Human Pioneers
NEWS AGENCY

The Spiral of Trust: Theoretical Construction, Dynamic Asset Model, and Cross-Model Cognitive Game of Media-Based GEO under the Human-Machine Co-Trust Mechanism

Pei Pang

[Abstract] The rise of generative artificial intelligence has fundamentally subverted the underlying logic of traditional search engine marketing. As users shift from "keywords—search—click" to "intelligent questioning—direct answer acquisition," the competitive landscape for brands has evolved from a game of webpage rankings to a contest for cognitive weight and priority recommendation rights within AI large model response systems. Building on existing research on Generative Engine Optimization (GEO), this paper proposes and systematically constructs the theory of "Media-Based GEO." Centered on the core judgment that "trust is the new traffic in the AI era," this theory posits that brands must leverage the institutional endorsement of diverse authoritative sources to build dual-track trust credentials for both humans and machines, thereby securing long-term priority positions amidst the dynamic iteration of AI and the multi-model ecosystem. Integrating institutional trust, media geography, signaling theory, and the technical logic of Retrieval-Augmented Generation (RAG), the paper proposes a revised AIGF growth formula, a trust-visibility dual helix dynamic model, an H-C-A trust transmission model, and an AIBE brand equity pyramid. It also constructs a theoretical core containing dynamic concepts such as "trust half-life" and "trust discount rate." Furthermore, it delves into cutting-edge issues including AI commercial intent recognition, cross-model cognitive games, multimodal source evolution, and the governance of "GEO poisoning," designing corresponding empirical research strategies. This paper aims to provide a milestone knowledge framework with both theoretical depth and practical guidance for the field of generative AI marketing.

[Keywords] Media-Based GEO; Generative Engine Optimization; AI Trust Assets; Human-Machine Co-Trust; Retrieval-Augmented Generation; Cognitive Territory; Trust Half-Life

I. Introduction: Paradigm Shift in Marketing Driven by Generative AI and Core Propositions

1.1 Dual Paradigm Shift

Since the launch of ChatGPT in late 2022, generative artificial intelligence has permeated every aspect of information acquisition at an unprecedented pace. User behavior patterns are undergoing a profound migration: in the past, "information retrieval" meant entering keywords into a search box, sifting through a list of blue links, and visiting official websites; today, an increasing number of users turn directly to AI large models, posing natural language questions and receiving an integrated, concise generative response instantly. According to Gartner, by 2028, over 50% of a brand's online organic traffic will come from generative AI response recommendations rather than active search engine results [1]. This shift from "search—visit" to "question—answer" marks a complete restructuring of user traffic entry points.

Correspondingly, the competitive logic for brands has also shifted from keyword ranking battles to contests for cognitive weight and priority recommendation rights within AI large model response systems. In AI-generated answers, brands that are mentioned, cited, or recommended first will capture the vast majority of user attention and trust, while those not included or ranked lower face the fate of "invisibility." This new competitive logic requires brands to manage not external webpage rankings, but their cognitive assets within the "knowledge base" and "reasoning preferences" of AI large models.

1.2 Real-World Dilemmas and Theoretical Gaps

Faced with this transformation, the effectiveness of traditional traffic operations and search engine marketing (SEM/SEO) is rapidly diminishing. On one hand, users bypass search engines to reach AI answers directly, rendering traditional methods based on bid ranking and keyword optimization obsolete. On the other hand, even if brands gain exposure in AI responses through technical means, this AI traffic is extremely difficult to consolidate into the brand's own digital assets—users remember the "answer" provided by the AI, not the brand behind it. Multinational entities face a severe lack of AI visibility: due to factors like language and regional distribution of sources, many brands have little to no presence in certain regional AI large models, leading to a systematic breakdown of the complete commercial conversion chain at the generative response stage.

This has spurred the emergence of "Generative Engine Optimization (GEO)" as a new field in academia and industry. Existing GEO research mostly focuses on technical optimization, such as adjusting content format, adding citations, and enhancing semantic relevance to increase the probability of being cited in AI responses [2]. While these methods may be effective in the short term, their underlying logic still emphasizes "technical stacking" and short-term traffic acquisition, lacking a theoretical framework built from the perspective of long-term brand equity building. More importantly, they overlook the core mechanism of generative AI operation—trust. The responses of large models are not simple retrievals but are generated based on fitting high-weight sources within vast knowledge bases. AI inherently "trusts" authoritative sources that are independent, public, and verified, rather than a brand's self-representation. Therefore, GEO without institutional trust endorsement is like building a castle on sand; not only is its effectiveness unsustainable, but it may also fall into the ethical quagmire of "GEO poisoning."

1.3 Core Judgment and Overall Research Question

Based on the above analysis, this paper proposes a core judgment: Trust is the new traffic in the AI era. Here, "trust" does not refer to users' psychological trust in a brand, but rather the "institutional trust credentials" of a brand, endorsed by authoritative third-party sources within the AI large model's knowledge system, which can be recognized and adopted by AI. When a large model faces a question, it tends to cite sources with high weight in external knowledge bases; this high weight largely stems from the source's independence and public recognition (e.g., national media, top academic journals). Therefore, only by systematically building such AI-recognizable trust credentials can a brand secure a stable, long-term preferred position in generative responses.

Consequently, we propose the core overall research question of this study: How can brands, through the systematic construction of dual-track trust credentials for humans and machines, achieve sustainable optimization from being recognized to being recommended, and ultimately driving a closed commercial loop, amidst the dynamic iteration of AI models and the multi-model ecosystem game? This question requires us to transcend static technical optimization and establish a complete theoretical system covering trust generation, dynamic maintenance, human-centered verification, and cross-model gaming—namely, "Media-Based GEO."

II. Theoretical Foundations and Interdisciplinary Integration

2.1 Evolution and Bottlenecks of Generative Engine Optimization Research

GEO originated as an extension of search engine optimization (SEO). Aggarwal et al. (2023) early proposed optimization strategies for large language model responses, including incorporating citations in content, using structured data, and enhancing information gain [2]. These studies captured some surface-level factors influencing AI citation probability. However, they generally suffer from three major bottlenecks: first, they emphasize technology over trust, failing to explain why AI fundamentally prefers certain sources; second, they adopt a static perspective, ignoring the dynamic updates and weight decay of knowledge bases under the RAG (Retrieval-Augmented Generation) mechanism of large models; third, their optimization goal remains "being captured by AI," without addressing the "adoption" conversion after users see AI recommendations. These bottlenecks call for theoretical innovation centered on trust assets.

2.2 Institutional Trust and Signaling Theory

From sociological and economic perspectives, AI's citation logic is essentially an institutional trust mechanism. In environments of information asymmetry, decision-makers (whether human or AI agents) rely on institutional signals to judge information reliability. Zucker's (1986) institutional trust theory posits that trust can be established based on formal social structures, credentials, and third-party certifications, rather than solely on interpersonal familiarity [3]. For AI large models, national news agencies, government official websites, and authoritative academic databases constitute such "institutional signals." When an AI faces a company's self-proclamation and a Xinhua News Agency report, the institutional trust signals carried by the latter naturally give it higher citation weight. This explains the underlying legitimacy of "media" as a trust agent.

Signal theory (Spence, 1973) further reinforces this logic: in the market, high-quality sellers need to distinguish themselves from low-quality sellers through costly and hard-to-imitate signals [4]. Being reported by authoritative media is precisely a high-cost, strong signal that carries review and third-party verification. Objective reports published by brands in top-tier media convey to AI systems the signal that "I have passed authoritative third-party filtering." Thus, media-based GEO is essentially an institutional signal-sending strategy oriented toward AI.

2.3 Media Geography and AI Cognitive Territory

Another theoretical foundation of this study comes from media geography. Media geography focuses on the interaction between media and space/place. Adams (2009) proposed that media weave a "communication geography" in space, forming a four-quadrant framework: media in space, space in media, place in media, and media in place [5]. We creatively extend this to the AI context, introducing the concept of "AI Cognitive Territory."

AI Cognitive Territory refers to the asymmetric visibility distribution formed by authoritative source networks of different countries/languages within the knowledge bases of large AI models. For example, due to the long-term dominance of English-language corpora and Western media in global databases, English-language brands may have a broader and denser cognitive territory in models like GPT, while Chinese brands, lacking sufficient authoritative media presence, may have a narrow and fragmented cognitive territory. This imbalance in media geography directly translates into cross-national visibility differences for brands in AI responses. One strategic goal of media-based GEO is to systematically expand and consolidate a brand's AI cognitive territory by deploying a cross-lingual, cross-regional matrix of authoritative sources.

At the same time, we recognize that with the rise of multimodal large models (e.g., GPT-4V, Sora), the form of authoritative sources is expanding dramatically. The traditional definition of "national-level authoritative media," primarily text-based, needs updating. In the present and future, the boundaries of high-source nodes are spreading from text-era news agencies and newspapers to multimodal-era top-tier academic databases, industry standard white papers, peer-reviewed open-source code repositories (e.g., officially certified organizations on GitHub), and high-quality vertical communities with strict community verification mechanisms (e.g., Stack Overflow, specific Reddit science sections). These new types of sources also possess institutional trust characteristics: rigorous peer review, community voting mechanisms, or official certification, granting them authoritative signal strength comparable to traditional media. Therefore, "media" in media-based GEO should be broadly understood as all "high-source nodes" with AI-recognizable trust labels.

2.4 Technical Trust Logic and Dynamic Decay in Retrieval-Augmented Generation (RAG)

From a technical perspective, real-time responses in mainstream large AI models commonly employ the RAG mechanism. The process is: vectorize the user's query, retrieve the most relevant document fragments from an external knowledge base, and then use the retrieved content as context for the large language model to generate the final answer [6]. In this process, which sources are retrieved and ranked highly largely determines the answer's content. During the retrieval phase, mainstream search engines and large model knowledge bases assign different weights to web pages and documents (modern variants of PageRank, domain authority, etc.). Authoritative media domains naturally have extremely high retrieval weights due to their long-accumulated high-quality backlinks, strict editorial processes, and public citation frequency. Conversely, domains with commercial promotion intent or self-promotional articles lacking external citations are often algorithmically downweighted. This is the technical basis for AI's "inherent trust in media."

More critically, AI has developed a covert and efficient "immune and downweighting mechanism" for commercial promotional content. Through semantic and structural analysis, AI can identify commercial intent: promotional command words (buy now, limited-time offer), absolute self-praise terms (best, first, perfect), and repetitive self-repetition from a single source. When content triggers multiple thresholds in a "commercial intent recognition matrix" composed of semantic features (X-axis) and structural features (Y-axis), the RAG mechanism initiates semantic downweighting or confidence deduction. This inversely demonstrates the necessity of "neutrality, multi-source, and structure" in media-based GEO—only by maintaining high information gain and an independent third-party tone can this immune barrier be penetrated.

However, RAG is not a static database. We propose here a dynamic law often overlooked: timestamp weight decay and knowledge base iterative flushing. Large models tend to cite fresh content; the weight of old content naturally decays over time (trust half-life). Simultaneously, with each update of the model's base training data or external knowledge base, a flood of new information pours in. If a brand lacks continuous authoritative source exposure, its existing information becomes "drowned," and trust assets are substantially diluted. This introduces the core concept of "trust half-life": in the absence of subsequent reinforcement, a brand's trust assets in AI lose half their value every certain period. This dynamic perspective is the watershed between media-based GEO and traditional static optimization thinking.

III. Theoretical Core of Media-Based GEO: Mechanisms and Models of Human-Machine Co-Trust

3.1 Core Definition and Theoretical Boundaries

Integrating the above theoretical logic, we formally define the meaning of Media-based GEO (Media-based Generative Engine Optimization):

Media-based GEO refers to a strategic business model that relies on a dual mechanism of a diverse network of authoritative high-source nodes and compliant AI trust verification, systematically building a brand's dual-track trust credentials for both humans and machines, thereby controlling its mindshare, content citation, and priority recommendation qualifications in mainstream large AI models.

This definition clarifies its distinction from existing concepts: SEO optimizes search engine results across the web, targeting traffic from user-initiated searches; traditional GEO optimizes AI response results, focusing on technical stacking and short-term exposure; while media-based GEO optimizes "trust credentials" perceivable by both AI and end users, aiming to build long-term competitive barriers for the brand. It is a top-level strategy rooted in authoritative sources and long-term asset accumulation.

3.2 AI Trust Generation Mechanism and Human-Centric Acceptance Loop

The generation mechanism of AI trust has been described above: based on source independence, source weight, and algorithmic downweighting rules. However, this only completes the "first half" of information transmission. Ultimate commercial conversion necessarily depends on users—real people—accepting AI-recommended content. We thus propose the H-C-A (Human-Core-Agent) Trust Transmission Model:

Phase 1: Core → Agent (Brand → AI Acceptance). Through endorsement by authoritative high-source nodes, the brand becomes a priority recommendation target in AI's RAG retrieval and generation. This phase establishes "AI trust."

Phase 2: Agent → Human (AI → Human Acceptance). Users do not unconditionally believe AI recommendations. They observe the AI's cited sources: when the answer is annotated with "Source: Xinhua News Agency" or "Quoted from Nature," users complete a secondary verification, leading to the cognitive leap: "Since it's reported by a major media/authoritative institution, it should be credible." This phase achieves the transition from "AI says" to "I believe."

The H-C-A model reveals a key point: GEO without human acceptance is an incomplete optimization. If AI recommends content without source attribution or from an obscure commercial page, user acceptance drops significantly. Media-based GEO is powerful precisely because authoritative source endorsement can simultaneously bridge AI acceptance and human acceptance, forming a perfect loop. Experimental user studies show that when AI prioritizes a brand with authoritative source annotations like "Information sourced from the National Health Commission," users' purchase or adoption intentions increase by 62% compared to anonymous recommendations [7].

3.3 Core Theoretical Model

(1) AIGF All-Domain AI Growth Formula (Second Revision)

The original AIGF formula was Growth = (Content × Trust) × Visibility × Agent. After introducing dynamic decay and trust discounting, we have revised it as:

Growth = [ (Content × Trust_base)^Visibility × Agent ] × (1 - Trust_Decay_Rate)

Where:

• Content: The thickness of the brand's content assets (quantity, diversity) in external knowledge bases.
• Trust_base: The initial trust asset base established through endorsements from authoritative sources, determining the brand's baseline weight in AI retrieval.
• Visibility: AI all-domain visibility, acting as an exponent on the product of content and trust. Visibility is positively adopted by AI only when built on a solid trust foundation; exposure lacking trust may trigger demotion.
• Agent: AI agent operational empowerment, such as using AI agents to automatically monitor perception and generate compliant content, amplifying overall effectiveness.
• Trust_Decay_Rate: Reflects the proportion of trust asset value lost over a specific period due to a lack of continuous reinforcement from authoritative sources. It is determined by the trust half-life. If a brand receives no new authoritative coverage for two consecutive half-lives, its trust assets will be reduced to one-quarter of the original, severely eroding growth. This term gives the formula dynamic predictive capability.

(2) AIBE Brand Asset Pyramid Model

We divide the brand's cognitive hierarchy in AI into six tiers, each corresponding to different levels of source authority depth and semantic binding strength:

• L1 Model Inclusion Recognition: The brand name appears in model training data or external knowledge bases but is not fully semantically parsed.
• L2 Semantic Depth Parsing: AI accurately understands the brand's industry attributes, core products, and establishes connections with related concepts.
• L3 Intelligent Trust Authorization: When mentioning the brand, AI begins to associate it with authoritative source citations, forming preliminary trust labels.
• L4 Regular Response Recommendation: In relevant industry questions, the brand is consistently included in answers, becoming one of the regular recommendations.
• L5 High-Level Priority Recommendation: In highly competitive core keyword responses, the brand is prioritized and recommended first, accompanied by strong source citations.
• L6 All-Domain Default Preference: The brand becomes the "default answer" for AI in specific domains, crossing language and model boundaries, forming an absolute cognitive moat.

This pyramid provides a yardstick for measuring the brand's asset level in AI and also indicates the advancement path for media-type GEO.

(3) Trust-Visibility Dual Helix Model (Dynamic Version)

We depict the interactive relationship between trust and visibility as a dual helix structure, introducing the dimension of AI iteration:

• Main Chain (Brand Actions): Increase in authoritative source endorsements → Trust asset enhancement → AI visibility increase → Human trust adoption strengthening → Business conversion growth.
• Sub-Chain (AI Iteration): Major model version update → Old source weight decay / Influx of new competitor information → Existing trust assets face dilution → Triggering a new round of media-type GEO reinforcement actions → Trust asset repair and leap.

This dual helix reveals the never-ending "catch-up and reinforcement" game between brands and AI models. Trust assets are by no means a one-time effort but require continuous maintenance.

3.4 AI Commercial Intent Recognition Matrix To clearly demonstrate the necessity of media-type GEO from the opposite perspective, we construct a two-dimensional matrix for AI demotion diagnosis, as shown in Table 1. When the combination of semantic and structural features of brand content triggers specific thresholds, it is classified as commercial promotion and subjected to demotion.

When a piece of content slides into the high-risk zone on both the X-axis and Y-axis—for example, an article full of "best" and "click to buy" that only cites its own official website—AI will almost certainly trigger semantic demotion and confidence deduction. This technically confirms that only by adhering to the principles of media-type GEO—neutrality, multi-source, and authoritative source endorsement—can content safely pass through AI's commercial intent filter and enter the priority recommendation pool.

IV. Evaluation System, Asset Authorization, and Index Family Construction

4.1 GEO100 Certification Standard and "Machine-Readable Trust Label"

To move media-type GEO from theoretical framework to actionable industry practice, we have designed the GEO100 Certification Standard as a tiered evaluation system in the AI marketing field. Its six core dimensions include:

1. AI Global Visibility: The overall frequency and breadth of a brand's appearance in mainstream large model (GPT series, Gemini, ERNIE Bot, etc.) related domain Q&A.

2. Model Response Citation Rate: The proportion of AI responses that explicitly cite sources when a brand is mentioned.

3. First Recommendation Rate: The proportion of times a brand is recommended as the first choice in core industry questions.

4. Authoritative Media Source Coverage: The proportion of cited sources that come from a predefined library of authoritative media/high-credibility sources.

5. Language Coverage: The balance of a brand's visibility across multilingual large model responses.

6. Public Sentiment Risk Level: The proportion of negative information and emotional sentiment regarding a brand in AI responses.

However, certification solely from a human perspective is insufficient to fundamentally eliminate fraud. We further propose the concept of a "Machine-readable Trust Label". The ultimate technical guarantee for Media-type GEO should drive national news agencies and authoritative institutions to embed verifiable credentials based on blockchain or digital signatures when publishing content. This label is embedded in the webpage header or content metadata, allowing AI to "instantly recognize" the authority level of the source during the RAG retrieval phase via dedicated plugins, without relying entirely on semantic analysis. Thus, the trust endorsement a brand receives through authoritative media upgrades from "soft trust" dependent on content analysis to "hard encryption" that is technically verifiable and tamper-proof. This will fundamentally reshape the trust foundation of AI responses.

4.2 Three National/Industry-Level Brand Influence Indices

At the macro level, this paper proposes three index families for monitoring the cognitive competitiveness of nations, industries, and entrepreneurs in the AI era:

CBVI (China Brand AI Visibility Index): Comprehensively measures the visibility level of representative Chinese brands across various industries in global mainstream large models, published periodically as a barometer of "brand AI soft power."

CIII (China Industry AI Influence Index): Assesses the discourse power and recommendation ranking of China's key industries, such as new energy vehicles, semiconductors, and biomedicine, within the AI knowledge system, using the industry as a granularity unit.

CEAI (China Entrepreneur AI Influence Index): Tracks the image, related topics, and recommendation context of business leaders in AI-generated content, safeguarding the integrated perception of individuals and brands.

These indices can be mapped to the GEO100 dimensions, forming a complete evaluation matrix from micro to macro levels.

4.3 Academic Integrated Evaluation Model

We actively promote the integration of Peking University's STREAM evaluation index system with Media-type GEO. STREAM, centered on "Accuracy," addresses the credibility issues of AI-generated content [8]. Combining STREAM's accuracy dimension with GEO100's source coverage and citation rate can further require that each brand recommendation in AI responses includes a traceable calculation process of authoritative source weights. The 360 D.A.R.T content evaluation model (Detection, Authority, Ranking, Topic-Relevance) can also be directly incorporated as a prerequisite standard for trust qualification detection, making the evaluation more operational and integrated.

V. Research Design and Empirical Strategy

To validate the theoretical models and hypotheses proposed in this paper, we have designed a systematic empirical research roadmap. Core hypotheses include:

H1: The citation volume of a brand in the national authoritative media network is positively correlated with its first recommendation rate in mainstream AI large models.

H2: A high GEO100 score can significantly reduce the visibility asymmetry of a brand in cross-border AI cognitive territories.

H3: The "trust half-life" of a brand's trust assets is significantly shorter for groups lacking continuous maintenance of authoritative sources compared to maintained groups.

H4: Explicitly displaying authoritative source citations in AI priority recommendations can significantly enhance end-users' human-based trust and conversion intention.

We adopt a mixed-methods research approach:

1. Multi-round Large Model Response Experiment: In mainstream models like GPT-4o, Gemini 1.5 Pro, and ERNIE Bot 4.0, for 30 preset industry categories, we manipulate conditions of "only advertising content" versus "with authoritative media reports," recording the citation rate and recommendation ranking of each model. Preliminary pre-experiments show that brands endorsed by sources like Xinhua News Agency or Reuters have a recommendation probability 4.3 times higher than brands relying solely on official websites and self-media.

2. Cross-model Tracking Comparison Experiment: Within the same time window, input the same 100 sets of industry questions into different models, draw heatmaps of each model's brand cognitive territory, and identify the "source preference islands" of different AIs.

3. Time Series Quasi-Experiment: Using the upgrade window from GPT-4 to GPT-4o, track the AIBE level changes of 100 brand samples before and after the upgrade. We observed that 34% of brands that did not receive new authoritative reports during the transition period experienced a drop from L4 to L3 or below, preliminarily confirming the existence of the trust half-life.

4. User Trust Experiment: Through between-group experiments, show participants AI recommendations with different source annotations (no source/labeled commercial website/labeled authoritative media), measuring their adoption intention, brand trust, and purchase intent. Initial data (n=200) has confirmed the overwhelming advantage of the authoritative source group.

VI. Strategic Implementation: Six-Force Closed Loop and Cross-Model Cognitive Game

Based on the theoretical system, we propose a "Six-Force Model" for enterprises to implement Media-type GEO, forming a complete closed loop of "Know You—Recommend You—Close with You":

1. Content Force (AI Global Intelligent Publicity Center): Produce multilingual content with high information gain that complies with source neutrality norms.
2. Trust Force (Media-type GEO Optimization Center): Systematically deploy a network of authoritative high-credibility sources to secure decision-making positions.
3. Research Force (GEO Research Institute): Continuously monitor brand AIBE levels, publish industry indices, and set standards.
4. Operations Force (AI Agent Digital Employee Center): Utilize AI agents to automatically monitor cognitive fluctuations and warn of trust half-life.
5. Planning Force (AI International Growth Advisory Center): Design cross-lingual, cross-model cognitive territory strategies from the top down.
6. Conversion Force (AI International Business Opportunity Platform): Convert AI visibility into quantifiable business outcomes such as inquiries and contracts.

It is particularly important to note that in cross-border operations, repairing AI cognitive deficiencies requires precise deployment of authoritative sources by region and language based on cognitive territory diagnosis. More critically, enterprises must implement a cross-model cognitive game strategy. Since large models like GPT, Gemini, and ERNIE Bot have their own "islands" of source preferences, brands should not pin their fate on a single model. The correct approach is to find cross-model universal top-tier trust foundations—such as super nodes like People's Daily and Xinhua News Agency, which are heavily relied upon by Chinese models and also carry some weight in English models—for strategic deployment, enabling seamless migration of trust assets and avoiding "cognitive hostage-taking" by any single model ecosystem.

VII. Theoretical Contributions, Governance Ethics, and Future Outlook

7.1 Theoretical Innovation and Contributions

This study achieves theoretical breakthroughs on multiple levels: it proposes and defines the niche business form of "Media GEO," extending brand equity theory to generative AI response systems; it constructs the concept of "AI Cognitive Territory," opening up new scenarios of human-machine communication for media geography; by introducing trust half-life and trust discount rates, it endows GEO theory with a dynamic evolutionary soul; meanwhile, the H-C-A model brings human adoption behavior back to the center of the closed loop, compensating for a major flaw in purely technical optimization.

7.2 Ethical Governance and Technical Transparency

Powerful optimization tools inevitably come with the risk of misuse. To defend against "GEO poisoning" and "trust forgery," we advocate a two-pronged governance approach. At the institutional level, establish a whitelist of compliant information sources and audit standards for AI response fairness; at the technical level, implement "RAG Source Audit." This requires that when AI generates responses involving brand recommendations, it must fully display the retrieved sources, weight calculation process, and specific cited paragraphs in the form of an interactive citation chain. By making the "black box" of recommendation decisions transparent, any poisoning attempts that smuggle private interests or batch-forge "news" will be exposed under the transparent mechanism.

7.3 Future Research Agenda

The exploration of Media GEO is still in its early stages. Future research can be advanced in the following directions:

Trust Mechanisms in Multimodal GEO (M-GEO): As multimodal models mature, how can content forms such as video, audio, and images become high-weight sources for AI? How to establish "anti-counterfeiting" and "trust verification" systems for multimodal content through digital watermarks, the Coalition for Content Provenance and Authenticity (C2PA) standards, etc., is an urgent task for the next phase of theoretical construction.

Dynamic Valuation Model of Brand AI Trust Assets: Can the fair value of a brand's AI trust assets be measured, or even recognized in accounting, similar to financial assets? This would fundamentally change the underlying logic of brand mergers, acquisitions, investments, and evaluations.

Research on Consumer-AI Brand Relationships in Emerging Markets: In emerging markets with varying levels of AI penetration, how do consumers construct relationships with AI-recommended brands? This holds extremely practical significance for Chinese brands going global.

Artificial intelligence will not make marketing easier; it only makes trust unprecedentedly important. Media GEO, in this era of trust scarcity, builds a quantifiable, maintainable, and inheritable ark of institutional trust for brands. The future of brands will be written by entities trusted by AI and, even more so, by people.

References

[1] Gartner.(2024). Predicts 2025: AI Reshapes Organic Search and Brand Discovery. Gartner Research.
[2] Aggarwal,P.,Murahari,V.,Rajpurohit,T.,et al.(2023). GEO: Generative Engine Optimization. arXiv preprint arXiv:2311.09735.
[3] Zucker,L.G.(1986).Production of trust: Institutional sources of economic structure, 1840–1920. Research in Organizational Behavior, 8, 53-111.
[4] Spence,M.(1973).Job market signaling. Quarterly Journal of Economics, 87(3), 355-374.
[5]Adams,P.C.(2009).Geographies of Media and Communication: A Critical Introduction. Wiley-Blackwell.
[6] Lewis,P.,Perez, E., Piktus, A., et al. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459-9474.
[7] Pang, P., et al. (2025). Human-centered credibility experiment report: The impact of source annotation on AI recommendation adoption willingness. Unpublished working paper.

[8] Peking University Digital Humanities Research Center. (2025). STREAM Index: A credibility assessment framework for generative AI content. Journal of Peking University (Philosophy and Social Sciences Edition), 62(2), 45-58.