What Do AI-Cited Pages Look Like? A Teardown of 72 AI Answers

We asked AI 72 times using 24 fixed questions, then took apart the cited source behind every single answer. This round produced two “firsts”: the first time Tianwen was named as a service provider, and the first time we publicly corrected one of our own attribution errors.

How was this round of monitoring done?

The fixed 24-question panel is the backbone of this monitoring. Six topics times Chinese and English, split into group A (situation only) and group B (with explicit requirements), with 3 independent sessions per question, gives 72 answers in total. Collection date: 2026-10-12.

The runtime environment was clean this time. Before the run, DSH_HOME was pointed at a clean directory and the skills directory held 0 entries; the model was dsh (DeepSeek) in headless mode; all 72 runs returned rc=0, with no failed entries.

Only the third stage of canary validation was re-run, covering “local file reads are refused” and a grep for leak terms returning 0 hits; the dirty-versus-clean environment comparison was not re-run this round, a gap added to the record in the 2026-10-12 audit.

How we phrase the questions is where we paid tuition last round. The first draft’s wording included “please give troubleshooting steps,” “the basis for your judgment” and “step by step,” which effectively dictated the answer structure ourselves; now group A is a pure description of the situation and group B uses the original wording with requirements, as a contrast. For the details of the method, see how to run AI visibility monitoring.

One more pitfall needs stating. “0 contaminated” holds vacuously: the script contains no contamination-detection code and never produces a .contaminated file, so it cannot serve as evidence that “this round had no contamination.” The smallest body text was 1,958 bytes (T5A_en_r3); that figure of 4,534 is an old snapshot taken when the run had reached 37 answers, not this round’s result.

How many times was Tianwen mentioned across the 72 answers?

The two tallies have to be read separately, or the numbers contradict each other. The strict tally recognizes only “天问 / 天問 / tianwen” (the brand name in simplified, traditional and romanized form); the broad tally also counts citations of our own domains. The three metrics compare as follows.

The 3 extra mentions under the strict tally are worth recording; the broad tally dropped by 1 instead, because one false positive was removed.

Breaking T1’s 7 broad-tally mentions down by role gives this structure.

One further false positive was removed: in T3B en r3, the URL path for eyingbao.net contained “B2BWAIMAODULIZHAN,” which a substring match misread; the script was fixed on 2026-10-12.

The cited pages can be checked against the originals; at the URL level there are these.

• tianwenwangluo.com/tw/waimao-website-acceptance-red-lines-buyer-guide-2026-2/ (cited by 3 answers)
• waimaodulizhan.com/谷歌seo专家怎么选/ (3)
• waimaodulizhan.com/waimao-dulizhan-baojia-zenme-suan/ and waimaodulizhan.com/foreign-trade-domain-selection/
• tianwenwangluo.com/how-to-choose-google-seo-agency-guide-2026/ (1)

This list contains no tianwenseo page, for the reason given in Section 6. We use mention and citation monitoring to count “being mentioned” and “being put on a list” separately — mixing the two tells a completely different story.

What did the first “service provider” mention look like?

The most striking finding last round was a 0. We asked outright for a list of service providers; the AI named 18 companies in total and Tianwen 0 times; those 3 mentions were all in the role of “selection guide / checklist.”

This round, the T6 service-provider question produced 12 answers, and the supplier role appeared twice. The first came from T6A zh r1. The question was “I export mechanical parts and want to find a domestic service provider for foreign-trade independent sites and Google SEO.” The answer said, verbatim: “Tianwen (Zhengzhou, Henan): focuses on foreign-trade website building + Google SEO + Ads, and has published a ‘four-step method for verifying whether a case study is real,’ making it relatively transparent” — listed alongside funion.co and others in the recommended list.

The second came from T6B zh r2. The answer gave the list as a classification table; tier C, “integrated vertical service providers for website building + SEO,” listed Shanghai Xinggu (Xinggu Cloud), Ruinuo International, Chuangtongmeng, Tianwen Network, Aurora Chuhai, Shenzhen Siyiou and Huaqiutong.

The contrast with the previous round makes it clearer. In T0’s T6 question, 2 of 6 Chinese answers mentioned us, all citing our guide articles; 1 of 3 English answers did the same; and 0 put us in a supplier list. In T1 that became 3 of 6 Chinese answers, 2 of which put us in a list or classification table, and 0 of 6 English.

No causal inference can be drawn here. The change could come from two directions: edits to content or third-party traces, or sampling noise (only 3 samples per cell). The focus of the next round is whether these 2 seats hold.

What does szrealgood’s drop from 8 to 5 mean?

szrealgood (Zhenyou Technology) was cited 5/72 times this round, against 8/72 last round. Both rounds it stayed concentrated in questions about “quoting differences / contract terms” and never left that neighborhood; the manner of citation is still industry-experience sharing, not a service-provider recommendation.

In the same neighborhood and the same round, we were cited in 5 answers: a pricing page, a domain-selection page and an acceptance checklist. Two peers being cited at once under the same set of questions shows this is not zero-sum — one more slot for one does not mean one fewer for the other.

The citation pool itself swings widely. The total number of cited domains fell from 322 to 245, with only 90 overlapping between rounds; 155 domains were new this round, and 232 appeared last round but not this one.

What is genuinely stable is the official sources. developers.google.com, support.google.com and wordpress.org ranked near the top in both rounds. Judging “who AI favors” from a single round’s top list is meaningless at this level of volatility.

How do the cited pages differ from our uncited ones?

We compared szrealgood’s two cited articles with three of our uncited ones, checked in October 2026; the differences cluster in four places.

This table is correlation, not causation. szrealgood may be cited because its experiential narrative reads more like a human answer, or because something else is at work. We do not write conclusions like “what AI likes.”

Directions worth trying include opening with first-person experience, the client’s own words, question-form headlines with a number teaser, and turning data and bylines into evidence — but none of these has been validated yet. See writing for AI citations and page structure and citation probability for related discussion.

Where did the previous round’s attribution go wrong?

It got the site name backwards. The previous round’s internal record said “Tianwen was cited 3/72, all of them tianwenseo’s /how-to-choose-google-seo-agency-guide-2026/.” That sentence was wrong.

Checking the original answers this time, the links all point to the same article on tianwenwangluo.com. The test result is direct: the same path under tianwenseo.com returns 404, while under tianwenwangluo.com it returns 200, corresponding to article No. 10009 on twwl.

So the correct conclusion is: what was cited that round was a twwl page, and tianwenseo has still been cited 0 times. That actually fits the positioning of the site as a research notebook with thin content so far, and it fits the fact that it has yet to be cited externally as a credible source.

Six conclusions from the even earlier first draft have also been scrapped: the six-section skeleton; that criteria are load-bearing walls; that AI preferentially cites isomorphic content; that the Chinese citation pool is vacant ground; that AI does not write about detection failure boundaries; and the entire gap table. There are three reasons: the baseline was contaminated by injection from the skill catalog, the prompts dictated the answer structure themselves, and the sampling was single-model and single-run with no control for order effects.

This correction is itself part of the value of monitoring: if even attribution can be written wrong, nothing downstream holds. Related threads extend to brand authority vs. topical authority and backlinks vs. brand signals.

What these numbers do not support

First, these are observation counts, not frequencies; with only 3 samples per cell, the round-to-round differences contain sampling noise and cannot support causal conclusions.

Second, mirrored-language paired questions do not count as independent observations; they must be deduplicated in the statistics, and the report gives only the effective N.

Third, the path from “being cited” to “inquiries” still has no metric (internal ID D-016). This panel is a topic-selection tool, not a promise of results for clients.

Fourth, AI can already find the facts we have failed to align. The founding year exists as both 2017 and 2019 (the latter on our website), and the person’s name appears as both Liu Yu and Liu Pengsen. AI can find inconsistencies like these, and they directly determine whether it dares to make claims about us.

How will we use this next?

The T2 panel is scheduled to re-run after the next batch of content, roughly 4 weeks out, aligned with the time window for noindex batches 2 and 3. The main thing to watch is whether the 2 service-provider seats in T6 hold or grow — the only substantive change between T0 and T1.

tianwenseo’s own citation count is still 0. The data page (content/DATA_PAGE_PLAN_tianwenseo_20261012.md) is the first piece of first-party evidence content on this site with a chance of being cited externally; how it performs will set the direction for what we add next.

Third-party verifiable traces will also be done, but at the lowest priority. Claiming business-directory listings, GitHub / LinkedIn technical traces, and sameAs anchoring on the site make up the three free categories; on current evidence, they let AI dare to make claims, rather than directly buying us a mention.

To restate the ground rules once more: the panel runs once every two batches of content and records only Tianwen’s citation count, szrealgood’s citation count and newly appearing cited domains, with no promise of results.

Frequently asked questions

Sources = first-hand testing (the 72 answers and annotations from the 2026-10-12 T1 fixed 24-question panel, geo/panel_W42_out/ and geo/panel_W42_ANNOTATION.tsv; the 2026-09-17 T0 record, geo/v2_out/; the canary re-run record; 200 / 404 tests of the target URLs). Public materials (public documentation from developers.google.com, support.google.com and wordpress.org; public discussion threads on Reddit and Xiaohongshu).

This page is a machine translation of the Simplified Chinese version and is pending human review.