Research note · Research note 02 — current edition

The Anatomy of a Source-Linked AI Answer

Independent Content Strategist & Product-Marketing Writer · Published Feb 17, 2026

Primary audience: Product managers building AI features
Also useful for: AI engineers and technical founders
Tone: Authoritative
Reading time: 2 min
Published: Feb 17, 2026

Abstract

A source-linked answer is not an answer with links appended. This note breaks a well-formed answer into its parts — claim, support, provenance, confidence, and refusal — and argues that the structure of the answer object matters more than the wording of the prose.

Key questions

What is the smallest unit that should carry a citation?
How should an answer represent a claim it cannot support?
What does a reader actually verify when they click a citation?
Where should confidence live: in the prose, the data, or the interface?

How this note was developed

— Decomposition of answer formats used across public AI research and search interfaces.
— Prototype construction of an answer object with per-claim provenance.
— Review of what a reader needs in order to verify a claim in under thirty seconds.
— Analytical note; no external validation or user study is claimed.

Central arguments

01The sentence is the right citation unit

Paragraph-level citation lets an unsupported sentence hide between two supported ones. Sentence-level attachment is more work to produce and dramatically easier to audit, because a reader can check one claim without reading the whole source.

02A citation should point at a passage, not just a page

A link to a 6,000-word page shifts verification cost onto the reader. Carrying the supporting excerpt alongside the URL makes verification a glance rather than a search, and makes it possible to detect later when the page no longer contains that passage.

03Refusal is part of the output schema

An answer object needs a place to record 'not established by available sources'. If the schema has no such field, the model will fill the gap with prose, because prose always has room for another sentence.

04Confidence belongs in the data, expressed in the interface

Hedging language is unparseable and inconsistently interpreted. A structured confidence value with a stated basis — number of independent sources, agreement, source age — lets the interface present uncertainty consistently.

Limitations

— Per-sentence provenance increases cost and latency, which some products will reasonably reject.
— Excerpt storage has licensing implications that vary by jurisdiction and source.
— Confidence scores can be presented in ways that imply more rigour than the underlying basis supports.

citationsprovenanceanswer design