A Cited-Source Taxonomy for AI Answers
Classify every cited URL into one fixed type, and write down the rule you used: your own site, a direct competitor, a review or comparison site, a publisher, community or forum content, video, reference sources such as encyclopaedias and standards bodies, marketplaces and integration directories, and other. Settle the ambiguous cases before you start. A competitor's blog post comparing you is 'competitor', not 'comparison'; a sponsored listicle on a news site is 'publisher', flagged as sponsored. The output worth having is a count of citations by type across your prompt set, which shows where the answers in your category draw from and which of those sources you could realistically appear in. It tells you where to look, not why any source was chosen.
After a few weeks of sampling you'll have hundreds of cited URLs and no idea what they add up to. A list of domains sorted by count is hard to act on. A list sorted by type answers a better question: what kind of source do assistants reach for when your buyers ask these questions?
The nine types
| Type | Includes | Ambiguous case and the rule |
|---|---|---|
| Own site | Every property you control, subdomains and docs included | A partner page about you is not yours |
| Competitor | Every property of a direct competitor | Their comparison post about you stays 'competitor' |
| Review or comparison | Software directories, independent 'best of' lists | A vendor-owned list goes under that vendor |
| Publisher | News and editorial sites | Sponsored pieces stay here, flagged sponsored |
| Community or forum | Q&A sites, forums, discussion threads | Classify by platform, not by the poster |
| Video | Video platforms | A transcript page on a publisher site is 'publisher' |
| Reference | Encyclopaedias, standards bodies, government, academic | A vendor's glossary is not reference |
| Marketplace or directory | App stores, integration and partner listings | Your own listing is still 'marketplace' |
| Other | Anything that fits none of the above | If 'other' passes one in ten, add a type in a new version |
The log columns
cited_url,domain,type,sponsored,names_our_brand,prompt_id,surface_id,run,date
https://reviews.example/best-scheduling-tools,reviews.example,review_or_comparison,no,yes,P07,S-03,2,2026-09-29
(illustrative row)The names_our_brand column matters more than it looks. A review site that names you is a citation to them and a mention of you, and it's the most common way a brand appears in an answer without a single one of its own URLs being cited.
Reading the counts
An illustrative month: 180 citations across 300 sampled answers. Review and comparison sites 40, competitors 32, publishers 30, community 28, your own site 16, reference 14, marketplaces 12, video 5, other 3. Read each row as a question. Are you on the review sites being cited, and is what they say accurate? Which publisher pieces keep appearing, and are they current? Community threads can't be written for you, and shouldn't be; answering questions honestly under your own name is the only version that survives. Competitor citations are information, not a task.
What it can't tell you
Why a source was chosen. A high count for one type describes your sample on your surfaces in your window; it isn't evidence about how any assistant ranks sources, and it can change when a vendor changes its product. For your own pages there's now first-party data to set beside it: Bing Webmaster Tools' AI Performance report shows page-level citation counts for your site across Microsoft's AI surfaces. It won't show anyone else's URLs, which is what the taxonomy is for.
FAQ
Should subdomains roll up to the parent? — For your own and competitors' properties, yes: docs.competitor.example is the competitor. For platforms hosting many authors, classify by the platform's type, not the author's.
What about a page that names us but belongs to someone else? — Classify it by owner and tick names_our_brand. That's the mention-versus-citation distinction applied to sources: their citation, your mention.
How often should the taxonomy change? — Rarely. Adding a type breaks comparison with earlier months, so version it the way you version a prompt set, and recount old rows if you can.
Put this to work on your own website.
Fenn finds what your customers ask, drafts the articles and site fixes, and measures what ChatGPT, Claude, Gemini, Perplexity and Grok say about you — with every change waiting for your approval.