Research Note: A Repeatable AI Citation Study

Research Note: A Repeatable AI Citation Study

Question

What must a small publisher preserve so that observations about AI citations can be compared over time?

What the sources establish

The 2025 paper News Source Citing Patterns in AI Search Systems analyzed more than 65,000 responses and 366,000 citations from an evaluation arena. It found differences in the news sources cited by systems from OpenAI, Perplexity, and Google, alongside shared concentration patterns. Its scope is news-oriented conversations collected during a dated study window. It supports the need to identify the system and corpus, not a permanent personality label for each product.

The peer-reviewed 2026 paper Auditing Citation Behavior in AI-Generated Search Summaries models retrieval and citation as observable processes over query-document pairs. Its case study concerns Google AI Overviews and a defined set of Your Money or Your Life queries. It supports a measurement approach that preserves query, retrieval position where observable, source provenance, and citation outcome.

The 2026 preprint From Citation Selection to Citation Absorption separates whether a system selects a source from whether the final answer appears to use its language, evidence, structure, or factual support. That distinction is analytically useful, but the paper remains a preprint and its proposed absorption measures should not be presented as an industry standard.

Google's generative-AI Search Console announcement introduces a first-party visibility measure for a limited subset of properties. It reports where a site's URLs appeared in Google's generative features. It does not expose every prompt, retrieval trace, answer passage, or cross-platform citation.

Disagreement and uncertainty

AI answer products can vary with the prompt, conversation history, location, language, account state, product tier, model, retrieval index, date, and interface experiment. Some products do not expose the model or retrieval trace. A repeated result may still change even when the tester preserves every visible condition.

There is no evidence-backed universal interval for repeating a small citation study. A monthly baseline can be an operating choice, not a scientific finding. Material product changes, major site releases, and reporting changes justify additional runs.

An observed citation cannot establish why the system selected a page. It does not reveal training inclusion, hidden weights, or causation. A brand mention without a link, a linked citation, an answer passage that uses the page, a search impression, a referral visit, and a conversion are separate events.

Editorial use

The guide should define one observation as one prompt submitted to one named product under one preserved environment at one time. The minimum record should include prompt identifier, exact prompt, topic and intent, system, visible model, location, language, login state, date and time, conversation state, answer capture, every cited URL, link placement, brand mention, apparent answer use, and any referral or first-party visibility evidence available later.

The public schema should use rows rather than prose notes so runs can be compared. The guide can recommend a baseline run and a monthly operating cadence while plainly labeling the cadence as a practical default.

Automation should occur only through a supported interface and within the relevant terms. Manual tests remain valid when their environment and limitations are recorded.