Expert Perspective | Guidelines for Trust Tree in the AI Search Era – Interpretation of Three Group Standards by the China Advertising Association GEO

When the entry point for information acquisition shifts from"search box"to"dialog box", when users' consumer decision-making increasingly depends on an answer generated by large models, a completely new question arises for the entiremarketingindustry: In the new information environment dominated by generative engines, how shouldservices be optimized? And how should outcomesbeverified?How should content source entities be classified and graded?The China Advertising Association issued and officially implemented in October 2026three group standards regardingoptimization of generative engines(GEO)）'sself-regulatory rulethat is followable, measurable, andidentifiablefor the emergingservicefield of GEO.IdentificationUnifiedSelf-Regulatory Rules.

For any emerging service to mature, the primary question is always"Who provides it, what they do, and how they do it.The "Generative Engine Optimization Service Requirements" specifically address the service entity, service process, and behavioral boundaries, providingGEO practitioners at the three levels of service entity, service process, and behavioral boundaries.

Firstly,the standard begins with four basic principles: truthfulness and accuracy, user value and fair competition, transparency and verifiability, and prudence and commitment, as well as information disclosure and avoidance of conflicts of interest. These four principles requireGEO servicesto beconducted around real, accurate, and legal information; any content related to brand, product, qualifications, honors, cases, or evaluations musthave factual basis or supporting evidence and must notfactual basis or supporting evidence; it is notshoulduse false, misleading, or unverifiable information; it requires practitioners to aim at improving information quality and the decision-making value for the requester, rather than at deceiving users, contaminating the information environment, or disrupting the engine's normal operation; it requires objective explanations for the uncertainty and attribution boundaries of the output from the generative engine, and does not"guarantee first"permanent recommendation"" such inappropriate performance promises.

Second,in the service entity section, the standard establishes a"negative list-style"access threshold. Any entity that exhibits a lack of authenticity, has received significant penalties for false advertising or data fraud over the past twelve months, has been found to have engaged in information poisoning, corpus bombing, falsification of credible content sources, falsification of monitoring results, prompt injection, or hidden text"black-hatGEO"", shall not claim compliance with the basicrequirements of this standard.Thisclausedirectly keeps those who profit fromdirectly"poisoning"mass generation"fabrication"activitiesoutside thethreshold required by the standard, making compliance a prerequisite for business operations.

Third,in the service implementation phase, the standards are nearly"process-oriented.the process has been standardized from service confirmation, entity and authorization verification, material verification, establishment of search request samples, current status assessment, toGEO corpus processing, construction of credible content sources, technical adaptation, outcome monitoring, risk management, and closure review—using an almostshouldaim to improve content accessibility and machine readability, andshould notaim to deceive crawlers, hide content, or bypass platform security mechanisms.

Over the long term,GEOfieldhas the largest trust deficit precisely due to"self-rating"and the absence of professional, authoritative backing, which leads toscams such as fake screenshots, inducement questions, sample manipulation, drift in monitoring dimensions, data tampering, environmental manipulation, self-optimization and self-verification, and exaggerated conclusionswhich occur frequently.Refer to the following translation: Frost & Sullivan Frost & Sullivan China LeadLeo LeadLeo Research Institute LeadLeo YUAN CAPITAL SULLIVAN TELE-TREND CLOUD TECHNOLOGY TradeGo MagnaTEC Automotive & Mobility Environmental Protection & Energy Saving Technology Logistics & Supply Chain MATERNAL AND INFANT Education & Training Real Estate & Property Catering & New Retailing Advanced Materials Healthcare & Life Sciences Semiconductor & Chip COMMERCIAL AVIATION Technology, Media and Telecom LANDSCAPING Big Data & AI Infrastructure Construction & Utilities Culture & Entertainment AGRICULTURE, FORESTRY ANIMAL HUSBANDRY AND FISHERY Food & Beverage Fintech SHIPPING AND PORTS Dual Carbon & New Energy Mining & Metals Public Sector Cross-Border E-commerce Trade Building Technology, Construction & Decoration Beauty & Fashion Smart Homes Digital Infrastructure Enterprise Services Consumer Electronics"Monitoring and Verification Requirements for the Optimization Effect of Generative Engine" precisely targets this common pain point, transforming the optimization effect from a self-talked-about promotional statement into something that can stand up to verification.outcomes..

The standard first clarifies a key understanding: given that generative responses inherently exhibit volatility, any single query resultbeused as proof of outcomes. For this reason, it introduces a series of strict scientific sampling rulesto minimizeis notclose to the actual effectshoulduse at least3 time windows of repeated sampling, with no fewer than 30 valid observations for a single key indicator; the intention phrase set must be jointly confirmed and sealed by all parties prior to monitoring, and during monitoring, it should notbeunauthorized to increase, reduce, or alterbedirectly compared. These rules directly target"problematic practices such as selecting friendly questions, hiding sample sizes, excluding unfavorable samples, and creating apparent improvements"and counter theseusing scientific methodology againstself-verifying fraud.

In terms of the indicator system, the standard establishes a six-dimensional framework covering visibility, citation quality, answer quality, stability, compliance, and process quality, providing clear definitions and calculation methods for indicators such as brand mention rate, proportion of target source citations, answer accuracy, negative misreference rate, evidence completeness, valid observation rate, sampling review pass rate, and original record completeness. It also requires that each valid observation be traceable to specific intention phrase sets, query identifiers, query time, platform entry, environment status, answer snapshot, collection log, and record file hash. Most importantly, the standard clearly defines value red lines: when dealing with indicators such as negative misreference rate and risk content handling rate,it should not—formal effect verificationshould"eliminating real negative information"as the optimization goal, and for real negative facts,they shouldbe handled in reasonableways such as fact correction, supplementary explanation, service rectification, or dispute appeal, rather thanbenotused as"suppression"blocking"the objectand thusand"compliance governance"and"malicious post-removal"to be distinguished.

More crucially, it features an independent design. The standardrequiresthat institutions that issue verification conclusions with external credibilitybeindependent of the optimized service being verified, and should notbeboth the optimization implementer and the external verification; inorganizations under the same legal entity that also operate two types of businesses, they shouldshouldestablish an isolation mechanism and disclose conflicts of interest in the report. During verification,when conductingNot less than the sealing intention phrase setshould be conducted, and deviations due to data fraudshouldDirectly determine that the verification fails. The verification report truthfully presents adverse outcomes and fluctuations, andshould notonly report favorable data, andshould notuse"guarantee"control"permanent"unique"official internal referral"such misleading statements. This set of repeatable, traceable, and independently verifiable rules breaks the"self-rating"self-verifying myth, returning outcomes to a real, reproducible path.

GEO’s essence is optimizing content identified, cited, and recommended by large models; the quality of content hinges on whether the content source entity itself is credible. When cross-domain content ecosystems generally face the common challenge of lacking an information traceability mechanism and inconsistent standards for verifying authenticity, false information and misleading content risks accumulate, undermining public trust in high-quality content. The 'Classification, Grading, and Selection Requirements for Content Source Entities in Optimization of Generative Engine' were created to address this issue.

The most innovative contribution of the standard is the establishment of a"six-dimensional scoring, four-level grading, and dynamic adjustment"trustworthiness assessment model. It evaluates content source entities across six dimensions: entity qualifications, authenticity records, traceability and machine interpretability, behavioral compliance, stability and continuity, and external independent verification, extracting quantifiable, verifiable indicators, calculating trustworthiness, and thereby classifying content sources into A-level highly credible, B-level generally credible, C-levelbasically credible, and D-levellimited credible"whether it is the official website"whether it is a major platform"vague intuition, and upgrade to an objective profile that is quantifiable, verifiable, and comparable—the channel identity itself is no longer inherently equivalent to credibility; the subject’s qualifications, genuine records, traceability, compliance performance, operational stability, and external cross-verification are the true criteria for determining the level of credibility.

What’s more, the grading is not a permanent status. For those who engage in serious illegal activities, are determined to have engaged in black-hatGEO, or cause major negative public sentiment due to content, they will directly not be rated and will not enter the candidate list; at the same timeit is recommendedto establish a full-cycle dynamic management mechanism including regular reviews every6 months, immediate reviews in case of major incidents, and options such as downgrade, exit, temporary rectification, and re-application, and even adding a"stable/fluctuating/oscillating"Stability indicator. Complemented by a hierarchical selection strategy:A-level is prioritized for inclusion but the key facts stillshouldrequire verification of the original source,B-level is selected after routine verification and cross-verification, C-level is limited to specific scenarios and used with caution, and D-level is not used, and is only for clue discovery. The standard also encourages independent third-party assessments, and advocates for industry associations to lead, with platforms and service providers working together to establish a hierarchical information sharing and joint oversight mechanism for content sources, enabling the joint prevention and control of cross-platform, organized false information and regulatory violations. This mechanism ensures that good currency receives the certainty of priority reference, while bad currency loses the conditions for contaminating the information ecosystem.

In the current context where generative artificial intelligence is profoundly reshaping the information dissemination landscape, the formal release and implementation of these three group standards by the China Advertising Association serve as both a benchmark and a trigger——Establishingstandards can foster trust, and fostering trust can enable long-term success.When norms become consensus, verifiability becomes a habit, and authenticity becomes the baseline,GEOfieldwill surely bid farewell to impetuosity and bubbles and embrace a new era of high-quality development centered on quality and trust.

(Author:Jiang ZhibinProfessor at Shanghai International Studies University)