Measure AI citations.
Keep the evidence honest.

By CitedLoop · Updated 2 October 2026

AI citation share describes how often a set of sampled answers cites your domain. It is not market share, a search ranking or proof of revenue. This guide explains how to measure it and how to make your expertise easier for people and search systems to find.

Start with a useful answer, an accessible page and a measurement plan. No special schema, file or backlink package guarantees inclusion in an AI answer.

1. Define the test before collecting answers

1. Choose real buyer questions. Use customer conversations and research. Separate branded questions, such as “CitedLoop reviews”, from unbranded category questions. Do not count an echoed brand name as evidence of discovery.
2. Freeze the comparison panel. Keep questions, provider, model, search settings and run counts comparable. Record timestamps and failures. If settings or questions change, label the comparison exploratory rather than equivalent.
3. Count sources, not impressions. Inspect source URLs and require an exact hostname match or an intentional subdomain match. A URL that merely contains a brand string is not enough. Track linked citations separately from unlinked mentions.
4. Keep the denominator visible. Report successful answers, cited answers and failed requests. Do not silently treat provider failures as uncited answers or hide excluded trials.
5. Verify the intervention. Save the baseline, the actual published URL, the content change and its publication time. A mock publication on another host is not a treatment of your main website.
6. Choose a follow-up window in advance. Allow for crawl and indexing delays, but do not promise a fixed time to visibility. Avoid repeatedly testing until one comparison happens to pass a significance threshold.

2. Show the uncertainty

Citation share = cited answers ÷ successful answers. A 95% Wilson interval expresses sampling uncertainty under a binomial model. Repeated AI responses can be correlated; this interval does not capture every source of model, question or engine variability.

Illustration, not a customer result: 0 citations from 16 successful answers gives a 0% observed share, with an approximate 95% Wilson interval of 0%–19.4%. A small sample cannot establish that the true rate is zero.

For before/after work, a two-proportion test has assumptions, including sufficiently informative samples and independent observations. Small counts, repeated questions and correlated responses may call for different methods or specialist statistical advice. Significance is not causality, and a non-significant result is not proof that nothing changed.

3. Make useful content accessible

• Answer a real decision. Explain what your product does, who it suits, limitations, setup requirements and verified compatibility. Avoid publishing hundreds of thin, near-identical pages.
• Build pages worth referencing. Publish original tutorials, reproducible methods, technical documentation and carefully labelled research. Include the underlying evidence and dates. Do not invent customer results or expert authors.
• Make discovery technically possible. Serve public pages successfully over HTTPS, use clear internal links and canonical URLs, maintain a sitemap, and check accidental noindex directives or blocked crawlers.
• Review crawler access deliberately. OpenAI says allowing OAI-SearchBot supports discovery in ChatGPT search. Check both robots.txt and hosting/security rules. Search crawling and model-training preferences are distinct decisions.
• Make FAQs useful to humans. Answer genuine questions in plain language, with enough context to be understood. If using FAQPage markup, keep it consistent with the visible questions and answers. Do not expect Google FAQ rich results: Google retired that feature in 2026.
• Keep identity consistent. Use the same real business name, product description and contact information across your website and profiles you control. Add only verified organisation details to structured data.
• Make the page easy to use. Provide readable text, meaningful headings, keyboard-accessible controls and mobile navigation. Content must remain readable if animation or JavaScript fails.
• Measure visits and outcomes separately. Track genuine referral traffic and conversions as well as sampled citations. A citation-rate change is not itself an ROI calculation.

Google says there is no special structured data required for generative AI search and that it ignores llms.txt for Search. Focus on content quality and standard search fundamentals rather than treating either as a ranking shortcut.

4. Earn links instead of manufacturing them

A backlink is a link from another website to yours. Adding outbound links here does not create backlinks to CitedLoop. Editorial placements depend on other publishers deciding that a resource is useful.

1. Offer this measurement checklist and calculator as a practical resource for articles that discuss small-sample AI visibility tests.
2. When a real customer study is ready, publish the question panel, dates, sample sizes, limitations and permissioned results—including null results. Invite independent review rather than prescribing a positive conclusion.
3. Help relevant communities with a complete answer first. Disclose your affiliation and share a link only when it genuinely adds value and the community permits it.
4. Keep partner or product listings factual. Do not claim platform partnerships, reviews or customer endorsements that do not exist.
5. Do not purchase ranking links, exchange links at scale, post automated comments or make fabricated forum recommendations. Sponsorships need appropriate disclosure and link qualification.

5. Use a release checklist

• Page answers the intended buyer question, with sources and current product limitations.
• Published HTML, canonical URL and internal links work.
• Visible content agrees with structured data.
• Crawler and security rules have been checked.
• Baseline, comparison settings and publication timestamp are recorded.
• Result can honestly be “no detectable lift”; no invented benchmarks or testimonials.

Download a blank citation-test log (CSV) →

Sources and further reading

The operational advice above is a CitedLoop synthesis, not an endorsement by these organisations.

• NIST: confidence intervals for proportions — the Wilson method and assumptions.
• Google: optimising for generative AI features — search foundations and common misconceptions.
• OpenAI: publishers and developers FAQ — OAI-SearchBot access and discovery.
• Google Search documentation updates — retirement of FAQ rich results in 2026.
• Google: link spam policies — avoid manipulative link acquisition.

More CitedLoop questions → · Ask about early access