MadRaven Lab / Field Notes
How to Verify AI Citations and Recommendations in a GEO Report
Direct Answer
Request the full question, original answer, actual citation URLs, search setting and sampling time. A brand mention is not necessarily a citation. A link to the website is not necessarily a recommendation. A recommendation is not an enquiry. Each claim needs its own evidence and denominator.
Separate Five Outcomes
| Claim | Minimum Evidence | Unsupported Inference |
|---|---|---|
| Brand mentioned | An identifiable brand name in the original answer | The website was used as a source |
| Website cited | A checkable URL on the domain in the answer's sources | The source was interpreted accurately |
| Brand recommended | Explicit presentation as a suitable option | Suitability for every customer |
| Search exposure grew | Relevant official report, dates and metric | Equivalent citation growth across platforms |
| Enquiry generated | Actual lead and source record | Attribution from a brand screenshot alone |
Keep a Reproducible Sample Record
Record a sample ID, platform and visible model, timestamp and timezone, language and region, web-search setting, complete prompt, new-conversation status, original answer, every source URL and conclusion. When the platform does not reveal a model or region, record unknown.
An unbranded question might ask how an Australian business should compare bilingual website proposals. A named question might ask what MadRaven Lab's website says about bilingual services. Keep those groups separate. Mentioning a brand that was already included in the prompt does not demonstrate an unsolicited recommendation.
Verify What Each Citation Supports
Open the cited page. Check that it exists, belongs to the expected domain, supports the nearby claim and has applicable information. If a page describes website design while the answer invents a customer's revenue improvement, the citation does not support that result.
Google's documentation states that its AI search features use foundational SEO eligibility and helpful content; satisfying requirements does not guarantee crawling, indexing or appearance. Keep robots, schema and answer-format checks separate from observed citations.Google AI features guidance
Compare Fixed Samples Over Time
Keep at least three independent new conversations per question and platform, noting search settings. This is a starting protocol proposed here, not evidence of statistical significance. Repeat at T+7, T+14 and T+30 with the same questions, language and settings. Record model or feature changes rather than attributing every difference to the website.
If two of six auditable samples cite the domain, report “2 of 6 samples cited the domain.” Do not turn that into a claim that AI has a 33% probability of recommending the brand. Explain whether failed requests or search-disabled samples are in the denominator. A cropped screenshot without its original answer should not count as a verified citation.
Connect the Evidence to Business Records
Keep lead sources, landing URLs, language and deduplicated IDs. Ask actual prospects how they discovered the service, recording self-reports separately from tracked sources. An unidentified source remains unknown. A conversation link may support a lead's account; it cannot replace records of consultations, agreements or received revenue.
For a GEO report review with MadRaven Lab, bring original answers, questions, source URLs, current website content and enquiry records. Establish which conclusions are evidenced and which describe technical readiness or untested hypotheses. There is no guarantee that a platform will cite the site.
Official source checked 10 October 2026. The sampling protocol is a recommended method, not a platform rule or outcome promise.