How LLMs Cite: The Attribution Mechanism Study (arXiv 2606.28358)

The paper is "How Do LLMs Cite? A Mechanistic Interpretation of Attribution in Retrieval-Augmented Generation" (arXiv 2606.28358), listed in our evidence base as ECIR 2026 work. It asks a narrow question: inside a model that answers from retrieved documents, which internal parts decide whether a source gets attributed?

The model studied is Llama-3.1-8B, an open-weight model. Mechanistic work needs open weights, because the researchers have to inspect and change internal components directly. That is why the study is not run on ChatGPT, Claude or Perplexity.

The setting is Stage 3 of the four-stage model: generation and citation. The documents are already in the model's context. The study says nothing about how they got there.

Mechanistic interpretability treats a model as a circuit. Researchers trace which components contribute to a behaviour, then intervene on them to test the trace.

Our evidence records the intervention at the component level:

• The researchers identified attention heads and MLP layers involved in producing citations.
• They amplified the contribution of specific heads and MLPs.
• They held the sources constant. The retrieved documents did not change.
• They checked whether answer accuracy suffered.

The last two points carry the weight. If citations change while the inputs and the answer quality stay the same, citation behaviour is something the model does separately from using the source.

The paper describes citation as produced by a "distributed, multi-stage 'attributional ensemble'". No single component decides it.

The headline numbers, quoted as the evidence base gives them:

• Amplifying specific heads and MLPs "repairs over 90% of missed citations".
• The same intervention "eliminates 69% of spurious ones".
• It did so "without harming answer accuracy".
• The sources were not changed.

Read those together. A missed citation is a source the model used but did not credit. A spurious citation is a source credited that did not support the answer. Both can be fixed by changing the model's internal weighting, not the evidence. So citation emission is mechanically separable from source influence.

This fits what platforms document. OpenAI's API docs say "the number of sources is often greater than the number of citations." Anthropic cites short text spans (a cited_text field of up to 150 characters), while Perplexity cites whole sources by integer index. Those are different units, which is why Claude and Perplexity cite differently. It also helps explain why AI citations sometimes do not support the claim.

The study is strong on mechanism and narrow in scope. Keep these limits in view.

• One model. Llama-3.1-8B is a small open-weight model. Nobody has shown the same components behave the same way in the frontier models consumer apps serve as of October 2026.
• No live web. Sources were supplied in context. Like nearly every GEO effect size, the finding sits at Stage 3. The critical survey of 45 GEO studies (arXiv 2607.14035) found no technique with a stable, cross-platform causal effect on discoverability.
• No publisher lever. Amplifying attention heads is something only a model operator can do. No page edit reaches it.
• Not a ranking study. It does not say which pages get retrieved. Retrieval and position in the context list are separate questions.

One inference the evidence base draws: citations likely undercount influence, and the undercount is unlikely to be random with respect to how well-known a source is. That is a reasoned caution, not a measured result.

The practical conclusion is about measurement, not optimisation. If you report AI visibility, treat citation counts as a noisy signal.

1. Count retrieved and cited separately. OpenAI's Responses API, Anthropic's web search tool and Perplexity's Agent API all expose the retrieval set apart from citations. A source retrieved but not cited may still have shaped the answer.
2. Never treat one citation count as proof of use. A citation shows the model credited a source. It does not show the source drove the answer.
3. Never treat a missing citation as proof of non-use. The study repaired over 90% of missed citations without touching the sources. Missed credit is a model behaviour.
4. Report per engine. Span-level and source-level citations are not the same unit. Do not add them into one number.
5. Disclose retrieval state. An answer with no visible citations may come from the model's training data, not the web.

What stays in your control is upstream: whether crawlers can fetch your pages, whether key facts are in the HTML before JavaScript runs, and whether specific facts are stated plainly enough to quote. For how the other studies fit, see what the studies show on LLM citation and the definition of an AI citation.

Can I change how an LLM attributes my page? Not through anything in this study. The intervention amplified attention heads and MLPs inside the model, which only the operator can touch. What you control sits earlier: crawler access, server-rendered content, and index eligibility. Those decide whether your page can reach the model's context at all. Once it is there, the evidence shows citation behaviour is driven largely by the model and by context position, not by on-page tactics.

Does a citation mean the AI used my content? Not necessarily. The study shows a model can cite a source that did not support the answer, and fail to cite one that did. It removed 69% of spurious citations and repaired over 90% of missed ones without changing the sources. A citation is evidence of credit, not proof of use. Pair it with the retrieved-sources list where the API exposes one.

Why do engines retrieve more sources than they cite? Retrieval and citation are separate steps. OpenAI's documentation states that the number of sources is often greater than the number of citations. The model reads a candidate set, then emits citations through its own internal process. This study shows that process is distributed across components and can drop credit for sources the model used, so the gap between retrieved and cited is expected.

Do these results apply to ChatGPT, Claude or Gemini? Unknown. The work was done on Llama-3.1-8B, an open-weight model, because the method needs direct access to internal components. Closed production models cannot be studied this way from outside. The finding that citation is separable from use is a reason for caution with any engine's citation counts, but the specific percentages should not be read as properties of other models.

Could someone game citations with hidden text instead? Injected or adversarial content does affect AI answers in lab studies, and it is illegitimate. Hidden instructions and cloaking for AI crawlers risk penalties and reputational damage. The defensive point matters more: third-party pages can misrepresent a brand. See the research on [LLM search agents and manipulated web content](/research/llm-search-agents-content-manipulation-study) for what that study measured and its limits.