Answer attribution

The Citead team, updated

In short. Answer attribution links claims in an AI answer to supporting source passages so readers can check the evidence.

Definition: Answer attribution connects claims in an AI assistant's answer to supporting source passages. A citation identifies the evidence associated with a claim so that a reader can check whether that evidence supports it.

Key facts

  • The ALCE paper, published in December 2023, defines a task in which a system retrieves passages, answers a question, and cites supporting passages for its statements.
  • Liu and colleagues' verifiability study distinguishes citation recall, the coverage of answer statements by supporting citations, from citation precision, the support supplied by individual citations.
  • As of 17 October 2026, Google's search grounding documentation describes inline citation annotations that link specified ranges of answer text to source URLs.

How does ALCE connect statements to evidence?

ALCE stands for Automatic LLMs' Citation Evaluation, where LLM means large language model. In Enabling Large Language Models to Generate Text with Citations, the system receives a question and a collection of text passages. Its answer consists of statements, each associated with a list of cited passages. Several passages can support the same statement.

The paper generally uses sentence boundaries to identify statements. For QAMPARI, a dataset whose answers are lists, each answer entity serves as a statement. The attribution unit therefore depends on the answer's structure.

ALCE divides its source collection into 100-word passages to make human verification easier and fit more retrieved passages into the model's available input space. It evaluates answer correctness separately from citation quality. Correctness concerns the accuracy and coverage of the answer; citation quality concerns support from the passages actually cited.

What does an attribution annotation look like?

Google's grounding guide provides a concrete representation: a url_citation annotation attached to the answer's text content block. The annotation includes a source URL and indices identifying the associated text segment.

Selected fields from the guide's sample annotation are:

{
  "type": "url_citation",
  "start_index": 0,
  "end_index": 56
}

The full sample also supplies url and title fields. The application can use these annotations to display inline citations. The indices locate the text being attributed, while the URL identifies the referenced source. This makes the relationship between a particular answer segment and a source explicit enough for an application to render it.

How can attribution be checked for a website?

The verifiability study's evaluation method supplies a practical review procedure: identify the answer statement, open its associated citations, and compare the statement with their contents. Its annotators classify each citation as providing full, partial, or no support. They also check whether multiple cited pages collectively support the complete statement.

For a website review, this means checking the exact claim attached to the site's link against the relevant passage. A sentence can contain independently verifiable claims, and a citation placed after several statements can have an ambiguous scope, as the study explains.

ALCE's automated evaluation uses natural language inference, a test of whether evidence entails a statement. It combines cited passages when checking statement support and separately detects irrelevant citations. A manual review can follow the same distinction: assess coverage across the answer, then assess the contribution of each citation.

FAQ

Can several sources support the same claim?

Yes. ALCE explicitly allows multiple passages for a statement and evaluates their combined support. When reviewing an answer, you can read those passages together before judging whether the whole claim is supported.

How do I distinguish partial support from full support?

In the verifiability study, full support covers all information in the associated statement. Partial support covers some information while other parts remain missing or contradictory. You can apply that distinction to the complete sentence associated with a citation.

Sources

  1. Association for Computational Linguistics, "Enabling Large Language Models to Generate Text with Citations", December 2023
  2. Association for Computational Linguistics, "Evaluating Verifiability in Generative Search Engines", December 2023
  3. Google, "Grounding with Google Search", updated 13 October 2026

Check how a page reads to answer engines. Open the GEO score analyzer

Test a page and see what to fix first

The analysis usually takes less than 10 seconds and needs no account.

400,000+ pages scored in the public benchmark