Citation authority: a practical guide to auditing AI sources

Before you start: define citation authority

A buyer can encounter your brand in an answer without visiting your website. That changes what a marketing leader needs to observe. It does not make organic rankings irrelevant.

Pew Research Center tracked 900 US adults in March 2025. Traditional results received clicks on 8% of visits to search pages with an AI summary, versus 15% without one; links in the summaries received clicks on about 1% of visits. This observational comparison does not isolate the effect of the summary from differences between searches.

I use citation authority to ask a narrower question: when an answer needs evidence about this topic, does it repeatedly reference a useful source we publish? A source can support a claim about a competitor, contain criticism or be misrepresented. A citation alone is neither endorsement nor proof of accuracy.

Step 1. Classify each appearance

Do this: Take one saved answer. Mark the brand mentions and source links separately; then classify every relevant source using the table.

Semrush’s June 2026 AI Visibility Index covered 126 million US prompts from January through April across four platforms and 22 industries. It distinguishes mentioned brands from cited domains; Gemini’s overlap could be as low as 30%. That is not agreement between engines, a ranking overlap or a probability that your page will be selected. Source: Semrush, data © Semrush.

For an expert’s personal site, an employer’s research remains affiliated evidence even when it sits on another domain. Record authorship, publisher and disclosed relationships. If ownership is unclear, mark it unknown. Open the cited page and check whether it supports the specific sentence beside the link.

Before moving on: Each source has an ownership category. Unknown relationships remain marked unknown.

Step 2. Inspect the cited claim

Do this: Open each source, find the supporting passage and record whether it supports the answer accurately. Keep platform observations separate.

A clear identity, a useful method and a maintained source page are assets you can reuse. Their value is not a guarantee that visibility transfers between ChatGPT, Gemini and Google’s AI experiences. Different questions, modes and dates can produce different sources.

Use patterns in your own captured answers to choose the next investigation. A high mention rate and low citation rate might reflect the questions asked, use of independent sources, a missing page, or an access problem. The rates identify a pattern to inspect; they do not establish its cause. Do not pursue Wikipedia, Reddit or another publisher simply because it appeared prominently in one aggregate study.

Before moving on: You can point to the evidence behind each assessment, including missing support.

Step 3. Choose one evidence improvement

Do this: Select a verified weakness in a page you can edit. Record the claim, missing evidence or identity conflict, and the correction required.

Start with the claim that matters to the buyer’s decision. Give it a source, date, population and limitation where relevant. Show how you reached the conclusion. Separate your interpretation from the underlying finding. A concise answer is useful only when it preserves those boundaries.

The GEO research paper by Aggarwal and colleagues tested content interventions in a benchmark and reported visibility improvements of up to 40% in that experimental setting. It offers hypotheses to test, not an expected lift on today’s live platforms. Generative Engine Optimization (GEO) here means work to improve how content can be discovered and used in generated answers.

Then check authorship and entity identity, fix contradictions and keep the substantive evidence accessible. Google’s AI-features guidance retains established SEO practices and does not require special AI markup. Useful editorial structure and accurate identity are maintenance decisions, not shortcuts to a universal trust score.

Before moving on: The proposed correction addresses an observed defect rather than a presumed platform preference.

Step 4. Record the baseline and next test

Do this: Calculate rates from the saved answers with the denominators shown below. Give the chosen correction an owner; preserve the baseline before applying it and keep the same panel for follow-up.

Verify that the existing panel represents actual audience decisions. Label new editorial questions as proposed rather than observed demand. Preserve a fixed comparison panel and record platform, mode, search setting, language, market, date, repetition, full answer and source URLs. Count a successful answer without your brand as a valid absence. Record failed requests separately.

These rates overlap and must not be added. A citation occurrence count is a different measure. Competitive share needs a declared competitor set and denominator, such as all recorded brand mentions in that set. None of these percentages is a percentage of topic ownership.

Review the context and accuracy of every relevant citation, then compare platforms separately. Repeat the same conditions and preserve changes to the panel. A before-and-after difference can also reflect platform changes or response variation. Track referrals and qualified enquiries separately in the campaign measurement workflow. Begin with a baseline you can inspect, not a score you cannot explain.

Before moving on: The register distinguishes answer rates, source quality and business outcomes.

Check your finished work

Copy-and-complete template

This is a working template, not observed results.

ANSWER RECORD, once per run
Question ID / run ID / repetition: [ ]
Platform / mode / search / language / market / date: [ ]
Status: [valid / failed / no overview]
Brand mention / domain citation / target URL citation / recommendation: [yes or no]
Complete answer record: [ ]
SOURCE RECORD, repeat for each relevant URL
Cited URL / publisher / relationship: [ ]
Claim and supporting passage: [ ]
Assessment: [supported / unsupported / unclear]
Proposed correction / owner / follow-up: [ ]

• Every relevant citation has a source URL and a checked supporting passage.
• Owned, affiliated, independent and unknown sources remain distinct.
• Rates show valid-answer denominators and excluded failures.
• The next test preserves the panel and does not promise causal uplift.