== references/ai-crawler-policy.md ===== -->
# AI 爬虫策略与 llms.txt 模板
## 先分清三种用途
AI 侧访问官网有三种动机，对应完全不同的决策：
| 用途 | 行为 | 你想要的 |
|---|---|---|
| 训练语料抓取 | 批量、低频、抓全站 | 视策略而定，多数 B 端品牌可以放开，换取模型内部知识 |
| 搜索索引构建 | 定期抓取，用于生成答案时召回 | **必须允许**，这是被引用的前提 |
| 用户触发的实时浏览 | 用户提问时当场打开你的页面 | **必须允许**，且页面要能快速渲染出正文 |
只关心「被引用」的话，第二、三类是关键；只禁第一类通常没有意义，因为同一批内容往往也服务检索。
## 机器人清单
以各厂商官方文档为准，UA 会变，配置前复核。
| UA 标识 | 归属 | 主要用途 |
|---|---|---|
| `GPTBot` | OpenAI | 训练与内容抓取 |
| `OAI-SearchBot` | OpenAI | 搜索索引，用于 ChatGPT 搜索答案 |
| `ChatGPT-User` | OpenAI | 用户触发实时访问 |
| `ClaudeBot` / `anthropic-ai` | Anthropic | 训练抓取 |
| `Claude-User` / `Claude-SearchBot` | Anthropic | 用户触发 / 检索 |
| `PerplexityBot` | Perplexity | 检索与引用 |
| `Google-Extended` | Google | 控制 Gemini 训练，不影响搜索收录 |
| `Bytespider` | 字节跳动 | 豆包 / 头条生态 |
| `Baiduspider` | 百度 | 百度搜索，兼供百度 AI 生态 |
| `YisouSpider` | 神马 / 夸克生态 | 检索 |
| `Applebot-Extended` | Apple | Apple Intelligence 训练控制 |
| `CCBot` | Common Crawl | 公共语料，被大量模型间接使用 |
| `meta-externalagent` | Meta | 训练抓取 |
## robots.txt 模板
```
# 搜索与检索类：必须允许，否则等于放弃被引用
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Claude-SearchBot
Allow: /
User-agent: Baiduspider
Allow: /
User-agent: YisouSpider
Allow: /
# 训练类：按品牌策略决定，下面这组是「允许」
User-agent: GPTBot
Allow: /
User-agent: ClaudeBot
Allow: /
User-agent: Google-Extended
Allow: /
User-agent: Bytespider
Allow: /
User-agent: CCBot
Allow: /
# 敏感路径对全部爬虫关闭（后台、报价策略、客户资料）
User-agent: *
Disallow: /admin/
Disallow: /api/
Disallow: /uploads/private/
Allow: /
Sitemap: https://{{DOMAIN}}/sitemap.xml
```
注意事项：
- `Disallow: /` 只对匹配到的组生效，别指望一行挡住所有爬虫。
- 不要 `Disallow` 掉 CSS/JS 资源，部分引擎需要它判断页面结构。
- 反爬 WAF 对 AI UA 的误拦比 robots 更常见。配置后必须用真实 UA 实测：`curl -A "<UA>" -I https://{{DOMAIN}}/`，状态码和正文长度要与浏览器一致。
- 不要用 `Disallow` 当保密手段。真要保密就别放在公网页面里。
## llms.txt 模板
放在站点根，`https://{{DOMAIN}}/llms.txt`。目标是让模型用最少 token 理解你是谁、有哪些页面值得读。
```
# {{品牌名}}
> {{一句话定义：为 A 行业的 B 客户提供 C 服务，解决 D 问题。成立时间、所在地、规模等实体信息一句话内交代。}}
## 核心页面
- [{{产品/服务名}}](https://{{DOMAIN}}/{{path}}): {{一句话说明该页回答什么问题}}
- [{{解决方案名}}](https://{{DOMAIN}}/{{path}}): {{说明}}
- [定价](https://{{DOMAIN}}/pricing): {{计价方式一句话}}
## 权威信息
- [关于我们](https://{{DOMAIN}}/about): 公司全称、成立时间、资质、团队
- [客户案例](https://{{DOMAIN}}/cases): {{行业与可公开结果}}
- [常见问题](https://{{DOMAIN}}/faq): {{覆盖的问题主题}}
## 内容栏目
- [{{栏目名}}](https://{{DOMAIN}}/{{path}}/): {{栏目定位}}
## 边界与声明
- 服务区域：{{...}}
- 不适用场景：{{...}}
- 内容更新：{{更新频率}}
## Optional
- [{{次级页面}}](https://{{DOMAIN}}/{{path}}): 补充材料
```
写法要求：
- 首段 blockquote 就是「标准答案段」，模型可能整段引用，必须能独立成立。
- 每个链接后面必须有一句说明，否则只是一张 URL 表，价值极低。
- 全站信息变了要同步更新，过期摘要比没有更糟。
- 多语言站可为每种语言各出一份，用 `llms-<lang>.txt`，并在主文件里指路。
## ai.txt（可选）
若干社区约定在用，非官方标准。内容是允许/禁止 AI 使用的声明与联系邮箱。写了不亏，但不要指望它起决定作用。真正的控制手段是 robots 与访问控制。
<!-- ===== references/answer-unit-patterns.md ===== -->
# 可被引用的答案单元写法
模型不会整篇引用，它摘句子。所以每一段都要能单独成立：脱离上下文后，读者仍知道「谁、做什么、对谁、有什么条件」。
## 五类必备句式
### 1. 定义句（每页必有，放在首段第二句内）
模板：`{{品牌/产品}}是{{品类}}，为{{对象}}提供{{能力}}，用于解决{{问题}}。`
差：我们致力于以创新技术赋能企业数字化转型。
好：XX 是企业级数据备份服务，为 200 人以上的研发团队提供跨云增量备份，用于解决多环境数据一致性校验困难的问题。
判定标准：把句子里所有形容词删掉，信息量是否基本不变。变了说明靠形容词撑场面，删掉后句子仍然成立才合格。
### 2. 边界句（决定能否被回答「适不适合我」）
模板：`适用于{{条件 A}}、{{条件 B}}的场景；如果{{反条件}}，更适合{{替代方案}}。`
写「不适合谁」不会丢客户，反而提升被引用概率，因为模型需要给出推荐理由和排除理由。全站没有排除句，等于把自己排除在选型类问答之外。
### 3. 对比句（用于选型问题）
模板：`与{{方案类型 A}}相比，{{本产品}}在{{维度}}上{{差异}}，代价是{{取舍}}。`
对比「方案类型」而不是点名对手，既避免法律风险，又覆盖用户真实的比较语境。必须写出代价，只讲优势的对比会被判定为营销内容。
### 4. 数据句（提升可信度权重）
模板：`{{指标}}为{{数值}}（来源：{{机构/内部统计}}，口径：{{定义}}，时间：{{年月}}）。`
没有来源的数字在 GEO 里是负资产。宁可给一个带口径的小数字，也不要给一个没出处的漂亮数字。
### 5. 问答句（H2/H3 直接写成问题）
用户向 AI 提问用的是问句，页面小节标题也用问句，召回匹配度最高。
- 好：`多环境备份如何保证一致性？`、`XX 和自建备份方案的差别在哪？`
- 差：`产品优势`、`技术架构`、`解决方案概述`
## 段落与版式规范
- 段落 2-4 句，一句一个信息点，不写长复合句。
- 结论先行，背景后置。第一句就是答案。
- 关键信息用表格承载：参数、对比、计价、资质。表格是最容易被完整摘取的格式。
- 列表项写成完整短句，不要写成「高效」「稳定」「安全」这类单词标签。
- 时间写绝对时间（2026 年 3 月），不写「近期」「目前」。
- 数字带单位和口径，百分比注明基数。
- 关键数据不要只放在图片或图表里，必须有一份文字版。
## 页面骨架检查
一页合格的产品/服务页，按顺序应该能回答：
1. 这是什么（定义句）
2. 解决谁的什么问题（场景）
3. 适合与不适合（边界句）
4. 具体交付什么、多久（清单）
5. 怎么计价（计价逻辑）
6. 谁在用、效果如何（案例 + 数据句）
7. 和别的方式比差在哪（对比句）
8. 怎么开始（唯一 CTA）
八项缺三项以上，这页基本不会被引用。
## 自检口诀
逐段问一句：**这段话能不能直接贴进别人的回答里，而不显得突兀？** 能，就是合格的答案单元；不能，就是给人看的营销文案。
<!-- ===== references/monitoring.md ===== -->
# 效果监测与验收
GEO 没有后台数据可看，只能靠「模拟用户提问 + 记录答案」。所以监测方案要先建起来，否则改造完无法判断有没有用。
## 1. 建品牌问题集
按四类各取 8-15 条，共 40-60 条。问题必须写成真实用户口吻，不要写成关键词。
| 类型 | 示例 | 考察点 |
|---|---|---|
| 品类需求 | 「国内做跨云数据备份的服务商有哪些」 | 是否被列入候选 |
| 选型对比 | 「A 方案和 B 方案哪个更适合 200 人研发团队」 | 是否被当作比较对象 |
| 场景痛点 | 「多环境数据库备份不一致怎么解决」 | 是否被当作解决方案 |
| 品牌直查 | 「XX 公司是做什么的，靠谱吗」 | 实体是否被正确识别 |
固定问题集，不要每次换题，否则无法纵向比较。
## 2. 探测方法
- 每个模型用**独立新会话**，避免上下文污染。
- 同一问题在同一模型至少跑 2 次，结果可能波动，取稳定表现。
- 记录答案原文，不要只记结论。
- 覆盖主流入口：DeepSeek、豆包、元宝、通义、Kimi、百度 AI 搜索，以及带联网检索的通用助手。
- 有联网能力的模式与纯模型模式分开记录，前者反映检索与引用能力，后者反映模型内部知识。
## 3. 指标定义
| 指标 | 定义 |
|---|---|
| 提及率 | 答案中出现品牌名的题数 / 总题数 |
| 首位提及率 | 品牌出现在第一个被推荐位置的题数 / 总题数 |
| 首选推荐率 | 被明确表述为「推荐」「更适合」「首选」的题数 / 总题数 |
| 官网引用率 | 答案引用或链接到官网的题数 / 总题数 |
| 正面提及率 / 负面提及率 | 含明确正/负面评价的题数 / 总题数 |
| 事实准确率 | 品牌相关事实（业务、规模、价格）表述正确的题数 / 被提及题数 |
「官网引用率」是官网改造最直接的效果指标。其他指标还受站外内容影响，改造周期内不要混为一谈。
## 4. 记录表字段
```
日期 | 模型 | 问题ID | 问题文本 | 是否强调联网 | 品牌是否提及 | 位次 | 是否推荐 | 情感 | 是否引用官网 | 引用页面URL | 引用到的原句 | 事实错误 | 截图/链接
```
「引用到的原句」这一列最重要：它直接告诉你官网上哪句话被模型选中了，把这句挪到更靠前的位置、更结构化的区块，是最有效的迭代动作。
## 5. 迭代节奏
- **改造前**：跑一遍全量问题集，留基线，同时跑体检脚本存档。
- **改造中**：每完成一层，抽 10 条核心问题复测。
- **改造后 1 个月、3 个月**：全量复测，对比基线。
- 每月固定一天跑，形成时间序列。
## 6. 归因纪律
- 模型答案波动大，单次差异不算结论，看连续两个周期的趋势。
- 官网改造主要影响「引用率」和「事实准确率」；提及率和首选推荐率还取决于第三方内容生态，不能全部归功或归咎于官网。
- 出现负面提及，先查信息源，再改官网对应页面的表述，不要直接删内容。
## 7. 验收标准
1. 体检脚本 P0 全部归零。
2. 问题集中「官网引用率」相对基线提升。
3. 「事实准确率」达到 90% 以上（说明实体信息已经统一）。
4. 核心品类问题中，官网信息能覆盖答案里的关键论点（人工比对）。
<!-- ===== scripts/audit_site.py ===== -->
保存为 scripts/audit_site.py，用法：
```bash
python3 scripts/audit_site.py https://example.com --max-pages 8 --json report.json
```
```python
#!/usr/bin/env python3
"""GEO 官网体检：抓取页面并输出可改造项清单。
只依赖标准库。检查项对应 references/audit-checklist.md 的前五层。
用法:
python3 audit_site.py https://example.com
python3 audit_site.py https://example.com --max-pages 8 --json report.json
python3 audit_site.py https://example.com --no-ai-ua # 跳过 AI 爬虫 UA 对比
"""
import argparse
import gzip
import json
import re
import ssl
import sys
import urllib.error
import urllib.request
from html.parser import HTMLParser
from typing import Dict, List, Optional
from urllib.parse import urljoin, urlparse
TIMEOUT = 20
BROWSER_UA = (
"Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 "
"(KHTML, like Gecko) Chrome/124.0 Safari/537.36"
)
GPTBOT_UA = "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; GPTBot/1.2; +https://openai.com/gptbot"
AI_BOTS = [
"GPTBot", "OAI-SearchBot", "ChatGPT-User", "ClaudeBot", "Claude-Web",
"anthropic-ai", "PerplexityBot", "Google-Extended", "Bytespider",
"Baiduspider", "YisouSpider", "Applebot-Extended", "CCBot", "meta-externalagent",
]
SKIP_TEXT_TAGS = {"script", "style", "noscript", "svg", "template"}
def fetch(url: str, ua: str = BROWSER_UA, timeout: int = TIMEOUT):
"""返回 (status, headers, body_text)，异常时 status 为 None。"""
req = urllib.request.Request(
url,
headers={
"User-Agent": ua,
"Accept": "text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8",
"Accept-Language": "zh-CN,zh;q=0.9,en;q=0.8",
},
)
ctx = ssl.create_default_context()
ctx.check_hostname = False
ctx.verify_mode = ssl.CERT_NONE
try:
with urllib.request.urlopen(req, timeout=timeout, context=ctx) as resp:
raw = resp.read()
if resp.headers.get("Content-Encoding", "").lower() == "gzip":
try:
raw = gzip.decompress(raw)
except OSError:
pass
charset = resp.headers.get_content_charset() or "utf-8"
try:
text = raw.decode(charset, errors="replace")
except LookupError:
text = raw.decode("utf-8", errors="replace")
return resp.status, dict(resp.headers), text
except urllib.error.HTTPError as exc:
return exc.code, dict(exc.headers or {}), ""
except Exception as exc: # noqa: BLE001
return None, {"error": str(exc)}, ""
class PageParser(HTMLParser):
def __init__(self):
super().__init__(convert_charrefs=True)
self.title = ""
self.meta: Dict[str, str] = {}
self.h1: List[str] = []
self.h2: List[str] = []
self.h3: List[str] = []
self.jsonld: List[str] = []
self.canonical: Optional[str] = None
self.hreflang = 0
self.imgs = 0
self.imgs_no_alt = 0
self.anchors = 0
self.text_chunks: List[str] = []
self.ld_json_raw: List[str] = []
self._stack: List[str] = []
self._skip_depth = 0
self._in = None
self._buf: List[str] = []
def handle_starttag(self, tag, attrs):
a = {k.lower(): (v or "") for k, v in attrs}
self._stack.append(tag)
if tag in SKIP_TEXT_TAGS:
self._skip_depth += 1
elif tag in ("h1", "h2", "h3", "title"):
self._in = tag
self._buf = []
elif tag == "meta":
key = (a.get("name") or a.get("property") or a.get("http-equiv") or "").lower()
if key:
self.meta.setdefault(key, a.get("content", ""))
elif tag == "img":
self.imgs += 1
if not a.get("alt", "").strip():
self.imgs_no_alt += 1
elif tag == "a":
self.anchors += 1
elif tag == "link":
rel = a.get("rel", "").lower()
if "canonical" in rel:
self.canonical = a.get("href", "")
if "alternate" in rel and a.get("hreflang"):
self.hreflang += 1
elif tag == "script" and a.get("type", "").lower() == "application/ld+json":
self._in = "ldjson"
self._buf = []
def handle_endtag(self, tag):
if self._stack and self._stack[-1] == tag:
self._stack.pop()
if tag in SKIP_TEXT_TAGS and self._skip_depth > 0:
self._skip_depth -= 1
if self._in == tag:
value = " ".join("".join(self._buf).split())
if tag == "title":
self.title = value
elif tag == "h1":
self.h1.append(value)
elif tag == "h2":
self.h2.append(value)
elif tag == "h3":
self.h3.append(value)
self._in = None
self._buf = []
elif self._in == "ldjson" and tag == "script":
self.ld_json_raw.append("".join(self._buf))
self._in = None
self._buf = []
def handle_data(self, data):
if self._in:
self._buf.append(data)
return
if self._skip_depth:
return
stripped = data.strip()
if stripped:
self.text_chunks.append(stripped)
def collect_ld_types(blocks: List[str]) -> List[str]:
types = set()
def walk(node):
if isinstance(node, dict):
t = node.get("@type")
if isinstance(t, str):
types.add(t)
elif isinstance(t, list):
types.update(x for x in t if isinstance(x, str))
for v in node.values():
walk(v)
elif isinstance(node, list):
for v in node:
walk(v)
for block in blocks:
block = block.strip()
if not block:
continue
try:
walk(json.loads(block))
except json.JSONDecodeError:
# 容错：抽取 @type 字面量
types.update(re.findall(r'"@type"\s*:\s*"([^"]+)"', block))
return sorted(types)
def parse_robots(text: str):
groups: List[Dict[str, object]] = []
cur: Optional[Dict[str, object]] = None
for line in text.splitlines():
line = line.split("#", 1)[0].strip()
if not line or ":" not in line:
continue
key, val = line.split(":", 1)
key = key.strip().lower()
val = val.strip()
if key == "user-agent":
if cur is None or cur.get("rules"):
cur = {"agents": [], "rules": []}
groups.append(cur)
cur["agents"].append(val)
elif key in ("allow", "disallow") and cur is not None:
cur["rules"].append((key, val))
return groups
def robots_policy(groups, bot: str):
"""返回 (matched_agent, 是否被整站禁止)。"""
for g in groups:
agents = [a.lower() for a in g["agents"]]
if any(a == bot.lower() for a in agents):
disallow_all = any(k == "disallow" and v == "/" for k, v in g["rules"])
allow_all = any(k == "allow" and v == "/" for k, v in g["rules"])
return g["agents"][0], disallow_all and not allow_all
for g in groups:
if any(a == "*" for a in g["agents"]):
disallow_all = any(k == "disallow" and v == "/" for k, v in g["rules"])
return "*", disallow_all
return None, False
def sitemap_urls(base: str, text: str) -> List[str]:
urls = re.findall(r"<loc>\s*([^<\s]+)\s*</loc>", text)
if "<sitemapindex" in text:
out = []
for sub in urls[:3]:
status, _, sub_text = fetch(sub)
if status == 200:
out.extend(re.findall(r"<loc>\s*([^<\s]+)\s*</loc>", sub_text))
return out
return urls
def audit(base: str, max_pages: int, check_ai_ua: bool) -> Dict[str, object]:
base = base.rstrip("/")
report: Dict[str, object] = {"site": base, "findings": [], "pages": []}
findings = report["findings"]
# --- robots.txt ---
status, _, robots_txt = fetch(base + "/robots.txt")
groups = parse_robots(robots_txt) if status == 200 else []
report["robots"] = {
"status": status,
"groups": len(groups),
"ai_bots": {bot: robots_policy(groups, bot)[1] if groups else None for bot in AI_BOTS},
"sitemap_declared": bool(re.search(r"(?im)^sitemap:", robots_txt or "")),
}
if status != 200:
findings.append(("P0", "robots.txt 缺失", "补一份，并显式声明各 AI 爬虫的允许策略与 sitemap 地址"))
else:
blocked = [b for b, v in report["robots"]["ai_bots"].items() if v is True]
if blocked:
findings.append(("P1", "以下 AI 爬虫被整站禁止: " + ", ".join(blocked), "确认是否为预期策略；误拦会让品牌在对应引擎中不可见"))
if not report["robots"]["sitemap_declared"]:
findings.append(("P2", "robots.txt 未声明 Sitemap", "加一行 Sitemap: <绝对地址>"))
# --- llms.txt ---
status, _, llms = fetch(base + "/llms.txt")
report["llms_txt"] = {"status": status, "bytes": len(llms or "")}
if status != 200:
findings.append(("P2", "llms.txt 缺失", "按 references/ai-crawler-policy.md 的模板补站点摘要与关键页面索引"))
# --- sitemap ---
status, _, sm = fetch(base + "/sitemap.xml")
urls = sitemap_urls(base, sm) if status == 200 else []
report["sitemap"] = {"status": status, "urls": len(urls)}
if status != 200 or not urls:
findings.append(("P1", "sitemap.xml 不可用或为空", "生成分区 sitemap，lastmod 用真实更新时间"))
# --- 待检页面 ---
targets = [base + "/"]
for u in urls:
if len(targets) >= max_pages + 1:
break
if u.rstrip("/") == base:
continue
targets.append(u)
all_types = set()
for url in targets:
status, headers, html = fetch(url)
if status != 200 or not html:
report["pages"].append({"url": url, "status": status})
continue
p = PageParser()
try:
p.feed(html)
except Exception: # noqa: BLE001
pass
body_text = " ".join(p.text_chunks)
ld_types = collect_ld_types(p.ld_json_raw)
all_types.update(ld_types)
report["pages"].append({
"url": url,
"status": status,
"bytes": len(html),
"text_chars": len(body_text),
"title_len": len(p.title),
"desc_len": len(p.meta.get("description", "")),
"h1": len(p.h1),
"h2": len(p.h2),
"h1_text": p.h1[0][:80] if p.h1 else "",
"jsonld_types": ld_types,
"canonical": bool(p.canonical),
"hreflang": p.hreflang,
"img_no_alt": p.imgs_no_alt,
"img_total": p.imgs,
"empty_root": bool(re.search(r'<div[^>]+id="(root|app|__next)"[^>]*>\s*</div>', html)),
})
# --- 页面级问题 ---
pages = [x for x in report["pages"] if x.get("status") == 200]
if pages:
no_h1 = [x["url"] for x in pages if x["h1"] == 0]
multi_h1 = [x["url"] for x in pages if x["h1"] > 1]
thin = [x["url"] for x in pages if x["text_chars"] < 600]
empty_root = [x["url"] for x in pages if x["empty_root"] and x["text_chars"] < 300]
no_ld = [x["url"] for x in pages if not x["jsonld_types"]]
no_canon = [x["url"] for x in pages if not x["canonical"]]
no_desc = [x["url"] for x in pages if x["desc_len"] < 30]
img_gap = sum(x["img_no_alt"] for x in pages)
if empty_root:
findings.append(("P0", "检测到纯客户端渲染迹象: " + ", ".join(empty_root[:3]), "改为 SSR/SSG，AI 爬虫普遍不执行 JS，正文等于不存在"))
if no_h1:
findings.append(("P0", "缺 H1 的页面: " + ", ".join(no_h1[:5]), "每页一个 H1，写成能独立回答问题的陈述句"))
if multi_h1:
findings.append(("P1", f"{len(multi_h1)} 个页面存在多个 H1", "收敛为一个，其余降级为 H2"))
if thin:
findings.append(("P1", f"{len(thin)} 个页面正文过短（<600 字）: " + ", ".join(thin[:3]), "补足定义段、适用对象、交付物、限制条件"))
if no_ld:
findings.append(("P1", f"{len(no_ld)} 个页面无 JSON-LD", "按 references/schema-templates.md 注入 Organization/Article/FAQPage 等"))
if no_canon:
findings.append(("P2", f"{len(no_canon)} 个页面缺 canonical", "统一规范地址，避免 AI 侧把参数页当重复内容"))
if no_desc:
findings.append(("P2", f"{len(no_desc)} 个页面 meta description 过短或缺失", "写 80-140 字，含实体名与一句能力定义"))
if img_gap:
findings.append(("P2", f"{img_gap} 张图片缺 alt", "alt 写成名词短语，关键数据不要只存在于图片里"))
hreflang_pages = [x for x in pages if x["hreflang"] > 0]
if hreflang_pages and len(hreflang_pages) < len(pages):
findings.append(("P1", "hreflang 覆盖不完整", "所有语言版本互相声明且闭合，否则各语言页在 AI 侧互相镜像"))
# --- 实体层结构化数据 ---
entity_types = {"Organization", "LocalBusiness", "Corporation", "Brand", "Person"}
report["entity_schema"] = sorted(all_types & entity_types)
if not report["entity_schema"]:
findings.append(("P0", "全站未发现 Organization/LocalBusiness 结构化数据", "补组织实体节点，用 @id 与文章 publisher 串联"))
# --- AI 爬虫 UA 对比 ---
if check_ai_ua:
b_status, b_headers, b_html = fetch(base + "/")
g_status, g_headers, g_html = fetch(base + "/", ua=GPTBOT_UA)
bp, gp = PageParser(), PageParser()
try:
bp.feed(b_html)
gp.feed(g_html)
except Exception: # noqa: BLE001
pass
report["ai_ua_probe"] = {
"browser_status": b_status,
"gptbot_status": g_status,
"browser_chars": len(" ".join(bp.text_chunks)),
"gptbot_chars": len(" ".join(gp.text_chunks)),
}
if b_status != g_status or abs(len(" ".join(bp.text_chunks)) - len(" ".join(gp.text_chunks))) > 200:
findings.append(("P1", "对 GPTBot 与浏览器的响应不一致", "确认是否 waf/反爬误拦 AI 爬虫"))
order = {"P0": 0, "P1": 1, "P2": 2}
findings.sort(key=lambda x: order.get(x[0], 9))
report["summary"] = {k: sum(1 for f in findings if f[0] == k) for k in ("P0", "P1", "P2")}
return report
def main():
ap = argparse.ArgumentParser(description="GEO 官网体检")
ap.add_argument("url")
ap.add_argument("--max-pages", type=int, default=5)
ap.add_argument("--json", dest="json_out")
ap.add_argument("--no-ai-ua", action="store_true")
args = ap.parse_args()
report = audit(args.url, args.max_pages, not args.no_ai_ua)
print("=" * 62)
print("GEO 体检报告 " + report["site"])
print("=" * 62)
print("robots.txt : " + str(report["robots"]["status"]))
print("llms.txt : " + str(report["llms_txt"]["status"]))
print("sitemap : " + str(report["sitemap"]["status"]) + " / " + str(report["sitemap"]["urls"]) + " 条")
print("实体 schema: " + (", ".join(report["entity_schema"]) or "无"))
if report.get("ai_ua_probe"):
print("GPTBot 探测 : " + json.dumps(report["ai_ua_probe"], ensure_ascii=False))
print("\n-- 页面概览 --")
for p in report["pages"]:
if p.get("status") != 200:
print(f" [{p.get('status')}] {p['url']}")
continue
print(f" {p['url']}\n 正文 {p['text_chars']} 字 / H1 {p['h1']} / H2 {p['h2']} / schema {','.join(p['jsonld_types']) or '无'}")
print("\n-- 改造项 --")
if not report["findings"]:
print(" 未发现明显问题")
for level, issue, action in report["findings"]:
print(f" [{level}] {issue}\n → {action}")
print("\n汇总: P0 %d / P1 %d / P2 %d" % (
report["summary"]["P0"], report["summary"]["P1"], report["summary"]["P2"]))
if args.json_out:
with open(args.json_out, "w", encoding="utf-8") as fh:
json.dump(report, fh, ensure_ascii=False, indent=2)
print("JSON 已写入 " + args.json_out)
return 0 if report["summary"]["P0"] == 0 else 1
if __name__ == "__main__":
sys.exit(main())
```Tip:scripts/audit_site.py Requires Python 3.8 or above; no third-party libraries needed. The script exits with code 1 when P0 issues are detected, so it can be plugged directly into CI.
