The Citation You Can't Trust: Why RAG Citations Name Sources the Model Didn't Use

What a citation in an AI answer actually is

A generative answer engine does not think the way its interface suggests. Underneath almost every one of them sits a pattern called retrieval-augmented generation, or RAG, first described by Lewis and colleagues in 2020: a pretrained language model is paired with a separate, searchable index of documents that is queried at the moment a question arrives. The model then writes an answer conditioned on the passages that search returned. The original work showed this produced more specific and more factual language than a model working from memory alone, and it let the system update its knowledge without retraining.

The important thing for anyone trying to be visible is that this is three separable operations, not one. There is retrieval, which pulls candidate passages from the index. There is generation, which writes the answer. And there is citation, which decides what to display beside the text as its apparent source. Because these are separate stages, the source a reader sees is a presentation artifact produced by a downstream step. It is not, by construction, a transcript of what the model read.

That gap between what looks like a footnote and what actually functioned as evidence is the whole subject of this piece. A footnote in an academic paper is written by the same author who used the source. A citation in an AI answer often is not.

Two ways a machine attaches a source

There are broadly two production patterns for producing a cited answer, and a 2024 study of trustworthiness in RAG describes why both leave room for the citation to be wrong about itself.

In a retrieve-then-generate pipeline, the system retrieves passages, writes the answer, and then, in a distinct post-hoc pass, decides which retrieved sources to attach to which sentences. In a generate-then-retrieve pattern, the model drafts first and evidence is fetched afterward to support what it already wrote. In both, the attribution step runs independently of the reasoning step. The source credited to a sentence is selected by a matching process after the fact, so it can be a plausible-looking passage that resembles the claim rather than the passage that produced it.

The authors of that study make the structural point plainly: only citation generated inline, at the same time and by the same process that writes the sentence, can guarantee that the credited source is the one actually used. Post-hoc attribution, which is the cheaper and more common production choice, offers no such guarantee. This is not a claim that engines are dishonest. It is a claim about what the architecture can and cannot promise.

The faithfulness gap, measured

The concern is not merely theoretical. A 2024 to 2026 survey of attribution techniques for RAG systems, which catalogs the methods proposed to keep answers grounded in their sources, reports that more than 95 percent of answers from the open-source language models it tested contained at least one unattributed sentence, meaning a sentence with no faithful supporting source behind it. The gap between the citations shown and the reasoning done is, on the available evidence, the normal case rather than the exception.

It is worth being precise about what is and is not established here. That the faithfulness gap exists, and that post-hoc attribution can decouple the credited source from the used one, is a well-supported finding in the current literature. Which mitigation reliably closes the gap in production is an open research question, not a solved one; the same survey exists precisely because the field is still proposing and comparing fixes. So the reading is asymmetric: the problem is established, the cure is emerging.

Why the dominant production pattern makes it worse

The RAG literature has matured into a shared taxonomy, usually described as Naive, Advanced, and Modular RAG, which is the vocabulary now used to talk about how AI Overviews, ChatGPT Search, Perplexity, and Gemini assemble answers. Within that taxonomy, attaching citations as a separate module after generation is attractive for ordinary engineering reasons. It is fast, it is cheap, it works across different underlying models, and it lets a system show sources without rebuilding how the model writes.

Those are good reasons to build it that way. They are not reasons the resulting citation is faithful. The same properties that make post-hoc attribution practical at scale, its separation from the generation step, are exactly what let the displayed source and the reasoning diverge. When you optimize a whole industry for the convenient pattern, the convenient pattern becomes the thing buyers are asked to trust, and its known limitation becomes an industry-wide blind spot.

Google AI Overviews: reading and crediting can be two passes

The clearest illustration of the split, in the surface most local businesses care about, is Google AI Overviews, though here the evidence needs an explicit caveat.

Technical analyses drawing on Google’s own disclosures describe AI Overviews assembling a query-specific set of candidate passages through embedding-based retrieval and pairwise ranking, generating the overview text, and then running a separate citation-extraction pass that scores and attaches the links a reader sees. If that description is accurate, the passages used to write the answer and the pages credited beneath it are not guaranteed to be the same set, and a page ranking well outside the top organic results can still appear as a cited source.

We flag this deliberately as emerging evidence. It is a synthesis of Google’s public technical documentation and patent filings via secondary industry analysis, not a single peer-reviewed primary source, and it should be verified against Google’s own generative-search documentation before any specific architectural or latency figure is asserted as fact. The direction is consistent across independent write-ups; the exact internals remain unconfirmed. What the pattern does support, cautiously, is the same conclusion the peer-reviewed RAG work reaches by a different route: rank and citation are not the same event, and a shown source is not a proof of use.

What this means for the claim "we got cited"

Put the pieces together and a common vendor pitch starts to look thinner than it sounds. "We got you cited in the AI answer" is offered as though it were proof that a business’s content drove the model’s recommendation. On the evidence, a citation is weaker than that. It establishes that the business appeared as a credited source in one answer, at one time, on one engine. It does not establish that the model reasoned from that page, because the attribution step and the reasoning step can be decoupled.

A second, independent finding compresses the claim further. The Pew Research Center’s 2025 behavioral study of 68,879 real Google searches found that when an AI summary was present, users clicked a traditional result in about 8 percent of searches, versus 15 percent without a summary, and clicked a link inside the summary itself only about 1 percent of the time. So even a faithful, correctly attributed citation is rarely the thing a reader clicks. The buyer-facing value of an AI citation is being named in the answer a person reads, not being a link in a list almost nobody follows.

None of this argues that citations do not matter. The founding study of generative engine optimization found that adding citations to credible sources, direct quotations, and specific statistics measurably raised whether a source was surfaced inside a generated answer. Earning citations is a real and worthwhile lever for presence. The precise, defensible claim is simply narrower than the marketing version: doing the work that tends to get you cited is sound; treating any single citation as proof that the model reasoned from you is not.

What an AI citation actually supports

If a citation is presence rather than proof, the response is to measure presence carefully and to stop claiming more than the evidence carries. Three states are worth keeping distinct, because collapsing them is where the overclaiming happens.

Where the evidence is settled, and where it is not

Because this domain is loud with confident claims, it is worth stating plainly which parts of the picture rest on solid ground and which are still moving. The table below separates them.

The pattern across the settled findings is consistent: a citation is a real, useful signal of presence and a poor proxy for causal reasoning or for traffic. Building content that earns citations is legitimate craft. Reporting a citation as evidence that a model relied on you, or guaranteeing a citation at all, runs ahead of what the research can support. On the AI-answer pillar, the credible position is the accurate one.