Bottom line: take 10 questions your customers would ask AI and test them one by one in Doubao, DeepSeek, and Qwen. Count how many times your brand is named and divide that by the total number of questions asked; the result is your AI mention rate. This article provides a free four-step manual method and a tracking-table template you can copy directly, plus a tool-based option for when manual measurement becomes unsustainable.
Why You Must Measure AI Mention Rate Now
CNNIC’s 57th report (2026-02): the number of generative AI users in China has reached 602 million, with a penetration rate of 42.8% and year-over-year growth of 141.7%. A Zhihu Research Institute white paper (as reported by China News Service) shows that 81% of consumers will partly follow AI recommendations when making a purchase.
In other words, your customers are now asking AI questions such as “Which XX company is best?” and “How should I choose XX?” Whether you appear in the answer—and whether your competitors are named—directly affects who receives the inquiry. Search engines are about rankings; AI Q&A is about mentions. That is the core metric of a brand’s AI visibility: mention rate.
There is also a clear priority order for which engines to measure. According to QuestMobile 2026Q1 data, Doubao has 345 million monthly active users, Qwen 166 million, and DeepSeek 127 million. Start with these three and you will cover China’s most mainstream AI Q&A entry points.
Free Manual Method: Measure AI Mention Rate in Four Steps
Step 1: Design 10 questions that customers would ask. Do not ask, “Do you know the XX brand?”—customers do not ask that. Design them in three categories: manufacturer-recommendation questions (“Which network transformer manufacturer is best?”), product-selection knowledge questions (“How should I choose a gigabit network transformer?”), and comparison questions (“Which is better for industrial scenarios, Solution A or Solution B?”). Newrank’s empirical study of 474,000 citation instances (2026-05) found that which content Doubao cites has a correlation of approximately 0 with an account’s follower count; the only significant variable was how well the title matched the intent of the question (0.23). The questions must therefore reflect how customers really ask them, or the measurement will not be meaningful.
Step 2: Ask each engine. Cover at least Doubao, DeepSeek, and Qwen; add Kimi and ERNIE Bot if you have the capacity. Start a new conversation for every question, with no follow-up questions and no steering. If the account has been linked to brand information, test with a clean account.
Step 3: Record whether you were named. Count a hit only when AI mentions your brand name on its own. Also record which competitors it named and which sources it cited. Tracking-table template (copy and use directly):
| No. | Question | Type | Engine | Date | Named us | Named competitors | Citation source |
|---|
| 1 | Which network transformer manufacturer is best? | Manufacturer recommendation | Doubao | 2026-07-12 | Yes | Brand A | An industry website |
| 2 | How should I choose a gigabit network transformer? | Product-selection knowledge | DeepSeek | 2026-07-12 | No | Brand B, Brand C | A Zhihu answer |
Step 4: Calculate the mention rate. Mention rate = number of times named ÷ total number of questions asked. 10 questions × 3 engines = 30 tests. In addition to the overall mention rate, always calculate it separately by engine and by question type—the structure is more informative than the total.
Three Pitfalls of the Manual Method
Pitfall 1: The sample is too small. Across 30 tests, a change of 1 result is already 3 percentage points, and you cannot tell whether it reflects the true level or luck. To draw a conclusion, the sample must reach the hundreds.
Pitfall 2: AI answers fluctuate randomly. With the same question and the same engine, you may be named today and omitted tomorrow. Large-model generation is inherently random, so a single result on a single day is only a reference point; tests must be repeated daily and averaged.
Pitfall 3: It cannot track continuously. Manual measurement produces a snapshot. Did the mention rate rise after you published new content? Did a competitor overtake you? Those questions require trends. When the data is maintained manually in Excel, most teams cannot keep it up for more than a month.
One-Click Tool Version: How YinJen Solves These Three Pitfalls
Disclosure: YinJen is a product developed by us (ZhiMaHang). The four-step manual method above does not depend on any tool; you can use only the free method if you prefer.
YinJen is a GEO monitoring desktop app (macOS / Windows) that automates the entire manual process above:
- Prompt auditing: define your own question set, configure hundreds of questions once, and test them automatically every day—solving both the sample-size and random-variation pitfalls;
- Coverage of all 12 AI engines: Doubao, DeepSeek, Kimi, Qwen, ERNIE Bot, Tencent Yuanbao, Zhipu Qingyan, and StepFun, plus ChatGPT, Claude, Gemini, and Perplexity, all aligned in one table across domestic and overseas engines;
- Mention rate and visibility score: Visibility Score (0-100) plus a trend line, aligning content actions with changes in mention rate and solving the inability-to-track pitfall;
- Competitor share-of-voice comparison + citation-source tracing: see not only your own score, but also which brands AI voted for and which content it used as evidence;
- Paragraph-level optimization recommendations: see how to fill content gaps, down to specific paragraphs.
One clarification is necessary: YinJen provides verifiable monitoring data and does not promise “guaranteed AI indexing or recommendation.” We recommend caution with any service that makes such a promise (see the FAQ). There is a 14-day free trial with no credit card required, and the Creator plan starts at ¥29.9/month. Official website: https://zhimahang.com/yinjen ; download: https://zhimahang.com/yinjen/download .
Real Case: What Is Hidden Beneath a 50% Overall Mention Rate?
From 2026-06-30 to 07-10, an electronic-components manufacturer in Suzhou (network transformers / RJ45 connectors) used YinJen to run 43 monitoring questions and 524 valid tests, producing an overall mention rate of 50.2% (263/524). That does not look bad—but break it down:
- The split by engine was polarized: Qwen 85%, Tencent Yuanbao 83%, Claude 74%, but ChatGPT 17%, DeepSeek 12%, and Gemini only 4%;
- Doubao 41%, DeepSeek 12%—Doubao is the AI-native app with the largest monthly active user base in China (345 million), while DeepSeek also has 127 million monthly active users (QuestMobile 2026Q1). Visibility was lowest precisely at the entry points with the most customers;
- “Manufacturer recommendation” questions had hit rates of 54%-75%, while the 19 “product-selection knowledge” questions produced only 1 hit (5%)—when customers asked AI during the selection stage, AI was using other people’s content to educate the manufacturer’s customers;
- DeepSeek named and recommended other brands in comparable questions, while this brand was completely absent.
Not one of these four conclusions would be visible from 30 manual tests. That is the difference between “measure once” and “monitor continuously.”
FAQ
Q1: How much does one manual measurement cost? It is free; the only cost is time. 10 questions × 3 engines takes approximately 1-2 hours. We recommend measuring one fixed round every month and reusing the same question set; otherwise, results cannot be compared over time.
Q2: What mention rate counts as passing? There is no industry-wide passing threshold. The right approach is to measure your own baseline first, identify structural gaps by engine and question type, then compare longitudinally against yourself and horizontally against competitors. The case above had an overall rate of 50% but only 5% for product-selection questions—a typical example of “the total passes, but the structure fails.”
Q3: What should I do if my mention rate is low? Can improvement be guaranteed? First locate the gap (which question category and which engine), then create content targeted at that gap. Citation-source tracing will tell you whom AI currently cites. But no compliant method can “guarantee” AI indexing or recommendation—the CCTV 315 Gala (2026-03-15) exposed the underground GEO practice of “AI poisoning”: 11 sponsored articles were enough to get a fictitious product recommended by AI. That path is neither acceptable nor sustainable. There is only one compliant path: monitor → find gaps → create content → monitor again to verify.
