Data · one engine (Claude + web search)
What AI answers actually cite
Real numbers from a live measurement panel — not estimates. Every figure on this page is generated from committed per-query data, and the raw citations are published. As of Jul 2026.
85%
of domains cited once
395 of 462
6.1
citations per answer
610 total
462
unique domains
across 100 queries
100
queries measured
commercial + GEO/SEO
85%of cited domains appear for just one query
Citations per answer, by topic
The number of sources an AI answer cites barely moves across subjects — but the domains do. Authority in one topic does not transfer to another.
| Topic | Queries | Avg citations/answer |
|---|---|---|
| technical-seo | 15 | 6.7 |
| marketing | 17 | 6.5 |
| content | 6 | 6.3 |
| geo | 24 | 5.9 |
| software | 20 | 5.8 |
| ai-tools | 12 | 5.8 |
| ecommerce | 6 | 5.8 |
Frequently asked
- What share of domains cited by an AI answer engine appear only once?
- In a 100-query panel run through one engine (Claude Sonnet 4.6 with web search), 85% of cited domains — 395 of 462 unique domains — were cited for only a single query. Citation is re-decided query by query, not a fixed leaderboard.
- How many sources does an AI answer cite on average?
- 6.1 citations per answer, across 100 queries (610 citations in total). Every answer in this run cited between 3 and 9 sources — a narrow, stable band.
- Which domains are cited most often by AI answers?
- semrush.com (cited in 12 of 100 queries), then medium.com (cited in 7 of 100 queries), followed by ahrefs.com, arxiv.org, blog.hubspot.com. No domain was cited in more than 12 of the 100 queries.
- Does the topic change how many sources an AI answer cites?
- Barely. Average citations per answer ranged from 5.8 (ecommerce) to 6.7 (technical-seo) across topics — the count is stable even though the specific domains cited are almost entirely different between topics.
- What engine and method produced these numbers?
- One engine — Claude Sonnet 4.6 with web search — run over 100 commercial and GEO/SEO queries. Every citation is captured verbatim from the model's answer (its web_search_result_location data). This is a single-engine pilot, not a cross-engine census; the raw per-query citations are published. Generated Jul 2026.
Method & provenance
100 queries · one engine (Claude + web search) · Jul 2026. Each query is sent to the model with web search enabled; we capture the exact URLs it cites, deduped by URL and ranked by first appearance — the same pipeline the live GEO Trace tool uses. This is a single-engine pilot (N = 100), not a cross-engine census; a second engine is being wired in.
