Is GEO only in our heads, or something real?
Is GEO only in our heads, or something real?
GEO（生成式搜索引擎优化）究竟是真实存在的，还是我们的一厢情愿？
I was deep in thought about how much of a model hallucination is RLHF-based assumption in the absence of training data or context. Pretraining gives the model the ability to confabulate. Post-training often influences whether it chooses to confabulate rather than say “I don’t know.” 我一直在深入思考：在缺乏训练数据或上下文的情况下，模型幻觉中有多少比例是基于 RLHF（人类反馈强化学习）的假设？预训练赋予了模型“胡编乱造”的能力，而训练后的微调（Post-training）往往决定了它是选择继续胡编，还是坦诚地说“我不知道”。
A base language model is trained to predict plausible continuations. If the evidence needed to answer is absent from its weights or context, there’s no fundamental mechanism in next-token prediction that says “stop, evidence unavailable.” It still has a probability distribution over tokens, so it generates the most plausible continuation. 基础语言模型被训练用于预测合理的后续内容。如果回答所需的证据在其权重或上下文中缺失，那么在“预测下一个 token”的机制中，并没有一个根本性的指令告诉它“停止，证据不可用”。它依然拥有一个关于 token 的概率分布，因此它会生成最合理的续写。
Hallucinations therefore exist without RLHF. TruthfulQA demonstrated this with pre-RLHF base models such as GPT-3, GPT-2 and GPT-Neo/J: they confidently reproduced misconceptions learned from text, and the larger models were less truthful, not more. 因此，即使没有 RLHF，幻觉依然存在。TruthfulQA 通过 GPT-3、GPT-2 和 GPT-Neo/J 等未经过 RLHF 的基础模型证明了这一点：它们自信地复述了从文本中学到的错误观念，而且模型越大，真实性反而越低。
But this is a slightly narrow point of view for web people like us. For us, when we think of hallucination we think of a model telling an untruth. Which the above is really a subset of. Seen in the broader sense, hallucination decomposes like this: 但对于我们这些从事互联网行业的人来说，这个观点略显狭隘。对我们而言，当我们想到幻觉时，我们想到的是模型在撒谎。而上述情况实际上只是幻觉的一个子集。从更广义的角度来看，幻觉可以分解为以下几种情况：
- Wrong information encoded during pretraining. The model learned a false claim because the corpus contained it, or because statistical compression distorted the underlying facts. 预训练期间编码的错误信息。 模型学到了错误的声明，因为语料库中包含了这些内容，或者统计压缩扭曲了底层事实。
- Imperfect retrieval from model weights. The correct information is effectively “in there,” but generation retrieves the wrong association: wrong year, wrong person, conflated events. 从模型权重中检索不完美。 正确的信息实际上“在里面”，但生成过程检索到了错误的关联：错误的年份、错误的人、混淆的事件。
- Inference errors. The premises are available, the model reasons incorrectly. 推理错误。 前提条件是可用的，但模型推理过程不正确。
- Confabulation under missing information. The model lacks the fact and constructs the most plausible answer anyway. 信息缺失下的胡编乱造。 模型缺乏事实依据，但依然构建出了最合理的答案。
- Post-training induced behaviour. SFT, RLHF and preference tuning make 1 to 4 more or less likely to be verbalised as an assertion, depending on whether the learned behaviour rewards answering, agreeing, sounding decisive, abstaining or checking uncertainty. 训练后诱导的行为。 SFT（监督微调）、RLHF 和偏好调整使得上述 1 到 4 点更有可能（或更少可能）被表述为断言，这取决于学习到的行为是奖励回答、赞同、表现果断，还是奖励克制或核实不确定性。
I know what you’re thinking, this is still narrow. We don’t interact with bare models like this. We interact with fully fledged ecosystems with so many moving parts: model, system prompt, memory, retrieval, tools, planners, execution loop, verifiers, reward shaping, UI constraints, external state. What fancy people call harnesses. 我知道你在想什么，这依然很狭隘。我们并不是在与这种裸模型交互，我们是在与拥有众多组件的成熟生态系统交互：模型、系统提示词、记忆、检索、工具、规划器、执行循环、验证器、奖励塑造、UI 限制、外部状态。这就是那些时髦人士所说的“工具链（harnesses）”。
In that environment an untrue response can arrive for a much wider variety of reasons: 在这种环境下，产生不真实回答的原因要广泛得多：
- The base model invents a fact. 基础模型编造事实。
- Retrieval returns the wrong document, stale information, or a semantically similar but incorrect result. 检索返回了错误的文档、过时的信息，或者语义相似但内容错误的结果。
- The model misreads correct retrieved evidence. 模型误读了正确检索到的证据。
- Tool selection is wrong: it searches the web when it should query an MCP server, queries the wrong database, or fails to use a tool at all. 工具选择错误：本该查询 MCP 服务器时却去搜索网页，查询了错误的数据库，或者根本没使用工具。
- The planner makes an unsupported assumption early, and every later step treats it as established fact. 规划器在早期做出了无根据的假设，后续每一步都将其视为既定事实。
- Memory injects an incorrect prior fact. 记忆注入了错误的先验事实。
- A tool succeeds partially, but the agent interprets partial output as complete. 工具部分成功，但智能体将部分输出解释为完整结果。
- The agent loses provenance across steps and can no longer distinguish observed facts from inferred facts. 智能体在步骤间丢失了来源信息，无法区分观察到的事实和推断的事实。
- A verifier checks linguistic consistency rather than correspondence with reality. 验证器检查的是语言一致性，而非与现实的对应关系。
- The system prompt or reward function pressures the agent to “finish the task,” so it bridges missing state with assumptions. 系统提示词或奖励函数迫使智能体“完成任务”，因此它用假设来填补缺失的状态。
- One agent invents something and another agent treats that output as evidence. 一个智能体编造了内容，而另一个智能体将其视为证据。
- The final response generator turns an uncertain intermediate state into a declarative statement. 最终的响应生成器将不确定的中间状态转化为陈述性语句。
Why even think about any of this? Well, because if you’re to believe GEO is a real discipline, you have to believe you can impact some of these factors in a controlled way as an outsider. Otherwise there’ll be nothing to optimise, right? 为什么要考虑这些？嗯，因为如果你相信 GEO（生成式搜索引擎优化）是一门真正的学科，你就必须相信作为局外人，你可以以可控的方式影响其中一些因素。否则，就没有什么可优化的了，对吧？
The important distinction is control over the model versus control over the evidence environment the model operates in. What can we do then? Leverage is qualified the same way throughout: where a remediation requires you to operate a machine endpoint, the leverage is conditional on your doing so. Rows 4, 7 and 8 all carry that condition, and an earlier version of this table applied it only to row 4. 重要的区别在于：是对模型本身的控制，还是对模型运行的证据环境的控制。那么我们能做什么呢？杠杆作用的定义是一致的：如果补救措施需要你操作机器端点，那么杠杆作用的前提就是你必须这样做。第 4、7 和 8 行都带有此条件，而该表的早期版本仅将此条件应用于第 4 行。
| # | Failure | Business leverage | Evidence | What the business can actually do |
|---|---|---|---|---|
| 1 | Base model invents a fact | Low | Strong that it happens; weak that businesses can fix weights | Improve the public factual corpus; correct authoritative third parties; allow training crawlers if you want to. Don’t promise that publishing something changes model weights. |
| 2 | Wrong or stale retrieval | Very high | Strong | Crawlability, indexing, canonical URLs, freshness signals, authoritative pages, remove stale duplicates, correct third-party sources |
| 3 | Model misreads correct evidence | High | Strong to moderate | Make facts explicit, atomic, unambiguous, scoped, easy to extract. Reduce contradictory and verbose pages. |
| 4 | Wrong tool selected | High, if you expose tools | Strong and growing | MCP and API discovery, precise tool descriptions, narrow responsibilities, obvious capability names, good schemas |
| 5 | Planner invents assumption | Medium-high | Strong underlying evidence | Publish direct answers to likely ambiguous questions, and explicit negative facts, so retrieval resolves rather than infers |
| 6 | Incorrect memory | Low externally | Mixed | Make the current authoritative state easy to retrieve so live evidence can override memory. You can’t normally edit another provider’s memory. |
| 7 | Partial tool result treated as complete | Very high, if you expose tools | Engineering evidence stronger than GEO research | Design API and MCP responses with pagination, completeness, timestamps, status, errors, coverage boundaries |
| 8 | Provenance disappears | Medium-high, if you return machine-readable facts | Strong problem evidence | Attach source URLs, IDs, timestamps and claim provenance to every fact you return |
| 9 | Bad verifier | Low | Strong that verification design matters | Provide ground-truth endpoints a verifier could use. You don’t control whether third parties use them. |
| 10 | Reward or system prompt encourages guessing | Very low | Strong phenomenon, outside business control | Almost nothing externally, except make uncertainty unnecessary |
| # | 故障类型 | 商业杠杆 | 证据 | 企业实际能做的事 |
|---|---|---|---|---|
| 1 | 基础模型编造事实 | 低 | 发生概率高；企业修复权重的能力弱 | 改善公共事实语料库；纠正权威第三方；如果愿意，允许训练爬虫。不要承诺发布内容就能改变模型权重。 |
| 2 | 检索错误或过时 | 极高 | 强 | 提升可爬取性、索引、规范 URL、新鲜度信号、权威页面，移除过时重复项，纠正第三方来源 |
| 3 | 模型误读正确证据 | 高 | 强到中等 | 使事实明确、原子化、无歧义、有范围、易提取。减少矛盾和冗长的页面。 |
| 4 | 工具选择错误 | 高（若你暴露了工具） | 强且在增长 | MCP 和 API 发现机制、精确的工具描述、狭窄的职责范围、明显的名称、良好的 Schema |
| 5 | 规划器编造假设 | 中高 | 强底层证据 | 针对可能存在歧义的问题发布直接答案，并明确负面事实，以便检索能直接解决而非推断 |
| 6 | 记忆错误 | 外部低 | 混合 | 使当前的权威状态易于检索，以便实时证据能覆盖记忆。通常无法编辑其他提供商的记忆。 |
| 7 | 工具部分结果被视为完整 | 极高（若你暴露了工具） | 工程证据强于 GEO 研究 | 设计带有分页、完整性、时间戳、状态、错误、覆盖边界的 API 和 MCP 响应 |
| 8 | 来源信息丢失 | 中高（若你返回机器可读事实） | 强问题证据 | 为你返回的每一个事实附加源 URL、ID、时间戳和声明来源 |
| 9 | 验证器糟糕 | 低 | 验证设计至关重要 | 提供验证器可以使用的地面真值（ground-truth）端点。你无法控制第三方是否使用它们。 |
| 10 | 奖励或系统提示词鼓励猜测 | 极低 | 强现象，超出商业控制 | 外部几乎无能为力，除非让不确定性变得没必要 |