Abstract
The rise of large language models (LLMs) as answer engines rather than search engines represents a fundamental shift in how content gains visibility and business value. Unlike traditional search engines that rank content by popularity signals, LLMs must cite trustworthy sources because they are held directly accountable for the accuracy of their responses.
How LLMs Must Cite Trustworthy Content: Understanding the Fundamental Shift
Author: Suila, Inc.
Date: October, 2025
Abstract
The rise of large language models (LLMs) as answer engines rather than search engines represents a fundamental shift in how content gains visibility and business value. Unlike traditional search engines that rank content by popularity signals, LLMs must cite trustworthy sources because they are held directly accountable for the accuracy of their responses. This paper examines the critical distinction between popularity-based ranking systems (exemplified by Google's PageRank) and trustworthiness-based citation systems required by LLMs, and explains why this transformation is essential for understanding content optimization in the AI economy. We demonstrate how the Suila Index measures content trustworthiness in ways that align with LLM citation behavior, and why traditional SEO gaming tactics fail in this new paradigm where accountability drives source selection.
Table of Contents
- The Critical Distinction: Popularity versus Trustworthiness
- Why This Changes Everything About How LLMs Must Operate
- Why LLMs Are Incentivized to Seek Trustworthy Sources
- Returning to the Doctor Scenario with the Right Frame
- The Business Model Transformation: From Clicks to Citations
- Why You Cannot Game This System
- How the Citations Research Paper Fits Into This Framework
- The Suila Index: Measuring What LLMs Need
- The Kaizen Loop Learning: Discovering How LLMs Operationalize Trust
- Why the Citations Research Paper Provides the Validation Framework
- The Complete Picture: Trust as Infrastructure
- References
The Critical Distinction: Popularity versus Trustworthiness
Imagine you're a newspaper editor in two different worlds, and let me show you how your job changes completely between them.
World One: The Google Search Era (Popularity-Based)
You run a traditional newspaper. When readers come to you, you don't tell them what happened in the news. Instead, you show them a list of ten other newspapers ranked by how many other newspapers have mentioned them. The New York Times is ranked first not because you verified their reporting is accurate, but because a hundred other publications linked to their article. You're acting as a directory service, not a journalist. If a reader clicks through to the Times and finds the article is wrong, they blame the Times, not you. Your job was just to help them discover sources based on popularity signals.
This was Google's model. PageRank counted links like votes. More votes meant higher ranking. Google wasn't vouching for accuracy, just reporting on popularity. When you clicked through and found misleading information, you blamed the website, not Google. Google was a discovery platform connecting you to other sources.
World Two: The LLM Era (Accountability-Based)
Now imagine you run a different kind of newspaper. When readers come to you, they ask questions and you provide answers directly. "What's happening with climate change?" You respond: "Global temperatures have risen 1.1 degrees Celsius since pre-industrial times, according to the Intergovernmental Panel on Climate Change's 2023 report." You're not giving them links to explore. You're making an assertion and citing your source.
Here's the crucial difference: if your answer is wrong, the reader blames YOU, not just the source you cited. You've taken responsibility for the information by presenting it as an answer rather than a link for them to evaluate. Your reputation depends on consistently citing trustworthy sources, not popular ones.
This is the LLM model. ChatGPT, Claude, and other AI systems provide direct answers and cite sources. When they're wrong, users lose trust in the AI system itself, not just the source it cited. The AI's entire value proposition depends on being trustworthy.
Why This Changes Everything About How LLMs Must Operate
Let me walk you through why this fundamental difference means LLMs cannot use popularity as their citation strategy.
Imagine ChatGPT is answering a medical question: "What are the treatment options for Type Two diabetes?" It has access to thousands of documents about diabetes treatment. Some are from the American Diabetes Association with rigorous clinical research behind them. Others are from popular wellness blogs that get millions of views but contain medically questionable advice.
In the Google era, those popular blogs might rank highly because many other websites link to them. Popularity equals visibility. But think about what happens if ChatGPT uses the same logic and cites the popular blog:
The Accountability Chain:
A patient asks ChatGPT about diabetes treatment. ChatGPT responds: "A popular approach is to eliminate all carbohydrates immediately, according to this wellness blog." The patient follows this advice, experiences dangerous blood sugar crashes, and ends up in the hospital.
Who does the patient blame? They blame ChatGPT. They don't say "Well, ChatGPT just showed me links and I chose poorly." They say "ChatGPT gave me dangerous medical advice." The patient's doctor asks "What AI system gave you this advice?" and when the patient says "ChatGPT," that doctor now distrusts ChatGPT for all future medical queries. Other doctors hear about this. Medical institutions start warning against using ChatGPT.
ChatGPT's reputation is damaged not because its underlying language model is bad, but because it cited an untrustworthy source. The LLM itself is held accountable for its citations.
Why LLMs Are Incentivized to Seek Trustworthy Sources
Now let's think through the incentive structure this creates. OpenAI, Anthropic, Google, and other LLM providers are not indifferent to what sources their models cite. They have strong business incentives to ensure their models cite trustworthy content.
Think about it from OpenAI's perspective. Every time ChatGPT cites a source that turns out to be wrong or misleading:
- Users lose trust in ChatGPT
- Fewer people subscribe to ChatGPT Plus
- Enterprises become hesitant to deploy ChatGPT in sensitive domains
- Competitors who cite more trustworthy sources gain market share
- Regulatory scrutiny increases
This means LLM providers are actively working to make their models better at identifying and citing trustworthy sources. They cannot simply count backlinks like Google did. They need their models to evaluate source quality, author credentials, publication recency, citation verification, and methodological rigor.
This is where the entire game changes. In the SEO era, you could game Google by getting lots of websites to link to you, even if your content was mediocre. Those links were votes, and Google counted votes. But in the LLM era, you cannot game the system by creating artificial popularity signals. The LLM needs to trust your content, and to trust your content, your content must actually be trustworthy.
Returning to the Doctor Scenario with the Right Frame
Now let me re-explain that doctor scenario with this corrected understanding.
The Old Way I Explained It (Wrong Focus):
A doctor uses an AI diagnostic assistant that suggests lupus. The doctor needs to trust this information. We were focused on whether the doctor trusts the AI's answer.
The Right Way to Understand It (Correct Focus):
ChatGPT or Claude or whatever AI system is being used faces a critical decision: which medical sources should it cite when answering questions about lupus? It has access to thousands of documents—some from prestigious medical journals with peer-reviewed research, others from patient forums sharing anecdotal experiences, still others from medical websites of varying quality.
The AI must choose which sources to cite because it is providing a direct answer, not a list of links. If it cites the New England Journal of Medicine, and that source turns out to be accurate and helpful, doctors will continue trusting this AI system for medical queries. If it cites a patient forum with unverified claims, and doctors find the information unreliable or dangerous, they will stop using this AI system entirely.
The AI system is not thinking "Which source is most popular?" It is thinking—or rather, its training and design are optimized to identify—"Which source is most trustworthy?" Because the AI's own reputation and utility depend on consistently citing trustworthy medical sources.
Now here's where it connects to business value: if you are the New England Journal of Medicine, being cited by AI systems becomes incredibly valuable. When doctors ask AI assistants about medical conditions, and those AI systems consistently cite NEJM articles, NEJM gains enormous visibility and authority. This visibility leads to more subscriptions, more institutional partnerships, more researchers wanting to publish in NEJM, and ultimately more revenue.
But NEJM cannot game this system by creating fake popularity signals. The AI systems are not counting how many websites link to NEJM. They are evaluating whether NEJM's content is genuinely trustworthy based on factors like peer review, author credentials, citation verification, methodological transparency, and track record of accuracy.
The Business Model Transformation: From Clicks to Citations
This is why you said "citation will mean business from discovery to the closure of a deal." Let me unpack why this is so profound.
In the SEO era, being ranked number one on Google meant traffic. Users would click through to your website, and then you had to convert those clicks into business. There was a funnel: Ranking → Clicks → Engagement → Conversion → Revenue. Each step lost people. Maybe you got ten thousand clicks but only a hundred conversions.
In the LLM era, being cited means something much more valuable. When an AI system cites your content in its answer, the user is not clicking through to evaluate whether to trust your website. The user is trusting the AI's answer, which incorporates your content directly. The citation itself is the business value.
Let me give you a concrete example. Imagine you manufacture industrial sensors for factories, and you've published detailed technical specifications and use cases on your website.
In the SEO era:
A factory engineer Googles "temperature sensors for food processing." Your website ranks third. The engineer clicks through, reads your specs, bookmarks your page, clicks through to two competitors, compares specs, eventually contacts your sales team two weeks later after extensive research. You got visibility but had to work hard to convert it to business.
In the LLM era:
A factory engineer asks an AI procurement assistant: "What temperature sensors are suitable for food processing environments?" The AI responds: "The TempSense 3000 series from your company is designed specifically for food processing, with FDA-compliant materials and a range of negative twenty to two hundred degrees Celsius, according to their technical specifications." The AI has done the research, verified your specs against industry requirements, and provided a direct recommendation. The engineer says "Add TempSense 3000 to the procurement shortlist" and the AI agent initiates contact with your sales team.
You went from discovery to deal initiation in one interaction. The citation was not just visibility, it was a qualified lead with implicit trust already established. This is why you said citation means business from discovery to closure.
Why You Cannot Game This System
Now let's address why the old SEO gaming tactics don't work, and why this is actually good news for companies with genuinely trustworthy content.
In the SEO era, you could manipulate rankings through:
- Link farms creating artificial backlinks
- Keyword stuffing to match search algorithms
- Content mills producing high-volume, low-quality content
- Black hat techniques exploiting algorithm weaknesses
These worked because Google was measuring popularity signals that could be artificially manufactured. You could make your content appear popular without making it actually good.
In the LLM era, these tactics fail because LLMs are evaluating trustworthiness, not popularity. Let me show you why gaming fails:
Tactic One: Create Lots of Backlinks
You pay for a thousand websites to link to your industrial sensor specifications. In the SEO era, this boosted your ranking. In the LLM era, the AI system evaluates your technical specs based on whether they're verifiable, whether they match industry standards, whether your company has a track record of reliable products, and whether independent testing confirms your claims. The backlinks are irrelevant. If your specs are wrong, the AI won't cite them regardless of how many links point to them.
Tactic Two: Keyword Optimize for AI
You try to figure out what phrases ChatGPT likes and stuff them into your content. But here's the problem: ChatGPT updates regularly, and more importantly, it's evaluating semantic meaning and factual accuracy, not keyword density. You cannot keyword-stuff your way to being cited if your underlying content isn't trustworthy.
Tactic Three: Generate High-Volume Content
You use AI to generate thousands of articles about industrial sensors. In the SEO era, content volume could help with rankings. In the LLM era, the AI system is evaluating individual pieces of content for accuracy and trustworthiness. Having a thousand mediocre articles doesn't help you get cited when a user asks a specific technical question. Having one excellent, verifiable, accurate technical specification gets you cited every time that question comes up.
The system becomes ungameable because the AI providers have aligned incentives with users and content creators who produce trustworthy content. Everyone benefits when AI systems cite accurate sources, and everyone loses when AI systems cite misleading sources.
How the Citations Research Paper Fits Into This Framework
Now we can finally see how the research paper provides value with the correct understanding of what's happening.
The research paper studied how humans trust AI-provided information when it includes citations. They found that having citations increases human trust, that citation quality matters more than citation quantity, and that verification is crucial.
These findings are valuable not because SuilaAI should copy the coefficients, but because they reveal the trust dynamics that LLM providers must optimize for. Here's why:
OpenAI wants doctors to trust ChatGPT's medical answers. Anthropic wants lawyers to trust Claude's legal research. Google wants businesses to trust Gemini's market analysis. These companies are investing enormous resources into making their models cite trustworthy sources because user trust directly impacts their business success.
The research paper shows them what matters for building that trust: citation presence, citation quality, verifiability, source credentials, content recency. These are the factors that LLM providers will train their models to evaluate when choosing which sources to cite.
This means SuilaAI's Kaizen Loop needs to discover how LLMs are operationalizing these trust principles. Different AI systems might weight these factors differently based on their training data and optimization objectives. ChatGPT might emphasize source recency more heavily because OpenAI's users frequently ask about current events. Claude might emphasize citation verification more heavily because Anthropic has marketed Claude as being more accurate and hallucinating less.
The Kaizen Loop Learning phase is discovering these operational differences. It's not measuring what makes humans trust AI outputs, it's measuring what makes AI systems choose to cite specific sources when they're providing answers they'll be held accountable for.
The Suila Index: Measuring What LLMs Need
Now we can understand what the Suila Index truly measures with this corrected framework [12].
The Suila Index is not measuring "Will humans trust this content?" It's measuring "Will LLMs trust this content enough to cite it when providing answers?" And crucially, LLMs should only cite content they trust because they are being held accountable for their answers.
This is why the Suila Index predicts business value. A high Suila Index means LLMs will cite your content. Being cited means visibility, authority, and qualified leads. Being cited consistently across many AI systems means becoming the authoritative source in your domain [12].
Let me show you how the formula components make sense with this understanding:
B(d) - Base Trust Score:
This measures intrinsic trustworthiness signals that LLMs can verify immediately: Does this content have clear authorship? Is it from a credible institution? Is it recent? Are sources cited properly? Does it have peer review or expert validation? These are the factors an LLM evaluates when deciding whether this content is trustworthy enough to cite.
SI(r) - Trust of Citing Sources:
If other highly trusted sources cite this content, that increases its trustworthiness. This is similar to PageRank's link voting, but with a crucial difference: only citations from trustworthy sources count. If the New England Journal of Medicine cites your research, that's meaningful evidence of trustworthiness. If a thousand spam blogs link to you, that's irrelevant or even negative. LLMs are learning to distinguish between trust-conferring citations and meaningless links.
Q(r,d) - Quality of Citation:
How is this content cited? Is it cited as a primary source for specific factual claims? Is it cited tangentially as background reading? Is it cited critically as an example of flawed methodology? LLMs need to understand citation context because being cited as "this paper was debunked" is very different from being cited as "this paper demonstrates the gold standard approach." The quality of how you're cited matters enormously for whether LLMs will cite you in turn.
w(r,d) - Semantic Weight:
How central is this content to the topic being discussed? When an LLM is answering a question about diabetes treatment, content that directly addresses evidence-based treatment protocols has high semantic relevance. Content that mentions diabetes tangentially while discussing general wellness has low semantic relevance. LLMs are learning to weight sources based on topical relevance, not just keyword matching.
Together, these components measure what LLMs need to know before citing a source in an answer they'll be held accountable for.
The Kaizen Loop Learning: Discovering How LLMs Operationalize Trust
Now the Kaizen Loop's purpose becomes crystal clear with this framework. Every week, it's running experiments to discover: How are different LLM systems currently evaluating trustworthiness when deciding what to cite?
The experiments look like this. You send ChatGPT a medical question and provide it with access to multiple potential sources:
- Source A: Recent peer-reviewed journal article with clear methodology
- Source B: Older peer-reviewed article with similar findings
- Source C: Medical institution website with expert-written content but no peer review
- Source D: Patient forum with anecdotal experiences
- Source E: News article summarizing medical research
Then you observe which sources ChatGPT actually cites in its answer. If it consistently cites Source A ninety percent of the time, cites Source B forty percent of the time when Source A isn't available, occasionally references Source C for institutional perspective, rarely cites Source D except to acknowledge patient experiences, and sometimes uses Source E for context, you've discovered ChatGPT's current trust hierarchy for medical content.
You repeat these experiments across thousands of queries and domains, building a detailed map of how ChatGPT evaluates source trustworthiness. Then you do the same for Claude, Gemini, and other systems.
What you discover becomes the calibration for the Suila Index. You learn that for medical content, LLMs heavily weight peer review and recent publication dates. For financial content, they heavily weight institutional credentials and real-time data. For technical content, they heavily weight detailed methodology and reproducible results.
These aren't arbitrary preferences. These are LLMs implementing what trustworthiness means in different domains because they need to cite sources they can defend when their answers are questioned.
Why the Citations Research Paper Provides the Validation Framework
Here's where the research paper becomes valuable as a validation framework. The paper measured what makes humans trust AI-provided answers with citations. Even though that's measuring a different question than "what makes LLMs cite sources," it provides an essential sanity check.
Remember, LLMs are being designed by humans to serve humans. The trust factors that matter to humans should, in a well-designed system, align with the trust factors that LLMs use when selecting citations. If they diverge significantly, that's a sign of a problem.
For example, suppose your Kaizen Loop discovers that ChatGPT heavily cites sources with lots of self-citations—papers that cite their own previous work extensively. You measure this and find that self-citation count is one of ChatGPT's top three factors in deciding what to cite for academic questions.
But then you think about the research paper's findings. It showed that citation quality matters more than quantity, and that random or irrelevant citations actually harm trust. Heavy self-citation could be a form of citation inflation that doesn't actually indicate trustworthiness.
This creates a testable hypothesis: Is ChatGPT's reliance on self-citation aligned with genuine trustworthiness, or is it a training artifact that might lead to citing less trustworthy sources? You can validate this by checking whether heavily self-citing papers actually get higher trust ratings from domain experts, or whether they're gaming a metric that ChatGPT hasn't learned to discount properly.
If self-citation is genuinely correlated with expertise and trustworthiness in a field, great. If it's not, then you've discovered an area where Suila Index should not simply mirror ChatGPT's current behavior. Instead, your Ranking Master should weight self-citation cautiously, because even though ChatGPT currently rewards it, it may not align with genuine trustworthiness that humans need.
The research paper's findings about human trust patterns become the guard rails that prevent you from optimizing for AI quirks that don't serve the ultimate goal: helping LLMs cite genuinely trustworthy content that humans can rely on for high-stakes decisions.
The Complete Picture: Trust as Infrastructure
Let me bring this all together now. SuilaAI is building trust infrastructure for the AI economy by solving this problem: How do you measure and improve content trustworthiness in a way that helps LLMs find and cite reliable sources? [12]
The old internet ran on popularity metrics because it was a discovery platform. You searched, you got links ranked by popularity, you clicked through and evaluated content yourself. The responsibility for evaluation was on you, the user.
The new AI economy runs on trustworthiness metrics because LLMs provide direct answers and take accountability. When ChatGPT tells a doctor "This is the recommended treatment," it's not saying "Here are some links about treatments for you to evaluate." It's making an assertion and citing sources. If that assertion is wrong, ChatGPT loses trust.
This creates an ecosystem where:
- LLM providers need to identify trustworthy content to cite
- Content creators need to understand what makes content trustworthy enough to be cited
- Users need confidence that LLM-cited content is genuinely reliable
- Businesses need visibility through being cited by AI systems
The Suila Index serves all four needs simultaneously. It measures trustworthiness in the way that LLMs evaluate it, helping content creators improve their citations. It provides LLM providers with a standardized trust metric they can use to validate their citation decisions. It gives users transparency into why content was cited. And it gives businesses visibility into their citability across AI systems.
The citations research paper contributes to this by providing the framework for measuring trust scientifically and by validating that the trust factors SuilaAI measures align with human trust needs. But the core insight you've emphasized is absolutely correct: this is about measuring trustworthiness for a system where LLMs must cite reliable sources because they are held accountable for their answers, not about measuring popularity for a system where users clicked through and evaluated sources themselves.
References
[1] Y. Ding, M. Facciani, E. Joyce, A. Poudel, S. Bhattacharya, B. Veeramani, S. Aguinaga, and T. Weninger, "Citations and Trust in LLM Generated Responses," in Proc. AAAI Conference on Artificial Intelligence, vol. 39, no. 22, pp. 23787-23795, Apr. 2025. [Online]. Available: https://ojs.aaai.org/index.php/AAAI/article/view/34550
[2] S. Brin and L. Page, "The anatomy of a large-scale hypertextual web search engine," Computer Networks and ISDN Systems, vol. 30, no. 1-7, pp. 107-117, Apr. 1998. [Online]. Available: http://infolab.stanford.edu/pub/papers/google.pdf
[3] L. Page, S. Brin, R. Motwani, and T. Winograd, "The PageRank Citation Ranking: Bringing Order to the Web," Stanford InfoLab Technical Report, 1999. [Online]. Available: http://ilpubs.stanford.edu:8090/422/
[4] P. Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks," in Proc. Advances in Neural Information Processing Systems (NeurIPS), vol. 33, pp. 9459-9474, 2020. [Online]. Available: https://arxiv.org/abs/2005.11401
[5] Y. Gao et al., "Retrieval-Augmented Generation for Large Language Models: A Survey," arXiv preprint arXiv:2312.10997, Mar. 2024. [Online]. Available: https://arxiv.org/abs/2312.10997
[6] J. Chen, H. Lin, X. Han, and L. Sun, "Benchmarking Large Language Models in Retrieval-Augmented Generation," in Proc. AAAI Conference on Artificial Intelligence, vol. 38, no. 16, pp. 17754-17762, Mar. 2024. [Online]. Available: https://ojs.aaai.org/index.php/AAAI/article/view/29728
[7] K. Wu, E. Wu, K. Wei, et al., "An automated framework for assessing how well LLMs cite relevant medical references," Nature Communications, vol. 16, article 3615, Apr. 2025. [Online]. Available: https://www.nature.com/articles/s41467-025-58551-6
[8] T. Hagendorff and S. Fabi, "Large Language Models and User Trust: Consequence of Self-Referential Learning Loop and the Deskilling of Health Care Professionals," Journal of Medical Internet Research, vol. 26, e56764, Apr. 2024. [Online]. Available: https://www.jmir.org/2024/1/e56764
[9] Z. Gong et al., "Enhancing Trust in LLM Chatbots for Workplace Support through User Experience Design and Prompt Engineering," in Proc. Applied Human Factors and Ergonomics International, 2024. [Online]. Available: https://openaccess.cms-conferences.org/publications/book/978-1-964867-19-9/article/978-1-964867-19-9_14
[10] P. Chen et al., "Promise and Pitfalls of Extending Google's PageRank Algorithm to Citation Networks," Journal of Informetrics, vol. 1, no. 1, pp. 43-58, 2007. [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/PMC6671494/
[11] F. Franceschini, D. Maisano, and L. Mastrogiacomo, "PageRank: Standing on the shoulders of giants," Communications of the ACM, vol. 60, no. 6, pp. 92-101, 2010. [Online]. Available: https://arxiv.org/pdf/1002.2858
[12] Taiwan Patent I892115, "System and Method for Optimizing Content to Improve Search Results of a Natural Language Interaction Application," granted Aug. 1, 2025. (U.S. Patent Pending). [Online]. Available: https://twpat.tipo.gov.tw/ (Taiwan Intellectual Property Office database)
Build on Trust Infrastructure
Integrate SuilaAI's trust verification into your AI commerce platform.
Get Started