July 2026 Chinese Generative Search Research (Part 1): Why a Retrieved Brand May Still Not Appear in an AI Answer

“A relevant page was retrieved, so the brand will appear in the answer” is a risky assumption. Retrieval, visible citation, brand surfacing, and contact-information surfacing are distinct stages in a generative-search pipeline.

The paper *What Do Chinese-Language Generative Search Engines Cite and Surface? A Large-Scale Empirical Study*, published on July 17, 2026, offers unusually large-sample evidence for Chinese GEO monitoring. Across Web and App interfaces for four major platforms, the researchers used 614 queries and three replications per query-platform-interface combination, collecting 214,119 raw records and cleaning them into 160,860 citation-level records.

This first part asks why entering a source pool is not the same as entering an AI answer.

What the study separates within visibility

The research jointly examines citation behavior, source attribution, entity exposure, and cross-interface consistency. The distinction matters: citations describe evidence that is visibly shown; entity exposure describes whether a brand, organization, or contact detail reaches the user. One cannot substitute for the other.

Under this study's definitions and sample, brands in the citation pool were selectively surfaced in answers at an overall rate of 8.3%; 12.4% of retrieved sources containing contact information contributed contact information to answers. These are not universal exposure rates for any brand, nor a promise that publishing a page creates an 8.3% chance of appearing. They show that many available source and entity signals were not adopted into the final answer in a controlled observation.

For a business, tracking only whether the official site was cited is therefore insufficient. Monitor whether the answer accurately communicates the brand, capability, eligibility, pricing scope, or contact path.

Which signals explain more than one composite quality score

The study's predictive analysis finds that content fit to the query, cross-source occurrence count, and semantic role are relatively important, while the 5118-Baidu Composite Quality Score was not the leading predictor for any outcome studied.

This should not be simplified to “quality scores do not matter” or “repeat publication wins.” It is a monitoring reminder: whether a page directly serves the question, can be consistently corroborated across sources, and provides a definition, comparison, fact, or action signal deserve separate audit beyond one external score.

Why freshness should be monitored by question type

For cited pages with publication dates, the paper fits approximate half-lives of 39 days for high-timeliness queries and 68 days for low-timeliness queries. These figures describe freshness in the research sample; they do not require a company to rewrite every page every 39 days.

A practical response is risk-based review: pricing, versions, offers, inventory, locations, policies, and service scope need more frequent verification, while stable concepts and historical explanations can use a longer window. An update must also change the underlying facts, not merely a displayed date.

Direct implication for GEO monitoring

A Chinese AI-search report should keep four layers distinct: entry into the source pool, visible citation, brand surfacing, and accurate surfacing of critical facts. Watching only one layer can mistake “the system saw it” for “the user understood it.”

GEO Radar at https://www.georadar.top can support fixed question sets, multi-platform analysis, and reporting on brand mentions, competitor co-occurrence, and answer differences. For answers with citations, retain both source links and answer text for human review of source-expression consistency.

Part two examines a harder finding: why Web and App surfaces on the same platform can have systematically different source sets, and why citations do not fully explain a surfaced brand or contact detail.

Sources for this article

• arXiv, July 17, 2026, *What Do Chinese-Language Generative Search Engines Cite and Surface? A Large-Scale Empirical Study*: https://arxiv.org/abs/2607.15771
• arXiv PDF, research design, citation-level data, predictive analysis, and limitations: https://arxiv.org/pdf/2607.15771
• arXiv record, publication date and author information: https://export.arxiv.org/api/query?id_list=2607.15771