理解系统
学习可读内容与 Agent 通过 API、feed 或协议执行任务之间的层级。
建立模型
阶段 08 · 改进
Agent-Ready Web：为机器发起的旅程做准备
从内容可读进一步走向授权 Agent 能发现能力、理解权限并安全行动。
结果
评估可发现性、机器可读能力、认证、权限、状态变化与人工监督。
- 先决条件
- 完成或复习上一阶段：GEO 运营系统
- 投入
- 60–90 分钟学习 + 一次工作时段
- 成果
- Agent 旅程图，以及区分发现与执行能力的准备清单。
学习 → 实施 → 证明
完成工作
绘制一个任务从发现到执行的路径，包括权限与失败状态。
制作成果
验证结果
展示受限原型并记录安全和运营边界。
检查证据
Concept boundaries
Machine-readable capability
A clearly described action, input, output, and constraint that software can discover and interpret.
Publishing a description does not make the underlying action safe or executable.
Agentic journey
A task path in which an authorized agent discovers information, makes bounded choices, and may request or perform an action.
Autonomy should not exceed the identity, permission, and approval model.
State-changing action
An operation that creates, updates, purchases, books, sends, deletes, or otherwise changes a system or external commitment.
Machine readability never removes the need for authentication, authorization, validation, idempotency, and recovery.
Core lesson
01
Separate readable information from executable capability
An agent may understand a page yet have no safe path to complete the user's job.
Map the layers: discoverable information, structured facts, capability description, interface contract, authentication, authorization, confirmation, state change, receipt, and recovery. A gap at any layer changes what the journey can responsibly promise.
Static files and metadata can improve discovery. APIs, feeds, tools, or protocols can expose capability. Neither should imply permission to act on behalf of a person without the required controls.
- Name the exact customer job.
- Mark where information ends and state change begins.
- Identify the human confirmation and recovery path.
02
Design the contract around failure
A useful interface declares what happens when data is missing, stale, duplicated, unauthorized, or partially completed.
Define inputs, validation, permissions, side effects, idempotency, rate limits, error states, confirmation language, receipts, and escalation. Treat observability and audit logs as part of the capability, not an afterthought.
Begin with read-only or reversible paths where possible. Expand autonomy only when the evidence shows the team can detect failure, contain impact, and restore a safe state.
- List failure states before the happy path is approved.
- Use least privilege for each action.
- Provide a durable receipt and escalation owner.
Decision framework
Discover → understand → authorize → act → recover
How far can this journey safely progress today?
- 01
Discover
Can an agent find the correct facts and capability description?
Fix canonical information and machine-readable discovery first.
- 02
Understand
Are inputs, outputs, constraints, and freshness explicit?
Define the contract and validation before execution.
- 03
Authorize
Can identity, scope, and user intent be verified?
Keep the path read-only or require human handoff until permissions are sound.
- 04
Act and recover
Can the system prevent duplicates, record the result, and recover from failure?
Permit bounded state change only with confirmation, idempotency, audit, and recovery.
Worked non-client example
A service company wants an AI agent to book consultations directly from a public capability page.
- Service information and availability are readable.
- The booking endpoint has no idempotency key or scoped authorization.
- Cancellation and human escalation are not represented.
Expose a read-only availability check and a human-confirmed booking handoff before allowing autonomous booking.
Discovery is ready, but the state-changing contract and recovery path are not.
A prototype can prove the bounded path works under test conditions; it does not establish production security or universal agent compatibility.
Reusable work template
Agent journey and capability contract
Map one customer job from discovery through recovery.
- 01
User job and authority
State the intended outcome, actor, and who may authorize each step.
- 02
Discoverable facts
List canonical sources, freshness, and machine-readable representations.
- 03
Capability contract
Define action, inputs, outputs, constraints, validation, and side effects.
- 04
Permission and confirmation
Specify identity, scope, least privilege, and human approval points.
- 05
Failure and recovery
List duplicate, timeout, partial, denied, stale, and cancellation paths.
- 06
Audit and evidence
Define logs, receipts, monitoring, owner, and production-readiness boundary.
Failure modes and corrections
Metadata equals capability
A file or schema is presented as proof that agents can complete the task.
Description does not provide an authenticated, reliable execution path.
Separate discovery evidence from executable capability evidence.
Happy-path autonomy
The flow handles success but not duplicates, partial completion, denial, or cancellation.
A state-changing failure can create real commitments or data inconsistency.
Design idempotency, receipts, recovery, and human escalation before autonomy.
Permission inherited from access
Because an endpoint is reachable, the agent is assumed to have authority to use it.
Reachability does not establish identity, scope, or user intent.
Require explicit authentication, authorization, and confirmation boundaries.
Practice exercise
Prototype one bounded agent journey
Choose a real customer job and stop the prototype at the last layer the current controls can support.
- 01Map discovery, understanding, permission, action, receipt, and recovery.
- 02Write the capability contract and failure states.
- 03Prototype a read-only, reversible, or human-confirmed path.
- 04Test denial, duplicate, stale-data, and handoff scenarios.
Proof artifact
An agent journey map, capability contract, bounded prototype evidence, and readiness backlog.
Completion rubric
- Information and execution layers are distinct.
- Authority and human approval are explicit.
- Failure and recovery paths are testable.
- The readiness claim stops at what the prototype actually proves.
GEO 学院知识库
完整知识库
先选阶段，再用主题与阅读意图选择深度。
considerationhubboth
什么是 AI Agent 策略？让品牌为智能体搜索做好准备的入门框架
AI 智能体正在代替 B2B 买家做功课、筛选供应商。想进入智能体搜索的候选名单，你需要建立一套让 AI 准确理解品牌信息的框架——这就是 AI Agent 策略。
阅读文章considerationhubboth
跨境电商 GEO 实战手册：让商品进入 AI 购物助手的推荐名单
AI 购物助手不是在决定把你的商品排在第几名，而是在决定要不要在回答中提到你。跨境卖家想进入推荐名单，第一步是让机器准确读懂你在卖什么。
阅读文章implementationhow-tosearchable
为 AI 提供 Markdown 版内容：/md 端点与纯文本兜底方案
与其反复优化 HTML 语义标记，再指望模型猜对页面结构，不如直接向 AI 提供一份干净的 Markdown 副本。
阅读文章considerationhow-tosearchable
产品页 GEO 的制胜关键：用结构化数据向 AI 购物智能体提供核心信息
当用户转向 AI 寻找产品时，智能体能否提取你的产品信息、正确引用价格与卖点，决定了你在这条新获客路径上的可见度。
阅读文章decisionhow-toboth
如何优化 ChatGPT Shopping 与 Perplexity 商品推荐：四步实战指南
AI 购物引擎不会被漂亮的品牌故事打动，它更看重商品数据是否一致，以及全网评价是否形成共识。做好这四步，提升商品进入 ChatGPT 与 Perplexity 推荐名单的机会。
阅读文章implementationhow-tosearchable
无障碍标记（ARIA）与 AI 可读性：一次设置，同时服务屏幕阅读器与 AI 智能体
为无障碍访问编写的 ARIA 标记，其实也是提供给 AI 引擎的高质量源代码——一次设置，两类读者都能受益。
阅读文章验证任务
验证任务：Agent-Ready Web
展示受限原型并记录安全和运营边界。
- 01保存起始证据 — 学习可读内容与 Agent 通过 API、feed 或协议执行任务之间的层级。
- 02完成阶段成果 — 绘制一个任务从发现到执行的路径，包括权限与失败状态。
- 03按目标复核 — 展示受限原型并记录安全和运营边界。
交付成果
Agent 旅程图，以及区分发现与执行能力的准备清单。
延伸资料库
继续主题学习
核心阶段保持开放；部分进阶白皮书继续使用原有解锁方式。
白皮书6 章 · 2026.06
从被读到被执行：MCP 与 B2A 的 agent-ready 基建
MCP 已成为各大 AI 供应商共用的连接器,WebMCP 让网站能对外暴露结构化功能,让 agent 直接执行任务。这份白皮书说明 agentic web 如何把营销从「被读」推向「被使用」,以及 B2A 基建怎么开始做。
从被读到被执行：MCP 与 B2A 的 agent-ready 基建指南6 章 · 2026.06
llms.txt 部署指南
从零开始给网站配置 llms.txt：语法、目录结构、与 robots.txt 的分工，以及在客户网站上验证过的常见坑。
llms.txt 部署指南指南6 章 · 2026.06
ChatGPT 商品 feed 设置指南:被 AI 选中的数据规格
OpenAI 砍掉 Instant Checkout 后改押商品探索,让商家提交 feed 让商品被“完整呈现”。这份指南给你 ChatGPT 商品 feed 的格式、字段与刷新规格,把商品做成 AI 探索层能读懂并选中的数据。
ChatGPT 商品 feed 设置指南:被 AI 选中的数据规格拆解6 章 · 2026.06
Agent 结账轨道:AP2、Visa、Mastercard 的协议拆解
2026 年 agent 付款的基建到位了:Google 开放 AP2、Mastercard Agent Pay、Visa Trusted Agent,连 Visa 都接进了 ChatGPT。这份拆解帮你看懂这些协议的差异,以及商家该支持哪些、如何让信任信号变成机器可验证的。
Agent 结账轨道:AP2、Visa、Mastercard 的协议拆解用工具完成本阶段
AI Agent Strategy
绘制一个任务从发现到执行的路径，包括权限与失败状态。
证据边界
Use the output for the decision it describes; do not treat a technical scan, self-assessment, or planning model as proof of live AI citations.
应用学习
进入下一个可检查的决策
只在证据基础符合决策时使用工具、诊断或服务。