What is source attribution in AI?

Source attribution is the link between a statement an AI system generates and the source that supports it. It answers one question: which document says this?

Three terms get mixed up here:

• Retrieval is the system fetching documents into the model's context.
• Use is the model actually drawing on a document to write the answer.
• Attribution (or citation) is the model marking a document as the support for a claim.

These can diverge. A page can be retrieved and never cited. A page can be cited for a claim it does not support. For the full picture of that gap, see cited vs retrieved sources.

In retrieval-augmented generation, the engine retrieves candidate documents, the model writes an answer, and the model emits citation markers alongside the text. The citation is produced by the model during generation. It is not a lookup the engine performs afterwards.

The format differs by platform, as of October 2026:

• Anthropic's web search tool for the Claude API cites text spans, with a cited_text field of up to 150 characters.
• Perplexity cites whole sources by an integer index.
• OpenAI's web search tool returns the full list of sources consulted separately from citations, and its docs say the number of sources is often greater than the number of citations.

The mechanism inside the model is the important part. A mechanistic study of Llama-3.1-8B (arXiv:2606.28358) found citation comes from a distributed, multi-stage "attributional ensemble" of components. Amplifying specific attention heads and MLPs repaired over 90% of missed citations and eliminated 69% of spurious ones, without changing the sources and without harming answer accuracy. Citation emission is partly decoupled from source use. The attribution mechanism study covers the method in depth.

Most AI visibility reporting counts citations. Attribution research says that count is a noisy proxy for influence, and the noise is unlikely to be random.

What this means for you:

• A page that shaped an answer may not appear as a citation.
• A cited page may not support the claim next to it. That is why AI citations sometimes don't support the claim.
• Citation units differ (spans vs whole sources), so citation counts from different engines don't add up cleanly.

The next move is to work on what you control. Citation happens at the generation stage, which the model controls. Before that, your page has to be reachable by the crawler and in the index, and the facts on it have to be in the HTML. The guide to how AI search picks sources walks through each stage. The free checker on this site shows whether AI engines name you for three buyer questions, which is a snapshot of what the answers say, not a measure of influence.

• "Cited means used." The mechanistic study shows citations can be repaired or removed without changing the sources. Citation and use overlap but are not the same.
• "A citation proves the claim." Attribution can point to a source that does not back the statement. Check the source.
• "You can rewrite a page to get cited." A critical survey of 45 GEO studies (arXiv:2607.14035) found citation-oriented rewrites can impair retrieval, and no technique showed a stable cross-platform effect on discoverability.
• "All engines cite the same way." They don't. Span-level and source-level citation measure different things.

• Faithfulness: whether an answer's claims are supported by its retrieved context. RAGAS scores it as supported claims divided by total claims.
• Retrieval-augmented generation: the pipeline in which attribution happens.
• Cited vs retrieved sources: why the retrieved set is larger than the cited set.
• How Perplexity chooses sources: one engine's retrieval and citation behavior.

Is source attribution the same as marketing attribution? No. Source attribution links an AI answer's claims to documents. Marketing attribution links a lead or sale to the channel that produced it. They meet when an AI citation sends a click to your site, but most AI answers produce no click at all. For the traffic side, see whether you can [attribute leads and sales to AI search](/questions/attribute-conversions-ai-search).

Can I make an AI engine cite my page? Not directly. Citation is decided by the model at generation time, and nobody outside the operator controls that step. What you control comes earlier: letting each engine's search crawler in, keeping your page indexable, and stating key facts in HTML that loads without JavaScript. Those remove blocks. They don't guarantee a citation.

Why do some AI answers cite a source that doesn't say what the answer says? Because citation emission is partly separate from source use. The 2026 mechanistic study found specific model components govern whether a citation appears, independently of the sources themselves. The model can attach a citation to a nearby or loosely related document. Always open the cited page before trusting a specific claim, especially prices, numbers and dates.

How do researchers measure attribution quality? Most metrics divide supported claims by total claims. FActScore breaks a response into atomic facts and scores the share supported by a knowledge source. RAGAS faithfulness scores the share of claims supported by the retrieved context. Both normalise by a count of claims rather than summing errors, so scores stay comparable across answers of different lengths.