An AI answer can mention your brand without linking to your website. It can also cite a page without mentioning the company behind it. These are different observations, and a useful citation report keeps them separate.
ProtoScrape’s Google AI Mode results provide answer text and available citation URLs and titles. Here is a practical way to turn those fields into evidence your team can review. The examples below are illustrative analysis methods, not measured customer results.
Keep the original observation
For each result, save the job ID, query, engine, country, language, collection timestamp, answer text and citation list. Keep any device setting from your original request alongside the result. Store the original URLs before making changes for analysis.
This record gives you a path back from a chart to the actual answer. It also helps explain a change later: a different prompt or market can change the meaning of the comparison. Export results before the approximately seven-day retention window ends if you need a longer history.
Separate mentions, citations and pages
Use three distinct measures. A brand mention means the answer text names the brand. A cited domain means at least one citation points to that domain. A cited page means a particular URL appears in the citation list.
Count a domain once per answer when calculating a domain citation rate. If one answer links to three pages from the same domain, counting three domain appearances would overstate how many answers included it. Retain the individual URLs for page-level analysis.
An empty citation list means no citations were returned for that observation. It does not establish that no external information influenced the answer. Citation order also should not be presented as an organic search ranking or a measure of source importance.
Normalize URLs without losing evidence
Create derived fields for hostname, path and your chosen domain grouping. Lowercase hostnames, but do not lowercase URL paths indiscriminately: paths may be case-sensitive. Decide whether www.example.com and example.com belong in the same reporting group and document that decision.
Remove tracking parameters only from a derived analysis field. Some query parameters identify different content, so deleting every parameter can combine distinct pages. Likewise, grouping every subdomain together can hide differences between a publication, support center and customer community.
Report a rate with its denominator
Suppose a fictional study schedules 20 observations. Eighteen succeed, two fail, and six of the successful answers cite a particular domain. Its observed citation rate is 6 / 18, or about 33%. Report all three counts, including the two failures.
Do not silently put failed jobs into the denominator as answers without citations. Do not remove successful uncited answers either: those are valid observations. When comparing time periods, keep the query set and collection settings consistent and show how many successful observations each period contains.
Use citations to choose what to review
A recurring source can suggest a useful next question: what information does that page provide, and is the same information missing or unclear in your own material? Read the cited page and answer together before recommending a content change. A citation pattern alone does not prove that one page feature caused inclusion.
Start with one engine and a stable collection schedule. Cross-engine analysis becomes useful only when each engine is available and its sample is comparable. Currently, Google AI Mode is available in ProtoScrape; Google Search, ChatGPT and Perplexity remain in validation. See the measurement guide for a broader reporting framework.