The one-sentence conclusion: this Suzhou manufacturer's overall mention rate across 12 AI engines was 50%, which looks like a passing result—but when broken down by engine, Doubao and DeepSeek, the two engines with the largest user bases, scored only 41% and 12% respectively, while DeepSeek had already begun recommending the company's competitors by name for comparable questions.

Background: The Anxiety Began with a Casual Conversation

The client is an electronic-components manufacturer in Suzhou. Its main products are network transformers and integrated RJ45 connectors, and its customers are primarily hardware engineers at communications-equipment and industrial-control manufacturers.

The concern began with something small. A purchasing engineer at a long-standing customer casually mentioned that, before selecting a product, he now asks DeepSeek, "How should I choose a gigabit network transformer?" The owner went back and tried it himself. The answer was highly professional and cited technical content from several peer companies, but it did not mention his company once.

This was not an isolated case. CNNIC's 57th report (February 2026) showed that China had 602 million generative-AI users, with a penetration rate of 42.8% and year-on-year growth of 141.7%. Data from a Zhihu Research Institute white paper, as reported by China News Service, was even more direct: 81% of consumers would partially adopt AI recommendations when making a purchase. Engineers in B2B procurement chains are among the earliest groups to turn "ask AI" into a professional habit.

Another figure is worth considering alongside those numbers: as of October 2025, about 70% of Baidu mobile search results pages contained AI-generated content, according to Baidu's Q3 earnings disclosure. Even when users are still "searching," more and more of the first screen is written by AI. An SEO ranking no longer automatically means being seen.

That raises the questions: what exactly does AI say about you? How often does it mention you? When it does not mention you, whom does it mention instead? Asking a few questions manually cannot answer them; a systematic checkup is required.

Checkup Method: 43 Questions, 12 Engines, 11 Days, 524 Valid Runs

We used YinJen to continuously monitor this client for 11 days (first measured run 2026-06-30, last 2026-07-10; the account itself was created on 06-21):

Question set: the client's question library holds 344 entries in total; 43 questions were actually deployed in this period, divided into two main categories—"manufacturer recommendation" questions, such as "recommended network-transformer manufacturers," and "selection knowledge" questions, such as "what is the difference between 100 Mbps and gigabit network transformers?";
Engine coverage: Doubao, DeepSeek, Kimi, Qwen, ERNIE Bot, Tencent Yuanbao, Zhipu Qingyan, StepFun, ChatGPT, Claude, Gemini, and Perplexity—12 engines in total;
Measurement method: automatically asked real questions of each engine every day and recorded the answers, producing 524 valid measured runs over 11 days (each question-engine pair ran 1.4 times on average, 4 at most; a further 236 runs failed and 82 were skipped, neither of which is counted);
Evaluation rule: an answer counted as one "mention" if it explicitly mentioned the brand or its official website; mention rate = number of mentions ÷ number of measured runs.

(The case data comes from monitoring authorized by the client and has been anonymized throughout.)

Finding 1: 50% Overall, but a Gap of More Than 20-Fold Between Engines

Across 524 measured runs the brand was named 263 times, an overall mention rate of 50.2%. Viewed alone, that number seems acceptable. Broken down by engine, however, the results describe two entirely different worlds. The 10 engines with usable sample sizes are listed below (the figures in parentheses are mentions over measured runs):

Qwen — mention rate: 85% (46/54)
Tencent Yuanbao — mention rate: 83% (38/46)
Claude — mention rate: 74% (39/53)
ERNIE Bot — mention rate: 67% (38/57)
Zhipu Qingyan — mention rate: 63% (34/54)
Kimi — mention rate: 53% (19/36)
Doubao — mention rate: 41% (24/59)
ChatGPT — mention rate: 17% (8/47)
DeepSeek — mention rate: 12% (7/57)
Gemini — mention rate: 4% (2/46)

Two further engines, StepFun (8 runs, 63%) and Perplexity (7 runs, 43%), had samples too small to compare and are listed here for completeness only.

The highest result was Qwen at 85%; the lowest was Gemini at 4%, a difference of roughly twenty-fold. That comparison deserves a caveat: Gemini's 4% is 2 hits out of 46 runs, so a single additional hit would shrink the ratio noticeably. What the contrast shows is that engines are not in the same order of magnitude—not a precise multiple. "AI visibility" is never one number; it is one number per engine. Looking only at the average hides the most consequential weaknesses.

Finding 2: The Weakest Results Were on the Largest Traffic Gateways

The significance becomes clear only when mention rates are considered alongside user volume. According to QuestMobile's 2026 Q1 data, native AI apps had 446 million monthly active users: Doubao had 345 million and DeepSeek had 127 million, making them the two largest gateways by user volume. Those are precisely the two engines on which this client scored only 41% and 12%.

Using the same measurement basis, the top-performing Qwen had 166 million monthly active users, less than half of Doubao's total. On the two engines with the broadest reach, the client was mentioned at most four times out of ten questions, and at the low end only a little more than once. Visibility needs to be weighted by traffic; once weighted, that 50% immediately shrinks.

Finding 3: Present for "Recommendations," Absent from "Knowledge"—Losing During Demand Formation

The contrast was even sharper when the results were split by question type:

"Manufacturer recommendation" questions: a hit rate of 54%–75%. When users directly asked "Which supplier is good?", the client often appeared on the candidate list;
"Selection knowledge" questions: only 1 hit across 19 questions, a hit rate of 5%.

Before engineers decide "which supplier to contact," they spend substantial time understanding "how to choose." During this demand-formation stage, AI was citing only other companies' technical content—AI was using competitors' content to educate this company's customers. By the time users reached the "manufacturer recommendation" stage with selection criteria learned from someone else, merely being present meant little more than making up the numbers.

Finding 4: DeepSeek Was Already Naming Your Competitors

The most striking record came from DeepSeek. For a comparable "manufacturer recommendation" question, it recommended other peer brands by name while omitting this brand.

Tracing the citation sources shows which pages the AI trusted. This record does not mean that "AI does not understand the industry." It means competitors' content had already entered DeepSeek's source pool while yours had not. A blank space does not remain blank forever; the only question is who fills it first.

Optimization Roadmap: Address Weaknesses One Engine at a Time

Different engines have different source structures, so the actions should differ as well. Our roadmap for this client had three steps:

Step 1, DeepSeek: open structured content on the official website. Turn product specification tables, selection guides, and application-scenario FAQs into clearly structured, crawlable pages so the engine at least has your content available to cite. This is the first priority for climbing above 12%.

Step 2, Doubao: Toutiao-ecosystem sources plus titles aligned with query intent. Newrank's empirical study of 474,000 Doubao citations (May 2026) provided clear guidance: citation likelihood had approximately zero correlation with an account's follower count. The only significant variable was the match between the title and the user's query intent (0.23). There is no need to spend heavily building a large account. Distribute content within the Toutiao ecosystem and phrase titles as the questions users actually ask—"How should I choose a PoE network transformer?" rather than "Our product receives another certification"—to align more closely with the citation mechanism.

Step 3, turn the 19 missed selection questions into content topics. Every unanswered "selection knowledge" question is a ready-made technical-content topic. Fill them one by one so AI can begin citing you during the demand-formation stage.

One boundary must be made explicit: these actions are fundamentally about "giving AI credible content it can cite." No one can guarantee indexing or rankings. In March 2026, CCTV's 315 Gala exposed the gray market for GEO "AI poisoning," in which just 11 advertorial articles could make AI recommend a fictitious product. That route is neither acceptable nor sustainable. There is only one compliant approach: keep producing genuine content, then use monitoring to verify whether each step works.

How to Run the Same Checkup for Your Own Brand

Every step in this checkup can be reproduced for your own brand:

Download YinJen for macOS or Windows: https://zhimahang.com/yinjen/download ;
Build your own question set. Think about how customers ask for "recommendations," but more importantly, how they ask "how to choose";
Enable automatic daily measurement and run it for two to three weeks to obtain per-engine mention rates, a Visibility Score, and comparisons of competitor share of voice;
Use citation-source tracing to identify whom AI currently trusts, then address weaknesses engine by engine using the roadmap above.

The 14-day free trial requires no credit card, and the Creator plan starts at ¥29.9/month. See https://zhimahang.com/yinjen for details. First understand your real position across 12 engines; then decide where to focus.

FAQ

Q: What is AI visibility (mention rate)? A: An explicit mention of a brand or its official website in an AI answer counts as one "mention." Mention rate = number of mentions ÷ number of measured runs, measuring how frequently a brand is seen in AI answers.

Q: How long should monitoring run, and how many questions are needed for useful results? A: This case covered 43 questions, 11 days, and 524 valid measured runs (the client's library holds 344 questions; only 43 were deployed in this period). We recommend covering both "manufacturer recommendation" and "selection knowledge" questions and measuring continuously for two to three weeks, because a single query is too susceptible to chance.

Q: Can content optimization guarantee an AI recommendation? A: No, and no one can make that guarantee. The "AI poisoning" gray market exposed by CCTV's 315 Gala demonstrates precisely the risk of shortcuts. What you can do is continue producing genuine content and use monitoring to verify the results.

An AI Visibility Checkup for a Suzhou Manufacturer: 524 Tests, Qwen at 85% and DeepSeek at 12%