DeepSeek, Doubao and AI answers: what changes for your site?
Your website still counts, but on its own it won't carry you, and much of what's sold as AI optimization comes without data.
The short answer
AI apps change your plan, but they don't replace your website. In one 2026 vendor study, brand sites earned 12.9% of citations, and DeepSeek now says it runs its own search engine. Foreign sites are barely measured. Build a crawlable Chinese site, earn real third-party mentions, and test answers yourself.
- CNNIC estimated more than 700 million generative AI users in June 2026. Its latest search count, for December 2025, was 782.06 million, down from 877.82 million.
- On October 17, 2026, DeepSeek said its search runs on an in-house engine that filters GEO manipulation.
- A July 2026 preprint from a GEO tool vendor gave brand sites 12.9% of 160,860 citations, and found 8.3% of brands in cited sources reached the answer.
- A small vendor test of Chinese B2B software prompts, run March to June 2026, put 14% of DeepSeek citations on global English sources.
- Regulators in Beijing and Shenzhen each fined a GEO provider RMB 50,000 in 2026, according to Chinese press reports.
You’ve probably heard that Chinese buyers now ask Doubao instead of searching. Part of that is true, and it’s being used to sell a lot of services with little public testing behind them.
This page sorts what’s documented from what’s claimed. It’s part of our guide to which Chinese search engines a foreign company should target.
Baidu’s own AI results have their own page: Baidu AI answers and foreign websites. So does WeChat search. Here we cover the standalone apps.
Are Chinese buyers moving from search to AI apps?
Partly, yes. Search engine use fell in 2025 while generative AI use more than doubled, and CNNIC says smart search products are weighing on classic search engines.1
| Measure | Figure | Date | Source |
|---|---|---|---|
| Search engine users | 782.06 million, 69.5% of internet users | Dec. 2025 | CNNIC 1 |
| Search engine users, a year earlier | 877.82 million, 79.2% | Dec. 2024 | CNNIC 1 |
| Generative AI users | 602 million, up 141.7% | Dec. 2025 | CNNIC 1 |
| Generative AI users | More than 700 million, over 50% of the population | June 2026, estimate | CNNIC 26 |
| AI-native apps, monthly users | 499 million, up 85.4% | May 2026 | QuestMobile 2 |
| Doubao, monthly users | 382 million | First half 2026 | QuestMobile 2 |
| Doubao, monthly users | 528 million | June 2026 | Xsignal, Beijing News 3 |
| DeepSeek, monthly users | 130 million | First half 2026 | QuestMobile 2 |
| DeepSeek, monthly users | About 165 million | June 2026 | Xsignal, Beijing News 3 |
| AI general assistants, uses in the month | 43.47 billion, up 162.8% | July 2026 | QuestMobile via CNNIC 26 |
| AI search engines, uses in the month | 38.50 billion, up 9.9% | July 2026 | QuestMobile via CNNIC 26 |
CNNIC found that answering questions is the top use of generative AI, at 76.0% of users.1 That’s the use that competes with search. Assistants are also growing far faster than the apps CNNIC files under AI search engines.26
For the buying side, see how Chinese buyers search for foreign services.
Where do DeepSeek and Doubao get their answers?
From different places, and that’s the core problem. DeepSeek documents its own search engine. We found no equivalent ByteDance documentation for Doubao.
| App | Owner | What the company documents | What outside sources report |
|---|---|---|---|
| DeepSeek | DeepSeek | An in-house engine with “hundreds of billions” of pages and multilingual search 4 | Numbered citations can appear when Search runs, says one guide 22 |
| Doubao | ByteDance | Nothing we could find on its index | Citations on the app lean on Douyin, says one study 6 |
| Yuanbao | Tencent | Its App Store listing advertises search of WeChat Official Accounts and Channels 28 | Sources differ widely between app and web 6 |
| Qwen | Alibaba | Nothing we could find on the app’s index | App and web barely overlap 6 |
DeepSeek’s change log says “web and API search capabilities are powered by our in-house DeepSeek search engine.”4
If that’s accurate, it undercuts the idea that DeepSeek simply reads Baidu. The note doesn’t name the mobile app.
Two more limits apply. Search has to run for citations to exist, and one guide tells users to turn it on before asking.22 An answer drawn only from the model involves no live page at all.
| Platform | App and web source overlap by domain (1 = identical) |
|---|---|
| DeepSeek | 0.507 6 |
| Doubao | 0.505 6 |
| Yuanbao | 0.299 6 |
| Qwen | 0.190 6 |
How much do these apps cite brand websites?
Some, but third parties lead in the best dataset we found, which gives brand and corporate sites 12.9% of citations.6
That’s enough to question “your site is invisible,” and too little to rely on your site alone.
The study is a preprint by Zhen and colleagues at Aidso, which makes an AI search monitoring tool.
It ran 614 Chinese queries, three times each, on the app and web versions of four platforms in June and July 2026.6 Aidso designed the queries itself, so read the shares as one sample, not the market.
- News media28.8%
- Vertical industry portals24.1%
- Social media and creators16.4%
- Brand and corporate sites12.9%
- E-commerce7.6%
- Encyclopedic knowledge5.7%
- Government4.5%
Zhen et al. 2026
| Source type | Share of citations 6 | Unique domains 6 |
|---|---|---|
| News media | 28.8% | 676 |
| Vertical industry portals | 24.1% | 810 |
| Social media and creators | 16.4% | 68 |
| Brand and corporate sites | 12.9% | 6,589 |
| E-commerce and transactions | 7.6% | 285 |
| Encyclopedic knowledge | 5.7% | 54 |
| Government | 4.5% | 1,225 |
- News media28.8%
- Vertical industry portals24.1%
- Social media and creators16.4%
- Brand and corporate sites12.9%
- E-commerce and transactions7.6%
- Encyclopedic knowledge5.7%
- Government4.5%
Read the domain column as well. Brand citations are spread across 6,589 sites, so each one earns few. Social citations come from just 68 domains.
Your sector shifts the mix. Government sources reached 14.7% in law and 8.2% in healthcare, and e-commerce sites reached 21.9% in hotels.6
Platforms differ too. Doubao’s app gave its top domain 28.6% of citations, and the authors call that concentration within the Douyin ecosystem. DeepSeek’s top domains took about 5%.6
One practitioner reports Doubao’s sources moving toward Douyin video through 2026, using screenshots.25 That fits, but it’s one person’s observation.
What gets a brand named in an answer?
Being cited doesn’t mean being named. Only 8.3% of brands found in cited sources reached the answer text.6 The other 91.7% didn’t make it in.
These are the main patterns in the study, strongest first. They’re associations from one dataset, not rules.
- Repetition across sources. Brands matched in one citation were selected 2.1% of the time. Brands matched in 11 or more were selected 59.9% of the time.6
- Fit with the question. Selection rose from 1.1% at the lowest content-fit score to 21.1% at the highest.6
- Site quality score. Selection rose from 2.7% in the lowest band of a composite SEO score, built on 5118 and Baidu webmaster ratings, to 16.5% in the highest.6
- Position in the source list. Brands in the first citation were selected 18.5% of the time, against 2.4% after the 12th.6
- Recency, for citations. Half-life among dated cited pages was about 39 days for time-sensitive queries, and 68 days for others.6
That SEO score matters less than it first looks. It ranked third for brand selection, but it wasn’t the leading predictor in most of the study’s models, and it was a weak signal for citation depth.6
So classic SEO strength seems to help, without settling the outcome.
Do foreign websites appear in these answers?
Sometimes, but it’s barely measured. We searched in English and Chinese and found one small vendor test, and nothing on Doubao or Yuanbao.
MaxAEO, which sells AI visibility services, ran 45 Simplified Chinese prompts on B2B software, three times each, from March to June 2026.27
| Engine | Domestic sources 27 | Global English sources 27 | Citations logged 27 |
|---|---|---|---|
| DeepSeek | 82% | 14% | 1,120 |
| Qwen via Quark | 89% | 8% | 1,043 |
It also reports a brand’s own English .com cited in 6% of Chinese answers.27 The article doesn’t define “global English” precisely, and it covers three software categories, so treat this as a hint, not a benchmark.
The rest of what we have is indirect:
- DeepSeek says its engine is multilingual,4 and its API accepts a filter such as
lang:zh,en.5 That alone doesn’t show English pages get cited in answers to Chinese questions. - DeepSeek’s API can localize results by IP address.5 So what a person in China sees may differ from what you see abroad.
- The Aidso brand figures don’t separate foreign sites, so 12.9% may say little about a foreign company.6
- In an audit run from Milan, Baidu showed AI overviews on 82.0% of Chinese queries and 4.0% of English ones.19 That’s Baidu, not the apps, so it’s only a hint.
Verdict
Assume a Chinese question needs a Chinese page. English pages can get cited, as that test suggests, but the shares are small and the sample is narrow.
What the agencies claim, and what the evidence says
Most guides on this topic come from companies selling the service. Some of their advice may well be right. Very little of it comes with data.
| Claim | Who says it | What we found |
|---|---|---|
| “Chinese AI models consistently ignore official brand websites” | Marketing to China 7 | Disputed. One vendor study gives brand sites 12.9% of citations 6. Foreign brands’ own sites are barely measured 27 |
| Chinese AI weights Zhihu, WeChat and Baike far above your site | iClick, and The Egg for Baike and Zhihu 810 | Partly supported. Third parties lead in one study 6, but iClick names no research, and we found no platform-level ranking data |
| DeepSeek favors content over 2,000 words | KAWO 9 | No data given, and none found |
| Doubao treats video engagement as a trust signal | KAWO 9 | Unproven. The Douyin concentration is real 6, the mechanism isn’t shown |
| Schema helps you get cited | iClick 8 | Untested here. The Aidso crawl left out JSON-LD entirely 6 |
Marketing to China is candid on one point: it says data on AI adoption and GEO market size is thin, and that specific numbers should be treated as directional.7 We agree.
Free check
Is your website visible on Baidu?
Enter your website. We check Baidu, access from China and how your site is set up, then email you the report.
Request received
We will send your report by email.
What should you change on your site?
Less than you’d think. Most of this also helps classic search, which is why it’s safe to do without proof. None of it guarantees a citation.
- Publish real Simplified Chinese pages. Follow how to translate your website into Chinese, and check whether machine translation is good enough for each page.
- Answer one question per page. Content fit was among the strongest patterns above. Use the words your buyers use, and give the answer early.
- Keep key facts current and dated. Update pages when something changes, and show the real date. Don’t change the date alone.
- Write contact details in the body text. Plain text is easy for any system to read. This is our inference, not a tested result.
- Make sure the pages load from China. Start with whether your site is accessible in China. DeepSeek says its fetching depends on a site’s availability and access policies.5
- Do the classic search work. CNNIC still counted 782.06 million search engine users.1 See how to rank on Baidu from outside China.
Then work outside your site. News media and sector portals are the types that dominate the data.6 Trade press, partners and customers who’d genuinely mention you are the honest way in.
That takes time and real relationships. If someone offers a shortcut, read the section on GEO agencies below.
Should you block or allow AI crawlers?
Decide bot by bot, and don’t expect much control. Documentation is weak, and a user-agent string can be copied by anyone.
| Agent | Status | What’s reported |
|---|---|---|
| DeepSeekBot | No DeepSeek crawler documentation found | The name comes from public registries. Compliance listed as “not always” 24. DeepSeek’s guide names no crawler 5 |
| Doubaobot | Listed by third parties | Tied to Doubao, for training and answers 11. A rule for one agent doesn’t cover another 23 |
| Bytespider | ByteDance says it follows robots.txt | Independent evidence is disputed. It feeds Toutiao Search and model training 23 |
xSeek warns that blocking Bytespider removes a site from Toutiao and other ByteDance news surfaces.11 So a blanket block has a cost, if that’s accurate.
Your server logs are the most direct evidence you have, and Search Engine Journal recommends them for these platforms.12 Read a hit on these names as a clue, not as proof.
How do you test your own visibility?
By hand, on a schedule, with a fixed list of questions. You won’t need a tool to start, and the study’s design gives you a simple model.
- Write 10 to 20 prompts in Chinese. Mix category, problem and comparison questions, and add your brand name to a few.
- Choose your platforms. Start with Doubao and DeepSeek, then add others if your buyers use them.
- Test app and web separately. Make sure search is on where the option exists.
- Run each prompt three times. The study did the same, and one answer proves little.6
- Record four things. Whether you’re named, whether you’re cited, which domains are cited, and whether the facts are right.
- Repeat monthly. Compare the cited domains over time. Those domains are your outreach list.
Ask someone in mainland China to run the prompts, since results may vary by location.5 We found no current official information on whether foreign numbers can open Doubao or DeepSeek accounts.
Clicks are harder. Baidu’s AI Traffic Analysis report doesn’t cover Doubao, DeepSeek, Qwen or Kimi.12
The Egg says DeepSeek visits may show up as direct traffic,10 and we found no China-specific click data at all.
Is it risky to hire a GEO agency?
It can be, and enforcement has started. Regulators have fined GEO providers over fabricated or unsupported claims. Those cases don’t settle how honest content and outreach sold under the same label would be treated, so ask a local adviser.
| Date | Event | Source |
|---|---|---|
| March 2026 | CCTV’s 3·15 gala exposes AI “poisoning,” including a nonexistent fitness tracker recommended by chatbots. 16 organizations sign a self-regulation convention | OECD 13 |
| May 2026 | Xinhua announces a project for a first GEO standard | Xinhua 17 |
| Reported June 2026 | Beijing’s Chaoyang regulator fines Baifu Technology RMB 50,000 under unfair competition rules, over unsupported figures and inflated cases | East Money 14 |
| August 2026 | The Xinhua-led standard, T/CAPT 026-2026, is released | Xinhua 29 |
| September 2026 | The China Commercial Advertising Association’s five standards, T/CAACCHINA 001-005-2026, are published, reported as effective October 10 | Sohu 31, 10jqka 30 |
| Reported October 2026 | Shenzhen’s regulator fines a small provider RMB 50,000 for fabricated ratings, treated as false publicity under unfair competition rules | IT Home 15 |
| Reported October 2026 | MLex reports fines in Guangzhou and Shenzhen for fabricated rankings and misleading content | MLex 16 |
| October 2026 | DeepSeek says its engine filters “SEO spam, GEO manipulation, and low-quality AIGC” | DeepSeek 4 |
Baifu is a useful warning. It sold top-three or top-five positions in AI search, and placed AI-drafted articles through media intermediaries without checking them.14
These cases concern the providers. The reports we read don’t say whether a client brand shares any liability, so raise that with your adviser too.
The new standards are group standards, not laws, and each set is reported as a first.2930 The advertising association’s set is described as recommended.30 The Xinhua-led one separates white-hat from black-hat GEO, and prohibits corpus poisoning and fake consensus.29
Our guide to banned advertising words covers the related rules on claims.
Before you sign, ask any provider these questions:
- What exactly will you publish, and where? Made-up ratings, rankings and figures are what these firms were fined for.
- Can you prove each claim? You’d be putting your name on it.
- Do you guarantee a position? Nobody outside these platforms controls that.
- How do you measure it? Ask for prompts, dates, repeated runs and raw answers. The advertising association’s standards set metrics with minimum sample sizes.30
The bottom line
If you only keep one thing, keep this: AI apps raise the bar for your site, but they don’t cancel it. You need Chinese pages that load, and true claims repeated by sources other than you.
Foreign sites are barely measured in these answers so far. So test your own prompts, and be wary of anyone who sounds certain.
How Index in China can help
We translate your site into Simplified Chinese with a human reviewer, work on getting it indexed on Chinese search engines and send a monthly report on Chinese traffic.
We don’t promise rankings, traffic or mentions in AI answers. See what we do, or book a 30-minute call.
Book a 30-minute callCommon questions
Do Chinese people still use Baidu, or have they moved to DeepSeek and Doubao?
Both, and one person can use several. Baidu reported 644 million monthly app users in June 2026, against 708 million in September 2025.2021 Those are different months, so compare with care. No source we read shows how many searches moved from one to the other, or how that differs for B2B buyers.
Can an AI answer get my company wrong?
Yes, and it can be hard to trace. In the Aidso study, about 13% of brand mentions matched no cited source, and about 71% of contact details weren't in the crawled body text, which left out footers.6 Untraceable doesn't mean wrong, but check your brand regularly. Chinese rules require providers to run a complaints channel.32 We found no brand correction process at Doubao or DeepSeek.
Does llms.txt help with DeepSeek or Doubao?
We found no evidence that it does, and no DeepSeek or ByteDance documentation that mentions it. A file like this takes little time to add, so there's no great harm in it. Just don't count it as progress, and don't let it delay the Chinese pages that any of these systems would actually have to read.
Do I need an ICP to appear in AI answers?
We found no evidence either way. None of the sources we read links an ICP to citations in Doubao, DeepSeek, Yuanbao or Qwen. An ICP is tied to mainland hosting rather than indexing, so treat it as a hosting decision. Whether your pages load from China is something you can actually check.
Can I use AI to write my Chinese content?
You can draft with it, but have a person check every page. DeepSeek says its engine filters low-quality AI-generated content, without defining the term.4 Chinese rules also require users who publish AI content through distribution services to declare it.18 Ask a local adviser how that applies to your own site and accounts.
Sources
- CNNIC: 57th Statistical Report on China's Internet Development (Chinese)
- IT Home: QuestMobile report on AI apps (Chinese)
- The Beijing News: Xsignal ranking of AI apps (Chinese)
- DeepSeek: API docs change log, Opening up the in-house search API
- DeepSeek: Search and Fetch API guide
- Zhen et al.: What Do Chinese-Language Generative Search Engines Cite and Surface? (arXiv preprint)
- Marketing to China: GEO in China, why your website is invisible to DeepSeek
- iClick: AI search optimization in China, the 2026 GEO guide
- KAWO: DeepSeek, Doubao and Qwen comparison
- The Egg: The 5-step GEO strategy for China's AI search era
- xSeek: Doubao user agents
- Search Engine Journal: Baidu now measures AI citations
- OECD AI Incidents Monitor: AI data poisoning via GEO in China
- East Money: Beijing Baifu Technology fined over GEO claims (Chinese)
- IT Home via Phoenix News: Shenzhen GEO provider fined RMB 50,000 (Chinese)
- MLex: Two businesses in China's Guangdong fined for manipulating AI search results (summary only)
- Xinhua: GEO standard project announced in Beijing (Chinese)
- Cyberspace Administration of China: Measures for Labeling AI-Generated Synthetic Content (Chinese)
- Li et al.: Auditing Source Exposure in Baidu and Google AI Search (arXiv preprint)
- Baidu: second quarter 2026 results, Form 6-K
- Baidu: third quarter 2025 results, Form 6-K
- FixAEO: Can DeepSeek search the web?
- PPC Land: Bytespider
- Trakkr: DeepSeekBot
- Baiyang SEO on NetEase: Doubao's sources (Chinese, practitioner opinion)
- CNNIC: Generative AI Application Development Report 2026 (Chinese)
- MaxAEO: Regional AI answer engines, which sources they cite
- Tencent Yuanbao: App Store listing, Hong Kong storefront (Chinese)
- Xinhua: GEO trusted information standard T/CAPT 026-2026 released (Chinese)
- 10jqka: GEO group standards announced by the China Commercial Advertising Association (Chinese)
- Sohu: five GEO group standards take effect (Chinese)
- Cyberspace Administration of China: Interim Measures for Generative AI Services (Chinese)