What exactly did we compare?

This note comes from the GEO panel and page comparison of the same project. Baseline round T0 was September 17, 2026; monitoring round T1 was October 12. Prompts were fixed by topic, language, and phrasing, and each round recorded 72 answers. That design lets me go back to the original sentences instead of picking a flattering screenshot and telling a story.

Strict brand mentions in T0 were 3/72, and the guide page involved sits on tianwenwangluo.com, not on this research site. The answer treated it as a selection reference; Tianwen did not appear on the supplier list. Attribution has to be stated clearly: the research site organized the evidence, which does not mean this site itself earned those citations.

In the same round, a peer site, szrealgood.com, was cited in 8/72 answers, mostly around pricing differences and contract terms. I compared its cited short articles with our own articles; what I saw were differences in form and position, not the search system’s internal reasons for choosing it. The detailed research entry point is in our field-tested research and method notes, so that editorial judgments and data records can be read separately.

What visible differences separate the list pages from the long-form articles?

The cited short articles are shorter and rely less on large tables. In the original comparison, the two peer articles ran about 1.5k words, while our comparison pages ran about 13k, 13k, and 3.8k words. These are only features of this particular set of pages; they cannot stand in for all Chinese-language SEO content.

I care more about one concrete question: after reading this far, can an export business owner immediately know what materials to ask a service provider for? A sentence like “technical capability should be assessed comprehensively” is perfectly correct, but you cannot take it to an acceptance review; a sentence like “please open a real project and explain the problem before the changes, the changes made, and the records afterward” at least gives you something to check.

This is the same approach I used when reviewing how to write answer paragraphs that can be cited. Put the conclusion, the conditions under which it applies, and the evidence in the same paragraph; readers then do not have to finish the whole article to understand which situation a sentence actually covers.

Was it the answer that got cited, or the service capability?

A guide being cited does not mean the company is being recommended. What the T0 answers endorsed was information of the kind “how to choose a service provider.” I cannot, just because our own domain appears at the end of an answer, write that “AI has endorsed our delivery capability.” When a buyer asks for a list, an answer may cite a screening guide without ever listing its author as a candidate.

T1 showed a different identity: of the 12 answers to the service-provider question, 2 placed Tianwen on a list or in a category table. That is worth recording. Like strict mentions moving from 3/72 to 6/72, it is an observed change in the answers, not a verified closed deal.

I record “mention in the body text,” “source link,” “provider list,” and “actual contact” separately. The reason is practical: the first few can appear together, or they can be entirely separate. If you mash them into a single visibility score, an owner cannot tell what the budget actually bought. You can use our monitoring definitions for brand mentions and citations to keep the original text for each item.

Why can’t we declare that we have found the citation formula?

Between “short articles were cited and long articles were not” and “short articles caused the citations” lies a credible causal comparison that we still lack. Domain history, prompt match, author information, page accessibility, and citations of the page by other sites could all explain the result. I did not hold these conditions constant.

Across the two rounds, the total number of cited domains went from 322 to 245, with only 90 overlapping. That reminds me that the same question can produce a very different set of sources. Look at a single screenshot, and you will mistake random fluctuation for a writing technique.

The environment check also has a gap. T1 ran in a clean environment, but that round only re-ran the local-read-refusal portion of the canary validation; the full dirty-versus-clean comparison was not repeated. A successful run does not by itself prove the absence of context contamination. I would rather keep that limitation than write a neat but overreaching conclusion.

Nor does “a page without JSON-LD was cited” prove that “JSON-LD is useless.” It only refutes the absolute claim that “without this markup, a page can never be cited.” When I need to check page status, I start with the technical SEO audit checklist and then record technical conditions and citation results separately.

How do you turn a purchasing question into a checklist that can be verified?

My starting point for rewriting is the decision the buyer is actually making, not the keywords I plan to stuff in. For example, “how to choose an SEO provider” can land on case attribution, data access, scope of work, acceptance criteria, and exit handover. Every item should let the reader take an action or rule out a misunderstanding.

What should follow a single judgment?

Evidence and boundaries. If you say “do not choose a team on the strength of a traffic screenshot,” then explain what to ask for instead: the project domain, the measurement period, when the provider’s involvement started and ended, the query and page composition, and whether inquiries were verified by sales. That way readers know how to verify, rather than having to accept the author’s attitude.

I would not make “we have been in business for many years,” “we write a lot of articles,” or “we buy expensive tools” core items. Those may have reference value, but they do not directly address the buyer’s risk. Put SEO provider risk signals and the scope-of-service and contract acceptance checklist side by side to find which promises have no matching deliverable.

What should long-form articles still keep?

Mechanisms that need explaining, counterexamples, and complex conditions should stay. A checklist is good for helping someone decide; a long-form article is good for explaining the reasons behind the decision. My approach is to put the main answer early and use the rest of the space for the conditions under which it applies. I will not cut the limitations to make the piece shorter, nor fill it with jargon to look expert.

What evidence should we keep next time?

Keep the original question sentence and the full answer first, then mark where the brand appears and which page a source link points to. Language, model, and sampling time also need to be recorded. Only repeated observations of the same question make comparison meaningful; once the question changes mid-stream, new results cannot be forced onto the old curve.

Source identification also means going back to the full URL. This round, similar words contained in another site’s path were initially misread as citations of our own brand, and were removed only later. Checking only whether a string appears makes it easy to credit someone else’s page to yourself; domain, path, and in-text identity should be verified together.

Content change logs also need to be specific. You cannot just write “optimized GEO”; you have to state which page changed in what way, what old content was kept, and what evidence could overturn the current judgment. If citations appear afterward, the most I will say at first is “citations were observed after the change,” not immediately write the sequence as cause and effect.

For the next comparison, I will try to choose pages that answer the same kind of purchasing question, rather than comparing a contract checklist directly against a product encyclopedia. Record the original passages, citation status, and page conditions first, then state what is going to change. That way, even if causality still cannot be established, we can at least rule out the explanation that “the things compared were fundamentally different,” and avoid forever chasing whatever writing style showed up once.

Counterexamples must be kept too. A list that was not cited, or a long-form article cited later, both weaken the claim that “one format is inherently better.” If you keep only the answers that support your judgment and then use them to guide all your writing, the experiment easily turns into self-affirmation. To me, discovering that an earlier judgment does not hold is also a result this research should deliver.

Business records need a separate column. Whether anyone arrived from a relevant source, whether they sent an inquiry, and whether that inquiry fits the product and purchasing conditions all need their own evidence. Our internal research has not yet connected the verification metric from “cited” to “inquiry.” For an export business owner, that gap matters more to the investment decision than any single mention.

What I want to test is not “whether the list wins” but “whether this page makes a real purchasing decision easier to check.” Citations are only one observation window. If a reader can use the checklist to vet suppliers, and if the conclusions hold up when traced back to the original text, then this piece is worth maintaining.

What are the questions owners ask most often?

Is a list-style article always more likely to be cited by AI than a long-form article?

No such conclusion can be drawn. We compared only a limited set of pages, with no random assignment, and the domain, the topic, and existing citation records could all affect the result. The list format is worth testing, but there is no proof that it always wins.

Can a page without structured data still be cited?

In this comparison, the cited peer pages had no JSON-LD. That means citations in this sample did not require JSON-LD; it does not follow that structured data has no effect.

If a name appears on a provider list, does that count as getting an inquiry?

No. A name entering a list only proves that the answer presented a provider identity; real visits, contact records, and sales verification are still needed to judge whether it produced a business lead.

Should we delete all our long-form articles and replace them with short checklists?

Rewriting the whole site on the strength of this one observation is not advisable. Pick one real purchasing question, keep the existing content that works, build a checklist that has sources and can answer the question on its own, and then record the cited text and what happens next.