Measure citations of your publication in AI

Which questions belong in a publisher audit?

Choose questions that your journalism is equipped to answer. Test an investigation into municipal transport against relevant information needs. A question about the publication’s reputation measures a different task. Include evergreen explanations and timely stories in separate groups, because their useful life and expected updates differ.

Record why each question matters: audience feedback, search queries, subscription research or an editor’s documented hypothesis. Keep the exact question, language, model or interface and collection date with every answer. A publisher-name prompt and an audience-topic prompt measure different forms of visibility.

How do you count a citation fairly?

For a specific question in Rankfor, open Answer Trail and choose the available model configuration. Enter Your Prompt and the relevant optional brand or region, then inspect the completed response and its surfaced sources. Keep this custom audience panel distinct from a general brand scan.

Keep the original link and the resolved article URL where available. Map approved hostnames and obvious URL variants to the same publication, while retaining subdomains or licensed partner sites separately when ownership differs. Confirm that the citation URL resolves to your content. Record an unlinked mention of the publication’s name separately.

Start with the proportion of eligible answers containing at least one citation to your publication. Also report the number of distinct cited articles. If you calculate a share of all citations, define whether multiple links to one article or domain in an answer are deduplicated.

An illustrative sample has 30 eligible answers. Your publication is cited in 6 answers through 4 distinct articles. That supports an observed answer-level citation rate of 20% for this panel. It says nothing about 20% of all AI answers, audience reach or advertising inventory.

What should you check behind each link?

Read the surrounding claim and the linked page. Confirm whether the page supports the statement, whether the attribution is accurate and whether the article date fits the question. Research on citation verifiability distinguishes citation coverage from citation correctness; the mere presence of a link is not a complete quality assessment.

Compare models and languages separately before producing a combined figure. The public brand-sourcing preprint supports examining citation composition across markets and languages, but its sample is not a benchmark for every newsroom. A rival’s higher citation count does not by itself establish stronger domain authority or explain why the system selected it.

How does the report guide editorial work?

Classify an observed gap before commissioning another article. Your site may already contain the answer, the topic may be outside the publication’s remit, the page may need a factual update, or technical access may merit investigation. An absent sampled citation is not proof that the article is poor or unread.

For a suitable improvement, attach the editorial rationale, primary evidence, owner and review date. Avoid producing a near-duplicate story for every phrasing of the question. Google’s current guidance favors substantive useful content and warns against search-driven query-variant sprawl.

Repeat the same panel when assessing change, preserving new exploratory questions separately. Report new, retained and lost cited URLs, plus changes in coverage and collection conditions. The sampling-completeness preprint cautions that source discovery depends on the repetition horizon; a finite collection should never be presented as the complete list of sources a model uses.