Do AI assistants cite creator content? Lessons from China

Peter Landa Founder, ShareHub LinkedIn

Published 2026-11-02 · Updated 2026-11-02

Two public studies on AI citation sources, plus our own experience improving AI visibility in China, and what brands can and cannot control; study data checked on 2026-11-02.

If creators publish honest content about your product, will AI assistants such as ChatGPT or Perplexity mention your brand more often? Public studies say it is possible, but no one can promise it. In China we tested models with fixed questions, added real content and tested again: it helped, and it faded when we stopped. This guide covers what carries over abroad, what a brand can control, and what no one can promise. Study data checked on 2026-11-02.

Key takeaways

• Both public studies show that AI assistants cite platforms where real people post (YouTube, Reddit and similar) for a large share of their sources, but the two studies rank them differently, and rankings shift by assistant and month.
• Our method in China has four steps: test models with fixed questions, see which sources they cite, add real and accurate content, and test again after a while. The method does not depend on any one platform, so it carries over.
• It has to be kept up. On one local-service brand in China we worked for about six months and reached first place on specific questions; after we stopped for about two months, we saw mentions fall.
• No provider can guarantee that an assistant will mention a brand, or set its rank. What a brand controls is whether content is real, facts are consistent, ads are disclosed, and visits can be traced.

What the studies say

The table lists only what each study reports itself. Samples, months and methods differ, so the numbers should not be compared side by side.

Read together, the studies support only a limited conclusion. Platforms with real people posting are cited often, but Reddit was not near the top for ChatGPT or Perplexity in June 2025 and ranked first in the March 2026 summary. "One platform always works" does not hold; change is normal.

What we did in China

In China we used the same idea to improve how AI assistants treat a local home-service brand that we ran ourselves. The steps were plain:

1. Fix a set of questions, including general questions and questions with a city name, and ask models such as Doubao and Qwen on a schedule. Record whether the brand is mentioned and where it ranks.
2. Look at which sources the answers cite, meaning where the model read this kind of information.
3. Fill the gaps: publish real, accurate articles on the platforms that get cited, covering service facts, common questions and real scenarios. A person reviews the content, and nothing is mass-produced.
4. Test again after a while, see what changed, and decide what to write next.

Here is a record from our monitoring system for this brand (Chengdu, home-based care services, 8 questions, each asked of Doubao and Qwen 6 times, 57 valid answers, tested October 14 to 15, 2026):

A few points stand out. First, results vary by question: the brand led on bathing-service questions and did not appear at all on the broader "home care" question, so improvement has to be judged question by question, not as a single rank. Second, the 428 citations came from 129 sites, mostly aggregator sites (about 28%), general media (about 21%), government and institutional sites (about 9%), authoritative media (about 8%) and content platforms (about 8%). Platforms where real people write, such as Zhihu columns and Sohu, were cited often, and the brand's own site made up only 0.5%. That matches the direction of the public studies: assistants cite third-party content more than what a brand says about itself.

By our internal records, more than 20 customers learned about us through AI models and then phoned our hotline after about six months of sustained manual publishing (internal records, not audited by a third party). We also saw mentions fall after we stopped publishing for about two months.

The case has clear limits. It covers one brand, one industry, one city and two Chinese models, with tests run within two days and a modest sample. AI answers are random by nature, and we cannot fully separate how much of the decline came from stopping and how much from model updates and competitor content. It shows that the method helps and needs upkeep. It does not show that you will get the same result.

What carries over and what does not

The method carries over. The platforms and rules do not: in each market, look again at who the models actually cite, then decide where your content goes.

What ShareHub does in this method

ShareHub (www.sharehub.fun) covers step three, where the real content comes from: it helps brands find and contact micro-creators who publish their own honest experience in their own language. Each creator gets a dedicated tracking link so the brand can see who drove clicks, and we help build an independent site or landing page where those visits can land. Some tools, such as tracking links, are free during early access; the paid creator-matching pilot starts from US$99 per creator, with details confirmed after we talk.

To be clear, ShareHub does not sell "being recommended by AI" and cannot control any assistant's answers. We provide the parts you can check: finding people, contacting them, and tracking. For how to find suitable micro-creators, see the guide to finding micro-creators; for tracking, see the guide to tracking referrals by creator.

Four things a brand can control

1. Keep product facts public and consistent. Names, specifications and prices should match across your site, product pages and the material you give creators, so assistants are less likely to read conflicting statements.
2. Work with real creators. They should have used the product and publish their own experience.
3. Disclose the relationship as the rules require. The US Federal Trade Commission's endorsement guide says that endorsements with a material connection must be clearly disclosed.
4. Keep testing and filling gaps, instead of doing it for a while and stopping. Fix your questions, test on a schedule, and adjust content by the results. For checks that make a site easier for AI to crawl and understand, see the website GEO checklist.

Common mistakes

• Saying "creators will post" means "you will be recommended by AI". No study supports that promise.
• Doing it for a while and stopping. In our own experience, the effect faded once we stopped.
• Having every creator post the same text. Identical content feels unreal and works poorly as independent sources.
• Betting on a single platform or model. The studies show rankings move by assistant and month.
• Not recording the source. Without dedicated links and visit records, you cannot tell whether a collaboration brought real visits.

Scope and limits

This guide summarizes two public studies and one case from China. It does not cover every assistant, region or industry. Study data goes out of date, so the update log gives the check date. The China case rests on our own monitoring records and internal data, from a small sample in a single industry, and is for reference only. ShareHub is in early access, we have not yet reproduced this case abroad, and we cannot promise the same result.

Frequently asked questions

If creators publish about my product, will AI assistants mention my brand?

Not necessarily, and no one can promise it. Studies show assistants often cite platforms where real people post, but which platform and which brand they pick changes with the assistant, the time and the industry. What you can do is keep product facts clear, work with real creators, and keep testing and adjusting.

Which platform do AI assistants cite most?

The two public studies disagree. In Ahrefs data from June 2025, YouTube had 16.1% in Perplexity, above Wikipedia at 12.5%; in Peec's March 2026 summary, Reddit and YouTube were first and second. The lesson is that platform rankings change, so test again with your own questions on a schedule.

Is doing it once enough?

No. In our experience in China, about six months of steady work showed results, and mentions fell after we stopped for about two months. A steadier approach is to fix your questions, test on a schedule, and treat it as ongoing work rather than a one-off campaign.

How do I know whether creator content brought real visits?

Give each creator a dedicated link, record the clicks, and compare them with the platform's own data. Whether an assistant mentions your brand needs fixed questions tested on a schedule; there is no reliable one-time method.

Sources and update log

• Ahrefs: most mentioned domains in AI assistants, published June 2025, checked 2026-11-02.
• Peec: most cited domains in AI search, March 2026, checked 2026-11-02.
• FTC: endorsement guide, checked 2026-11-02.
• ShareHub: current product status.
• China case: monitoring records of a brand we ran (October 14 to 15, 2026) and internal data, not audited by a third party.
• Update log: 2026-11-02 first published; corrected the same day: in the Ahrefs data YouTube is above Wikipedia in Perplexity and AI Overviews.