Chinese AI Search Is Hallucinating Your Competitors Into Existence

Generative search engines are now the gatekeepers of brand visibility in China — and they’re making up contact information 71% of the time. If your business depends on being found, cited, or called, this paper is a fire alarm.

What happened

Researchers ran a controlled, large-scale audit of four mainstream Chinese-language generative search platforms (Web and App interfaces — eight interface variants total), firing 614 queries across three replications each. From 214,119 raw records, they built a clean dataset of 160,860 citation-level observations and dissected how these systems select sources, attribute credit, and surface brand information in generated answers. The headline numbers are brutal: the overall brand-selection rate from the citation pool was just 8.3%, meaning the vast majority of retrieved sources never appear in the answer the user actually sees. More alarming for anyone running a local or service business: 71% of contact-information exposures (phone numbers, addresses) shown in answers could not be matched to the crawled body text of cited pages — a textbook hallucination problem layered directly onto source attribution. The study also found that AI visibility decays fast: fitted content half-lives were roughly 39 days for high-timeliness queries and 68 days for low-timeliness queries. And GEO practitioners take note: Baidu’s own composite quality score was not the leading predictor for any of the outcomes examined — content fit, cross-source occurrence count, and semantic role mattered more.

Cold read

This is a China-specific study of four unnamed platforms — the methodology is rigorous, but direct applicability to Google AI Overviews, Perplexity, or Western RAG products is an assumption, not a finding. The “71% of contact-information exposures unmatched to crawled text” figure sounds catastrophic, but the crawl methodology matters enormously: if pages were crawled at a different time than the query was run, or if JavaScript-rendered content was missed, some of that gap may be a measurement artifact rather than pure hallucination. The study covers query snapshots from a fixed period; benchmark contamination or platform-side A/B testing during the study window could introduce noise. Finally, 13% of brand exposures unmatched to the citation pool is striking, but without knowing how often those phantom brands are correct versus fabricated, the business risk is ambiguous — it could be signal (the model knows things the crawler missed) or pure confabulation.

What it means for you

Signal maturity: 4/5 — Large-scale, controlled design with real numbers; limited only by geographic and platform scope
Who gets hurt: Local service businesses, healthcare providers, and any brand relying on accurate contact details appearing in AI-generated answers in Chinese-language markets
What breaks if this is true: Your entire GEO / LLMO strategy built around “get cited = get surfaced” collapses — citation and answer inclusion are largely decoupled, and optimizing for citation rate vs. mention rate becomes non-negotiable
Why it might not land: Western operators may dismiss this as a China problem; platform behavior is also a moving target — these systems are updated constantly, and a 39-day content half-life means the study’s specific citation patterns may already be stale
Watch for: Any Western replication of this methodology on Perplexity, Google AI Overviews, or ChatGPT Search; specifically, whether the Web/App interface divergence finding holds — if it does, every brand audit needs to be run twice

Forecast as of 2026-07-20

By Q2 2027, at least two major Western generative search audits using comparable methodology will document contact-information hallucination rates above 40% on at least one mainstream platform, forcing a public response from the platform (correction features, citation verification labels, or explicit disclaimers) — this is too legally and reputationally exposed to stay invisible.

Source: What Do Chinese-Language Generative Search Engines Cite and Surface? A Large-Scale Empirical Study — Tao Zhen, Yue Liu, Gege Zhang, Yixuan Niu. https://arxiv.org/abs/2607.15771v1