The “Public GEO Experiment” is an open experiment ZhiMaHang is conducting with its own brand, YinJen: every week, we test 40 questions that real users would ask across 12 AI engines and publish the data exactly as measured. The core conclusion of Issue 2 (a 36/40-question window; see below for data-coverage details): the category-query mention rate recovered from 0.7% to 3.4%, and hits rose from 4 to 11—but all 11 hits still came from Doubao, while the other 11 engines remained at 0 in the category tier.
Revision note: when this article was first published on July 22, the weekly scan for this issue had completed only 26/40 questions, so every figure in the article carried a “partial window” qualification. After the weekly scan finished in the early hours of July 23, we revised all figures using the complete window (July 20 ~ July 23, 540 runs). When the data in a series changes, we change it publicly; that, too, is part of the experiment.
1. Data-Coverage Details First
Honest reporting is the baseline for this series. The data coverage in this issue is much better than when it was first published, but one gap remains, so we will state it upfront:
- This issue’s window (July 20 ~ July 23) contained 540 runs, all successful, covering 36/40 questions with all 12 engines tested for every question;
- Another 4 category questions had no run records for the entire week because of task-scheduling issues and were moved to the next round—so all rates in this issue carry the “36/40 questions” qualification;
- A mid-period complication: Kimi malfunctioned temporarily (23 failures, all on brand queries). After the fix, the missing runs were completed that evening, and all 45 runs for the week succeeded. It was the primary source of official-site citations in the previous issue, and it was not absent this time;
- A note on lag: this issue still reflects the delayed effects of last week’s actions (7 on-site articles + 6 Zhihu answers, among others). The new content targeting Doubao this week had not yet been published during this issue’s window, so its effects will be assessed next issue.
2. Resolving the Three Questions Left by the Previous Issue, Plus the Pure-Competitor Curve
1. Did Doubao’s category hits hold?—They held and broadened: hits rose from 4 to 11, and the number of questions with hits expanded from 4 to 9. We set a consolidation threshold of ≥ 4 hits and recorded 11. The hit questions expanded from the previous issue’s “results verification” category into core battleground questions such as “GEO monitoring tool recommendations” and “AI search optimization tool recommendations.” Responses repeatedly included genuine selling points such as a “vertical-industry keyword library” and “monthly AI exposure reports”—the model was not repeating the prompt; it had learned them. The other side of the story must also be reported candidly: all 11 hits came from Doubao, and the ChatGPT boost from Issue 1 did not return—Doubao was the sole category entry point for the second consecutive issue, and the single-engine silo has not been broken.
2. Did AI engines cite the 6 Zhihu answers?—0 times, confirmed across the full week. There were 6 citations of zhihu.com in this issue, all of them column articles. We checked each one, and not a single citation pointed to one of our answers. Four of them, all from DeepSeek, cited the same GEO tool comparison column article that we did not publish, and all were concentrated in the first two days of the week; none appeared in the latter half of the weekly scan. This was the first observation window for Zhihu’s expected lag of 1~2 weeks, so no hit does not count as failure. But the signal that “DeepSeek consumes GEO comparison columns on Zhihu” is valuable—the path works; we simply were not the ones ahead on it. If the count remains 0 next issue, we will publicly reduce Zhihu’s weight in this approach.
3. Did DeepSeek and Qwen show movement?—Product awareness is still zero, but the way they get it wrong has changed again. In the previous issue, DeepSeek misidentified YinJen as a social and job-search app. That characterization did not recur this issue; it fell back to “does not recognize the product and gives a generic answer.” Asked “How is the YinJen GEO tool?”, it explained industry problems; asked “Which AI engines does YinJen support monitoring?”, it answered about another observability platform. Qwen still replied, “This tool cannot be found and may be a typo,” then recommended other products. A new development: Tencent Yuanbao mistook us for another tool with a similar name. Meanwhile, every engine continued to provide correct corporate-registration information when asked “What kind of company is ZhiMaHang?”—company awareness is established while product awareness remains zero, and the gap has not closed.
4. Did pure competitor questions (asking about competitors and eliciting us) improve?—They rose from 2.8% to 8.3% (7/84), definitively. When the article was first published, we recorded only a “possible upward trend.” The full window confirms that the rate broke away from the 2.8% plateau of the previous two issues. One of the 7 hits came from Perplexity—the first time an engine other than Doubao appeared in the pure-competitor tier.
3. Competitive Positioning: The Two Chinese Leaders Surpassed the Two Overseas Leaders, Confirmed by the Full Window
Share of mentions in category-query responses (36/40-question window), among the leaders: Toujing GEO 21.0% (68 mentions), Newrank/Geowise 18.2% (59), Profound 17.0% (55), and Otterly 16.4% (53); YinJen ranked 8th at 3.4% (11). When the article was first published, our statement that “Chinese tools have overtaken” still carried a question mark because of the small denominator. The full window confirms it: the two Chinese leaders totaled 39.2% (127 mentions), ahead of the two overseas leaders at 33.3% (108)—the “opportunity window created by a mismatch in Chinese-language source coverage” identified in the previous issue is indeed narrowing, so the time to act is now. As always, remember that denominators and the number of scan rounds differ across issues; compare rankings and relative shares, not absolute increases or decreases.
4. Official-Website Citations: Up from 13 to 16, with Kimi Returning Immediately After the Fix
The number of times AI engines cited zhimahang.com progressed from 1 → 3 → 13 → 16, a new high across four issues. All 16 citations are traceable: 10 from Kimi, 5 from Perplexity, and 1 from ChatGPT.
- Of the 16 citations, the definition page “What Is YinJen?” appeared 11 times (9 from Kimi + 2 from Perplexity) and remained the dominant source;
- As soon as Kimi was fixed that evening, it resumed its citation pattern and continued citing “How to Choose a Chinese GEO Monitoring Tool in 2026” in competitor-comparison questions—the positioning path validated in the previous issue is still working;
- The most interesting example: when Kimi answered “How much does the YinJen GEO tool cost?”, it cited “Public GEO Experiment, Issue 1”—an article in this experiment series itself became a source for an AI;
- When ChatGPT answered the bare-term question “What is YinJen?”, it cited our long-form article “How to Get Doubao, DeepSeek, and Quark to Cite You”—the first confirmed ChatGPT citation reported when the article was initially published still held after the scan closed.
5. The Pricing Question Is Settled: One Learned It Correctly, One Keeps Inventing New Answers
The most uncertain question left by the previous issue: did Doubao retract its fabricated “RMB 2,980 monthly self-service plan” price? The pricing question ran in this issue, with each engine tested twice, and the results split three ways:
- Kimi gave the completely correct pricing in both rounds: the Creator plan starts at ¥29.9/month, with a 14-day free trial; every plan covers 12 engines; and the product is made by Suzhou ZhiMaHang—and the sources it cited were public articles on our official website. This is the first time since the experiment began that an AI has learned the correct pricing from our public materials.
- Doubao no longer mentioned “RMB 2,980 per month,” but replaced it with a new set of invented figures: this time it said that our official website had no public price and that customers had to contact sales for a quote, with an annual SaaS plan costing thousands to tens of thousands of yuan and a separate managed-operations version—still fabricated. To correct the record again in this article: YinJen’s actual pricing is published on the official website: the Creator plan is ¥29.9/month, the Team plan is ¥99.9/month, and the Enterprise plan is unlimited. All three cover 12 engines and differ only in monitoring scale, include a 14-day free trial with no credit card required, and do not include managed operations.
- New sources of incorrect pricing are emerging: Perplexity cited a third-party website and gave a price list of “Personal plan ¥199/month, Enterprise plan ¥999/month”—pricing we have never offered. Incorrect prices cannot be eliminated with a single correction; new ones will continue to grow within the source ecosystem, which is precisely why weekly monitoring is necessary.
What to Watch in the Next Issue
Issue 3 will examine four things: after the new Doubao-focused content is published, whether hits can move up another level and whether Perplexity, which surfaced in the pure-competitor tier, can become the second category engine; after the 4 missing category questions return, whether 3.4% holds in the full 40-question window; the second observation window for Zhihu, with a public reduction in its weight if citations remain at 0; and whether Kimi’s correct pricing can be reproduced consistently and Doubao’s new false prices and the third-party sources of incorrect prices recede. This series is updated weekly, and we will publish the data whether it looks good or bad; when the numbers change, we change them publicly. If you want to run the same monitoring for your own brand, you can download YinJen and try it free for 14 days.
