Expert Perspective | Establishing Trust Principles for the AI Search Era — An Interpretation of the Three GEO Group Standards Released by the China Advertising Association

As the entry point for information access quietly shifts from the “search box” to the “dialogue box”, and users’ consumption decisions increasingly depend on answers generated by large language models, a brand-new proposition is put forward for the entire marketing industry: in the new information environment dominated by generative engines, what criteria should optimization services follow? How can results be verified? How should content source entities be classified and graded? In October 2026, the China Advertising Association issued and officially implemented three group standards for Generative Engine Optimization (GEO): Requirements for Generative Engine Optimization Services, Requirements for Effect Monitoring and Verification of Generative Engine Optimization, and Requirements for Classification, Grading and Selection of Content Source Entities for Generative Engine Optimization. These standards set unified, followable, measurable and identifiable self-regulatory rules for the emerging GEO service sector.

I. Establish Rules for Services: Clear Boundaries and Red Lines for Practitioners

For any emerging service to mature, the core questions to answer are always “who delivers it, what to deliver, and how to deliver it”. Requirements for Generative Engine Optimization Services defines clear behavioral boundaries for GEO practitioners from three dimensions: service entities, service workflows and behavioral limits.

First, the standard opens with four basic principles: truthfulness and accuracy, user value and fair competition, transparency and reviewability with prudent commitments, information disclosure and conflict-of-interest avoidance. These principles require GEO services to be built on truthful, accurate and legal information. All content related to brands, products, qualifications, honors, cases and reviews must be supported by facts or supporting documents. False, misleading or unverifiable information shall not be used. Practitioners shall aim to improve information quality and the decision-making value for requesters, rather than deceiving users, polluting the information ecosystem or interfering with normal engine operations. They shall objectively explain the uncertainty of outputs from generative engines and the attribution limits, and shall not make improper performance promises such as “guaranteed top placement” or “permanent recommendation”.

Second, for service entities, the standard adopts a negative-list-based access threshold. Entities that lack authenticity, have received major penalties for false advertising or data fraud within the past twelve months, or have been confirmed to conduct black-hat GEO practices including information poisoning, corpus bombing, forging trusted content sources, fabricating monitoring results, prompt injection and hidden text, are not allowed to claim compliance with the basic requirements of this standard. This clause directly blocks bad actors profiting from poisoning, volume manipulation and fraud from qualifying under the standard, making compliance a prerequisite for business operations.

Third, for service implementation, the standard defines a nearly process-oriented full SOP covering service confirmation, entity and authorization verification, document inspection, search query sample creation, status assessment, GEO corpus processing, trusted content source development, technical adaptation, effect monitoring, risk handling and post-project review. It mandates that records of key confirmations, version changes, content reviews, publication logs and exception handling be retained for review. This full-traceability design means GEO services are no longer one-off deals or opaque black boxes, but standardized deliverables that can be traced, reviewed and held accountable. Notably, the standard imposes stricter qualification checks and special review requirements for high-risk sectors including healthcare, pharmaceuticals, finance, franchise investment and education & training. It also clarifies technical boundaries: technical adaptation shall aim to improve content accessibility and machine readability, and shall not be used to deceive crawlers, hide content or bypass platform security mechanisms.

II. Eliminate Inflated Performance Claims: Verifiable Outcomes and Traceable Data

For a long time, the biggest trust deficit in the GEO industry stems from self-assessment without authoritative professional grounds. This has given rise to rampant malpractices including forged screenshots, leading questions, sample manipulation, shifting monitoring dimensions, data tampering, environment manipulation, self-verification by the same optimization provider and exaggerated conclusions. Requirements for Effect Monitoring and Verification of Generative Engine Optimization addresses this common pain point, turning optimization performance from self-promotional claims into verifiable results.

The standard first clarifies a key understanding: given the inherent volatility of generative responses, results from a single query shall not be used as proof of performance. To this end, it introduces strict scientific sampling rules to approximate real-world performance. Formal effect verification shall adopt repeated sampling across no fewer than three time windows, with no fewer than 30 valid observations for a single key indicator. Before monitoring, all parties shall jointly confirm and seal the intent phrase set, which must not be altered arbitrarily during monitoring. Monitoring shall be conducted in a clean environment free of interference from login status, historical conversations, personalized recommendations, local cache, device fingerprints and geographic differences. Baselines shall be rebuilt when major changes occur to models or product versions; data before and after such changes shall not be directly compared. These provisions target long-standing abuses such as cherry-picking favorable queries, concealing sample sizes, excluding unfavorable samples and fabricating performance gains, and use scientific methodology to counter self-attested fraud.

For the indicator framework, the standard builds a six-dimensional system covering visibility, citation quality, response quality, stability, compliance and process quality. It defines and specifies calculation methods for indicators including brand mention rate, target source citation ratio, response accuracy rate, false negative citation rate, evidence completeness rate, valid observation rate, sampling review pass rate and raw record completeness rate. Every valid observation shall be traceable to the specific intent phrase, query identifier, query time, platform entry, environment status, response snapshot, collection log and hash value of the record file. Critically, the standard draws a clear value red line: when referencing indicators such as false negative citation rate and risk content resolution rate, the optimization goal shall not be “eliminating truthful negative information”. Verified negative facts shall be addressed through factual corrections, supplementary explanations, service rectification and dispute appeals, rather than suppression or removal. This distinguishes compliant governance from malicious deletion of posts.

Equally important is the independence requirement. Entities issuing verification conclusions for external proof shall be independent from the optimization service being verified. The same entity shall not both deliver optimization and conduct external verification. Where both services are operated within one legal entity, segregation mechanisms shall be established and conflicts of interest disclosed in reports. Independent sampling review covering no less than 20% of the sealed intent phrase set shall be carried out. Verification shall be rejected if discrepancies arise from data fraud. Verification reports shall truthfully present unfavorable results and fluctuations, instead of only reporting positive outcomes. Misleading wording such as “guaranteed”, “controlled”, “permanent”, “exclusive” and “official internal promotion” shall not be used. This set of repeatable, traceable and independently verifiable rules breaks the myth of self-scoring and grounds performance assessment in real, reproducible results.

III. Distinguish Authentic Content: Tiered Sources with Valid Adoption Basis

The essence of GEO is optimizing content that can be identified, cited and recommended by large language models. The root of content quality lies in whether the content source entity itself is trustworthy. Cross-domain content ecosystems commonly suffer from missing information traceability mechanisms and inconsistent authenticity verification standards. Risks of false and misleading content keep accumulating, undermining public trust in high-quality content. Requirements for Classification, Grading and Selection of Content Source Entities for Generative Engine Optimization is developed to tackle this challenge.

The standard’s pioneering contribution is a credibility evaluation model featuring six-dimensional scoring, four-tier grading and dynamic adjustment. Evaluating content source entities, it extracts quantifiable and verifiable indicators across six dimensions: entity qualification, content authenticity track record, information traceability and machine parsability, behavioral compliance, stability and sustainability, and external independent verification. Credibility is calculated from these indicators, and sources are classified into four tiers: Level A Highly Trustworthy, Level B Generally Trustworthy, Level C Basic Trustworthy and Level D Limited Trustworthy. This framework upgrades the previous vague judgment based on whether a source is an official website or a major platform into an objective, quantifiable and verifiable profile. Channel identity alone does not equal trustworthiness. Entity qualifications, factual track records, traceability, compliance performance, operational stability and external cross-verification are the real benchmarks for trust tiers.

Notably, grading is not permanent. Entities with serious violations, confirmed black-hat GEO conduct, or major negative public opinion triggered by their content shall not be graded or included in candidate lists. The standard recommends regular reviews every six months, ad-hoc reviews triggered by major incidents, and a full-lifecycle dynamic management system covering downgrade, removal, rectification deadlines and reapplication. Stability labels (“Stable / Fluctuating / Volatile”) may even be assigned based on the standard deviation of scores across consecutive evaluation cycles. Matching grading rules are provided for adoption: Level A sources are prioritized, though key facts still require original-source verification; Level B sources may be used after routine checks and cross-verification; Level C sources are restricted to limited scenarios and used with caution; Level D sources are not adopted and may only serve as discovery clues. The standard also encourages independent third-party assessments, and advocates joint information-sharing and supervision mechanisms led by industry associations together with platforms and service providers, to jointly prevent and control cross-platform, organized false information and violations. This mechanism rewards legitimate quality sources with priority citation, while removing the ground for bad-faith content to pollute the information ecosystem.

At a time when generative AI is profoundly reshaping information communication, the release and implementation of these three CAA group standards serve both as a measuring stick and a starting gun. Trust can only be built on clear standards, and sustained progress can only be achieved with trust. When compliance becomes consensus, verifiability becomes routine and truth becomes the bottom line, the GEO industry will move away from impetuosity and hype, and enter a new phase of high-quality development centered on quality and trust.

Author | Professor Jiang Zhibin, Shanghai International Studies University
Source | China Advertising Association