Citation Rules

When a chat program answers a question, the answer usually sounds very confident. But the text itself doesn’t reveal where the information came from. Citation rules are specifications for when such a system must name its sources and in what form. This means links, footnotes, or small numbers after a sentence that point to a website or document. Some of these rules come from the companies themselves, others from laws or from contracts with publishers. The term “citation” simply means a quotation or source attribution.

What’s at stake: verifiability and clicks

Language models occasionally invent details that sound convincing but are wrong. Experts call this hallucination. A source citation is the simplest countermeasure, because the reader can check for themselves. Without evidence, all that remains is trusting the system. For school, university, and journalism, that isn’t enough.

On top of that, there’s an economic dispute. News sites depend on people visiting their articles and seeing advertising there. If an AI answers the question directly, that visit doesn’t happen. Publishers have therefore filed lawsuits or entered into licensing agreements in which visible source citations are explicitly stipulated. A mention with a clickable link is part of the consideration.

The rules are also legally relevant. Anyone who takes over substantial portions of someone else’s text touches on copyright. A correct citation does not replace permission, but it does make the origin transparent. In academia, moreover: anyone who conceals a source commits plagiarism, regardless of whether an AI was involved.

How a model gets its footnotes

A language model normally doesn’t know where its knowledge comes from. It has learned statistical patterns from huge amounts of text, not a library with shelf numbers. From memory alone, it therefore cannot produce an honest footnote. If you ask for one anyway, the result is often invented titles or dead links.

That’s why the model is coupled with a search function. This method is called Retrieval Augmented Generation, or RAG for short: the system first searches for matching documents, presents them to the model, and has the answer built only from this material. Each text passage retains an identifier in the process. This makes it possible to place the document number behind every sentence it came from. The footnote is thus not a memory fragment, but a record of the search process.

In addition, some providers check after writing whether the statement really appears in the cited source. A second model compares the sentence and the evidence and removes citations that don’t fit. This isn’t perfect. A common mistake is to assume that every displayed footnote has been verified. The link can exist and still assert something different from the sentence before it.

Footnotes in search engines, schools, and publisher contracts

Citation rules are most visible in AI search services. Systems like Perplexity or the AI answers in Google and Bing place small digits or card tiles next to the text. A click leads to the original page. In pure chat programs without web search, however, such references are often missing entirely.

Schools and universities now have their own guidelines. Many examination regulations require that the use of an AI tool be disclosed, sometimes with the name of the program and the date. This is a special form of source citation. It does not replace the obligation to name the actual specialist sources and to read them yourself.

In the business press, citation rules mainly come up in negotiations and lawsuits. When a publisher signs a contract with an AI provider, visible attribution is a central point alongside money. The European Union’s AI Act also demands more transparency about the data used. Anyone reading such reports will understand them more easily knowing that this isn’t about politeness, but about control and revenue.