Revisiting GEO Reports Today: Why Do I Now Ask an Extra Question?
This morning, I reviewed several AI search updates from July again, then went back to GEO reports, and paused on one question: What we see — is it the entire process, or just the result left after the answer is generated?
Whether a brand appears, how it's described, and who's being cited all need attention. But if AI calls search and filters web pages before answering, only looking at the chat box means the diagnosis might miss a layer.
Search entry points are shifting from pages to APIs.
Why Is This Change Not Just About the Chat Interface?
On July 27, 2026, Tencent Cloud's developer page introduced the "Web Search MCP (Standard Version)". The official description mentions it's built on the Web Search API, involving indexing, retrieval, and ranking, and can return data fields.
Baidu Qianfan's "Intelligent Search Generation" documentation was updated on July 14, describing a process that first searches real-time web information, then summarizes it with a model, supporting API, tools, MCP, SDK, and other invocation methods. Note: This refers to a documentation update, not the initial release of the capability.
Two examples don't mean all AI products use the same mechanism. The common signal is: search is extending into a capability that Agents can call.
The answer layer is just the result, not the whole process.
In the Yahoo Era, What Were We Competing For?
I started working in SEO at a Hong Kong digital marketing company in 2005. In the early days, Yahoo, portals, and directories were still important; companies competed to be part of the entry points where users found information.
What the Google Era Changed Went Beyond Rankings
In the Google era, businesses understood competition around the SERP. Indexing, relevance, page structure, content quality, links — these factors influenced search visibility. Brands started competing for "which pages could enter the candidate pool".
In the Agent Era, What's the New Layer of Competition for Businesses?
Previously, users saw search results first, then decided what to click. Agents can now search first, and then compile answers for the user.
| Era | Typical Entry Point | What Users See | What Brands Focus On |
|---|---|---|---|
| Portal/Directory | Directories, portals | Websites and categories | Whether they enter the discovery scope |
| SERP | Search engine results pages | Rankings, snippets, links | Indexing, ranking, clicks |
| Agent | Search API, MCP, retrieval tools | Answers, partial citations | Whether it's retrieved, what information enters the context |
Many people understand GEO as "competing for a spot in AI answers". This is only half right. In some Agent scenarios, there are steps like retrieval, filtering, and ranking before answering; implementations vary by product, and it can't be generalized as uniform "internal AI rules".
What is the "Machine Retrieval Candidate Set"?
I use "machine retrieval candidate set" to describe: the set of web pages, snippets, or external information sources obtained via search capabilities before an Agent or model forms an answer. It's not an official, publicly recognized technical term across platforms.
- Was it retrieved?
- What content did the machine get?
- Did it later make it into the answer or citations?
Being answered doesn't mean being retrieved; being retrieved doesn't mean being used.
Why Should GEO Monitoring Look One Step Beyond the Answer Layer?
Brand mentions, cited sources, and answer changes still need monitoring. But I'd also ask: Did the official website or third-party factual sources appear in the relevant search results? Did the machine get information from the official site, media, or outdated info? If the brand didn't make it into the answer, is it more of a representation issue, or a retrieval visibility issue?
When my team and I proposed AIPO, we emphasized connecting the dots across the entire process. Today, we still need to combine performance monitoring, data analysis, website optimization, content strategy, and platform publishing.
Here's the Question: Can Businesses Actually See the Agent's Retrieval Layer?
Some might ask: users can't see these processes, so how can they be monitored? This is a fair question. But it depends on the scenario.
- For self-built Agents or some API scenarios, you can see fields like search results, URLs, and cited content.
- For Agents controlled by the enterprise, you can correlate and analyze call logs, retrieval results, and answers.
- For public AI products, you can't claim to know the entire internal candidate set; you can only diagnose indirectly using external evidence like public citations, answer changes, and site tests.
Not being able to see doesn't mean you can pretend you can; nor does it mean this structure isn't worth studying.
If Monitoring Boundaries Shift Earlier, What Should Businesses Check Now?
- Discoverable: Can brand, product, service, and factual pages be normally discovered.
- Understandable: Are entity relationships, product information, time, location, and authors clear? Use Schema to assist when necessary.
- Verifiable: Do key business facts have public sources, dates, and consistent descriptions?
- Updatable: After information changes, do old pages still create conflicts?
This isn't about pursuing so-called "MCP optimization". What's more valuable is ensuring the same facts remain consistent, verifiable, and retrievable across public sources.
Where Should Different Roles Focus Their Attention Now?
If you're a CMO or digital marketing lead, you don't need to overturn existing GEO KPIs. First, check if your reports only stop at the answer layer, then ask: if the brand didn't get into the answer, did the problem happen before retrieval or after generation?
If you're an SEO or GEO lead, you can document the correspondence between "search results — cited sources — answers", and don't treat SEO and GEO as two isolated workflows.
If you're a GEO industry practitioner, you can expand "prompt → answer" to "prompt → retrieval evidence → answer", and mark unobservable steps as "unknown" or "inferred".
When search becomes Agent infrastructure, GEO is no longer just about whether a brand appears in the answer, but whether brand information enters the scope that machines can retrieve, understand, and verify.
Related Questions
What exactly is the relationship between MCP and GEO?
MCP is not a GEO method in itself. Since web search can be called by Agents, GEO also needs to pay attention to the retrieval process before the answer.
In the future, do we no longer need to monitor who AI cites for GEO?
No. Brand mentions, answer performance, and cited sources remain important; we just can't only look at the end result.
What does "machine retrieval candidate set" mean?
It's an analytical concept from this article, referring to a set of external information sources obtained via search capabilities before an Agent or model generates an answer.
Can businesses directly know if AI has retrieved them?
It depends. In self-built Agent or some API environments, it might be visible; for public AI products, it's usually only possible to judge indirectly.
Which AI platforms should be monitored for GEO in China now?
It depends on your business. Observational targets can include Doubao, Kimi, ERNIE Bot, Tongyi Qianwen, Quark AI, and Yuanbao.
Now that AI search is becoming Agent infrastructure, does SEO still matter?
Yes. Discoverability, information clarity, and factual consistency remain crucial when Agents call upon search capabilities.
If you're responsible for marketing decisions, first see if your reports combine answers, sources, and website discoverability; if you're responsible for execution, you can start with a round of GEO Audit to establish a diagnostic baseline.