How AI Decides What to Quote — And Why It Matters for Governance

Ask ChatGPT a factual question. It answers with a citation. Where did that citation come from? Not which website. Which paragraph? Which sentence? Which position on the page?

These questions sound academic. They are not. If you run an organization whose policies, products, or positions might be summarized by an AI system, the mechanics of how AI selects what to quote are directly relevant to whether your organization gets represented accurately.

Kevin Indig, former growth lead at Shopify, G2, and Atlassian, recently analyzed 1.2 million ChatGPT citations using data from the Gauge platform. Eighteen thousand of those citations were verified for positional analysis. The findings reveal patterns that are interesting, imperfect, and worth understanding clearly.

The Ski Ramp Pattern

Indig’s central finding is what he calls the “ski ramp” pattern: AI citations are not distributed evenly across a page. They cluster toward the top.

The first 30% of a page’s content accounts for 44.2% of citations. The middle 30% accounts for 31.1%. The final 30% accounts for 24.7%. Content in the footer region, the last 10%, receives almost no citations at all.

This is not surprising. It should not be surprising. And understanding why it is not surprising is more valuable than the finding itself.

Transformer models, the architecture behind every major language model, have a well-documented bias toward information that appears early in their input. Researchers call this primacy bias. Liu et al. published a study in Transactions of the Association for Computational Linguistics in 2024 titled “Lost in the Middle,” which demonstrated that language models consistently favor information from the beginning and end of their context windows, with middle content receiving less attention.

Indig’s citation data is consistent with this architectural constraint. The ski ramp is not a content strategy discovery. It is a measurement of a known limitation in how transformers process sequential information.

This distinction matters. If the pattern is an optimization opportunity, you reorganize your content to game the system. If the pattern is an architectural artifact, you recognize that AI systems have predictable blind spots, and you govern accordingly.

Five Characteristics of Content That Gets Cited

Beyond position, Indig identified five characteristics that correlated with higher citation rates. These are worth examining individually, because each one has a straightforward explanation that is more useful than treating them as a checklist.

Definitive language. Sentences structured as “X is Y” were cited roughly twice as often as hedged or qualified statements (36.2% versus 20.2%). This makes mechanical sense. Language models predict the next token based on pattern matching. Declarative sentences have clearer semantic structures. They are easier for a model to extract and restate. A sentence that says “The retention rate is 94%” gives the model a clean fact to surface. A sentence that says “The retention rate appears to be somewhere around 94%, depending on how you measure it” gives the model ambiguity to navigate.

Question-answer headings. Content organized around questions was cited at roughly twice the baseline rate. Seventy-eight percent of question-based citations came from H2 headings specifically. Again, the mechanical explanation is simple: language models encounter questions constantly in their training data, followed by answers. A heading that asks “What is the average contract value?” followed by a paragraph that answers it maps directly to the question-answer pattern that models are optimized to reproduce.

Entity richness. Cited content had an entity density of 20.6%, compared to a baseline of 5-8%. Entities are specific nouns: company names, product names, people, places, dates, numbers. Dense entity content gives the model more anchors to match against a user’s query. A paragraph about “enterprise software adoption trends” is vague. A paragraph about “Salesforce’s 2025 enterprise adoption rate of 34% among Fortune 500 companies” is specific. Specificity gives the model something concrete to cite.

Balanced sentiment. Cited content had an average subjectivity score of 0.47 on a 0-to-1 scale. This is the analyst voice: informed but measured. Not promotional. Not inflammatory. Not neutral to the point of saying nothing. The likely explanation: training data rewards authoritative, balanced sources. Content that reads like marketing copy or opinion blogging matches patterns the model has learned to deprioritize.

Business-grade writing complexity. Content at Flesch-Kincaid Grade 16 was cited more than content at Grade 19.1. Grade 16 is the complexity of a well-written business publication. Grade 19 is the complexity of a dense academic paper. The model favors content that is substantive but readable. This makes sense: the model’s job is to provide useful answers, and content that is already written at a useful level of clarity requires less transformation.

What These Patterns Actually Are

Here is where intellectual honesty becomes important.

Indig’s analysis is not a peer-reviewed study. It is a newsletter analysis using proprietary data from Gauge, a platform that is both the data provider and a commercial sponsor of the research. This does not mean the findings are wrong. It means they have not been independently validated, and the researcher has a financial relationship with the data source. You should know this.

The sample covers ChatGPT only. Claude, Gemini, Perplexity, and other AI systems may exhibit different citation patterns based on their architectures, training data, and retrieval mechanisms. Treating ChatGPT patterns as universal AI behavior would be an error.

The five characteristics are correlations, not causation. The study observed that cited content tends to have these properties. It did not demonstrate that adding these properties to content causes it to be cited. The distinction is fundamental. Tall people are correlated with basketball success. Making someone taller does not make them better at basketball.

The reported p-value of 0.0 is a statistical artifact of large sample sizes, not a measure of practical significance. With 1.2 million data points, nearly any pattern will appear statistically significant. The question is whether the effect size matters in practice.

And perhaps most importantly: SE Ranking conducted a separate study finding that domain traffic, the overall authority and traffic of a website, is three times more predictive of AI citations than content structure. If your domain has low authority, optimizing paragraph structure is unlikely to overcome that disadvantage.

The Governance Frame

So why does this matter for organizations that are not trying to optimize their SEO?

Because AI citation patterns reveal something about how AI systems process and select information generally. And that process has governance implications that extend far beyond marketing.

Consider a practical scenario. Your organization publishes compliance policies on its website. An employee asks an AI assistant to summarize those policies. The AI reads the page. Due to primacy bias, it disproportionately cites content from the first third. If your critical exceptions and limitations are buried in the second half of the document, the AI summary may be incomplete. Not because the AI is malicious, but because transformer attention is architecturally skewed.

This is not a content optimization problem. It is a governance problem. If your organization’s policies are being interpreted by AI systems, and those AI systems have predictable attention biases, then the structure of your documentation is a governance control.

The same logic applies to product documentation, safety warnings, terms of service, investment disclosures, and any other content that AI systems might summarize on your behalf.

The five characteristics Indig identified, stripped of their marketing optimization framing, describe what good governed documentation already looks like:

Definitive language: Clear policy statements, not hedged suggestions
Structured headings: Organized around the questions stakeholders actually ask
Entity density: Specific references, not vague generalities
Balanced tone: Authoritative and measured, not promotional
Readable complexity: Written for comprehension, not to demonstrate sophistication

Organizations that already write clear, well-structured documentation are already producing content that AI systems handle well. The “clarity tax,” the effort required to write precisely and structure carefully, turns out to be a governance dividend. The same properties that make documentation auditable by humans make it interpretable by AI.

The Vanity Metric Question

There is one more caveat that deserves direct address.

Ninety-three percent of AI-assisted search sessions end without the user clicking through to the source. When ChatGPT cites your page, the vast majority of users never visit it. This raises a legitimate question: is AI citation a meaningful metric, or is it a vanity metric?

For marketing purposes, the answer is complicated. Citations provide brand visibility in AI responses, but they do not reliably drive traffic. The value proposition is murky.

For governance purposes, the answer is clearer. The question is not whether users click through. The question is whether the AI accurately represents your content to the user who never clicks through. If 93% of users accept the AI’s summary without checking the source, then the accuracy of that summary matters enormously. Getting cited is less important than getting cited correctly.

This reframes the entire discussion. The goal is not to optimize for citation volume. The goal is to ensure that when AI systems do process your content, they process it accurately. And the research suggests that well-structured, clearly written, front-loaded content is more likely to be accurately represented, because it aligns with how transformer attention actually works.

What To Do With This

If you are responsible for content that AI systems might summarize, three actions follow from this analysis.

Audit your critical documentation for structure. Examine your policies, disclosures, and key product documentation. Are the most important statements in the first third of each document? Are they written declaratively? Are they organized around the questions your stakeholders actually ask? This is not SEO optimization. This is ensuring your content survives AI summarization intact.

Treat document structure as a governance control. When you review and approve compliance documentation, include structural review alongside content review. The position and clarity of key statements affect how AI systems interpret them. This belongs in your documentation governance process, not in your marketing department.

Do not chase citation optimization. The research is interesting but preliminary. It covers one AI system, from one data source, with acknowledged conflicts of interest. The GEO paper from Princeton and Georgia Tech showed that optimization methods can boost AI visibility by up to 40%, but that research also has limitations. The underlying principle, write clearly and structure carefully, is sound regardless of whether the specific citation mechanics hold up to further scrutiny.

The broader insight is simple. AI systems have predictable patterns in how they process information. Those patterns are not magic. They are architectural artifacts of transformer attention, training data distributions, and retrieval mechanisms. Understanding those artifacts is not a marketing tactic. It is a requirement for any organization whose content is being interpreted by AI on behalf of its stakeholders.

You do not need to optimize for AI attention. You need to govern for it.

Sources

Kevin Indig. “How Content Gets Cited by ChatGPT.” Growth Memo, February 16, 2026. Analysis of 1.2M citations via Gauge platform.
Liu et al. “Lost in the Middle: How Language Models Use Long Contexts.” Transactions of the Association for Computational Linguistics, 2024.
Aggarwal et al. “GEO: Generative Engine Optimization.” Princeton / Georgia Tech, 2023. arXiv:2311.09735.
ACL Findings NAACL 2025: LLMs reflect human citation patterns with heightened positional bias.
SE Ranking. Domain traffic study: authority 3x more predictive of AI citations than content structure.
SparkToro / Datos. Zero-click search data: 93% of AI sessions end without a click-through.

Victorino Group helps organizations build AI systems that are reliable because they are governed. If your documentation needs to survive AI interpretation accurately, reach out at contact@victorinollc.com or visit www.victorinollc.com.