aicitation.hu / Articles / Q&A

AI Citation FAQ: Why Links Break and Sources Get Fabricated

A direct explanation of why AI models generate non-existent URLs, fake DOIs, and shifting facts, based on current research into retrieval and parametric memory.

By Zoltán Varga2026-11-055 min read

Why do cited links sometimes not work?

When an AI system generates a response, it often creates a list of sources to support its claims. These sources frequently include URLs that look structurally perfect but lead nowhere. The model constructs these addresses by predicting the most likely sequence of characters for a web address, not by retrieving a real page. This process results in links that appear valid but point to non-existent domains or broken paths.

Research indicates that a significant portion of these generated links are entirely fictional. Some studies have identified that between 3% and 13% of cited URLs in AI outputs are hallucinated. These links do not simply rot over time like old web pages; they never existed in the first place. Search engines and archive tools like the Wayback Machine cannot find them because no server ever hosted the content.

The distinction between a stale link and a hallucinated link is critical for verification. A stale link once worked but the content was removed or moved. A hallucinated link is a fabrication of syntax that mimics a real URL structure. Users who click these links encounter immediate errors or generic domain parking pages. This behavior stems from the model's training to complete text patterns rather than its ability to browse the live web for every single citation it produces.

Can a model invent a DOI or a study?

Digital Object Identifiers (DOIs) are unique alphanumeric strings assigned to academic papers and datasets. AI models can generate strings that follow the correct DOI format but correspond to no actual publication. These fake DOIs often include realistic-looking journal names, author lists, and publication years. The model combines these elements based on its training data, creating a citation that looks authoritative but contains no real record.

This fabrication extends to entire studies that never took place. An AI might cite a paper titled 'The Impact of X on Y' by a famous researcher, even if no such paper was ever written. The text of the citation is statistically probable, yet the underlying document is absent from all academic databases. This creates a specific risk for researchers and professionals who rely on these citations for further reading or validation.

Search engines like Google Scholar may inadvertently index these ghost references if the citation appears in enough places online. Once indexed, these fake entries can persist in search results, confusing users and polluting the academic record. The problem is not just that the AI makes a mistake, but that the error is formatted perfectly to bypass initial human scrutiny. Verification requires checking the DOI against official registries, not just trusting the visual appearance of the citation.

Why do citations change between runs?

Large language models operate on probabilistic principles, meaning they do not retrieve a fixed set of facts for every query. Instead, they calculate the most likely next word or token based on the input prompt and their internal parameters. This inherent randomness causes the output to vary even when the user asks the exact same question twice. One run might cite a news article, while the next cites a blog post or a different report entirely.

The variation also stems from the interplay between two distinct knowledge systems within the model: parametric memory and retrieval augmentation. Parametric memory consists of facts learned during training and stored in the model's weights. This knowledge updates slowly, often taking twelve to twenty-four months for major changes. Retrieval-augmented systems pull fresh data from the web in real time. A model might switch between relying on its internal memory or an external search tool depending on the specific query phrasing or system configuration.

This duality leads to inconsistent answers about the same entity. A query about a company's leadership might return an outdated CEO name from the parametric layer in one instance, and the current leader from a retrieved source in another. The user sees two different facts delivered with equal confidence. This instability makes it difficult to establish a single 'truth' from an AI response without cross-referencing multiple runs and verifying the sources manually.

Can citations be bought or manipulated?

There is no mechanism for purchasing citations or forcing an AI model to cite a specific source in exchange for payment. The generation of citations is a function of the model's training data and its retrieval algorithms, not a marketplace. Attempts to manipulate citations through paid placements or artificial link building do not directly alter the model's internal weights or its immediate retrieval logic.

However, the underlying data that feeds these models can be influenced by broader content strategies. Entities that establish a strong presence in high-authority sources, such as Wikipedia, industry reports, and academic journals, increase their likelihood of being retrieved and cited correctly. This is a structural process of improving data availability and quality, not a transactional purchase of a citation slot. The model selects sources based on relevance and authority signals found in its training corpus or search index.

The concept of 'buying' a citation misunderstands the nature of generative AI. While you cannot pay for a specific output, you can invest in the infrastructure that makes your entity visible to the model. This involves creating structured data, securing mentions in trusted publications, and ensuring your digital footprint is robust. These efforts improve the probability of accurate retrieval but do not guarantee a specific citation or prevent the model from occasionally hallucinating a different source.

How reliable are AI citations today?

Current research suggests that AI citation reliability varies significantly depending on the topic and the prominence of the entity involved. For well-documented subjects like depression, citation accuracy can reach 94%. However, for less common or niche topics, fabrication rates can rise to nearly 30%. This disparity highlights that the model's confidence does not correlate with factual accuracy, especially in areas with less training data.

High-salience brands and entities face a unique risk known as the Brand Hallucination Paradox. Because these entities are familiar to the model, they become easy anchors for plausible but incorrect completions. Studies have shown that top-tier brands can experience fabricated citation rates exceeding 50% in certain contexts. The model's familiarity with the entity name leads it to confidently attach false claims or non-existent sources to that name.

The reliability of citations also depends on the specific query type. Regulatory-framed queries, which ask about laws or compliance, tend to have higher fabrication rates than simple factual queries. This suggests that the complexity of the task and the pressure to provide a definitive answer can increase the likelihood of hallucination. Users must treat every AI citation as a lead for investigation rather than a confirmed fact, verifying the source through independent channels.

• Fabrication rates range from 3% to 30% depending on the subject matter and entity salience.
• High-profile brands often suffer from higher false attribution rates due to the Brand Hallucination Paradox.
• Regulatory and compliance queries show elevated risks of generating incorrect citations compared to general facts.
• Verification requires checking URLs and DOIs against official registries, not just trusting the AI output.

AI-assisted: drafted with a language model from the author's research notes, then edited and checked before publication. Author: Zoltán Varga, AI visibility researcher and AI ecosystem architect, Budapest.