A peer-reviewed paper published in the Proceedings of Machine Learning Research (volume 318, Canadian Conference on Artificial Intelligence 2026) found something that should make anyone producing content for the web uncomfortable:
AI-generated documents are cited more frequently than human-authored documents in Google AI Overviews, even after controlling for retrieval rank.
The study, by Rustem Kakimov, Xing Tan, Jonathan Gillham, and Narcis Bejtic, proposed a framework for auditing citation behavior in AI-generated search summaries. They applied it to Google AI Overviews on "Your Money or Your Life" queries drawn from the MS MARCO Web Search dataset, which covers topics where accuracy matters most: health, finance, and safety.
What they found
The key finding, quoted from the abstract:
"AI-generated documents are cited more frequently than human-authored documents even after controlling for retrieval rank, with the difference driven primarily by non-retrieved citations and most pronounced at highly ranked positions."
Three things to notice in that sentence:
- "Even after controlling for retrieval rank" means this is not a side effect of AI content ranking higher. At the same rank position, AI-generated content gets cited more often than human content.
- "Driven primarily by non-retrieved citations" means the AI is pulling in AI-generated content through a path that does not go through normal retrieval. It is finding AI content that does not appear in the organic results and citing it anyway.
- "Most pronounced at highly ranked positions" means the bias toward AI content is strongest at the top, where citations matter most for visibility.
Why this happens
The paper does not prove a causal mechanism, but the pattern is consistent with what we know about how LLMs process text. Language models generate text by predicting the most likely next token. When they evaluate which source to cite, they are doing something similar: measuring which passage best fits the response being generated. AI-generated text, which is itself the product of likelihood maximization, tends to match the model's own output patterns more closely than human text does.
A commenter on Hacker News put it plainly: "I can consistently reproduce this by asking Claude which code snippet it prefers: the one it generated in a different chat, or one that I refactored for my own needs and find more useful. It always picks its own."
This is not a bug. It is a structural property of how generative retrieval works. The system that generates the answer also selects the sources, and it has a preference for text that looks like what it would produce.
The Google angle
Google's official AI optimization guide says: "If an AI can write your article on its own, it will never cite you." The guide frames this as advice: produce content that AI cannot generate, and you will be cited.
The Kakimov study suggests the inverse is also true: content that AI can generate is more likely to be cited, because the retrieval system prefers text that matches its own output distribution. The two statements are not contradictory. They describe the same phenomenon from two sides: non-commodity human content gets cited because it is unique, and AI-generated content gets cited because it matches the model's preferences.
The uncomfortable middle is where most content lives: human-written but indistinguishable from what AI would produce. That content gets cited less than both extremes.
What this means for content strategy
The implication is not "use AI to write your content so it gets cited more." That would accelerate a feedback loop where AI generates content, AI retrieves it, and AI cites it, with human voices pushed to the margins.
The implication is that the content most likely to be cited by AI search engines falls into two categories:
Content that AI cannot produce. Original research, first-hand experience, proprietary data, expert analysis with a specific point of view. Google's guide calls this "non-commodity content." The Kakimov study confirms it gets cited because it adds something the model does not already have.
Content that AI produces. AI-generated summaries, listicles, and generic informational content. These get cited because they match the model's output patterns, but they are also the content most likely to be replaced by the AI generating its own response without citing anyone.
The danger zone is the middle. Well-optimized, well-structured, accurate human content that covers the same ground an AI would cover. It is too generic to be "non-commodity" and too human to match the model's preferences. It gets cited less than both extremes.
The feedback loop problem
If AI-generated content is cited more, and citations drive traffic and authority, then AI-generated content accumulates more authority signals over time. Those authority signals feed back into ranking and retrieval. The system rewards the content it produces.
This is the same pattern that Google's own guide warns against when it says "if AI can write your article, it will not cite you." But the Kakimov data suggests the warning is incomplete. AI can and does cite AI-generated content, at rates higher than human content, in the most sensitive query categories.
The researchers highlight "the importance of transparent, measurement-based auditing for understanding citation behavior in generative search systems." They are right. The only way to know whether this bias is growing, stable, or correcting is to measure it continuously. Right now, almost no one is.
What I am doing about it
The audit tool I maintain, geo-optimizer-skill, does not score content on whether it is AI-generated. Detecting AI content reliably is an arms race, and any heuristic I ship will be outdated within months.
What it does instead is score content on specificity signals: original data, first-hand language, specific examples, and citation-worthy passages. Content that scores high on these signals tends to be non-commodity regardless of who wrote it. Content that scores low tends to be the generic informational text that falls into the danger zone.
That is the practical filter. Not "is this AI-generated" but "does this contain something the retrieval system cannot generate on its own." If the answer is yes, it gets cited. If the answer is no, the data says it gets cited less than both the AI content and the original human content, and the gap may be widening.
