By Curtis Pyke · Published October 5, 2026 · KACI Wave One v2026.10.05
We asked OpenAI, Claude and Gemini the same 1,200 questions twice via their official APIs. Across 7,200 completed answers, their average overlap in cited website domains was about 6%.
The average domain-set Jaccard was 0.061 for OpenAI/Claude (95% interval 0.053–0.070), 0.060 for OpenAI/Gemini (0.053–0.069) and 0.065 for Claude/Gemini (0.056–0.074). The respective averages include 1,718, 1,720 and 1,293 eligible pairs out of 2,400. A score compares shared domains with the union of both domain sets; it is not the fraction of questions sharing a source. Pairs where both answers cited no domain are excluded.
Citation coverage and changing source lists
OpenAI’s two-run mean domain overlap was 0.664 (95% interval 0.645–0.682), Claude’s 0.515 (0.494–0.537) and Gemini’s 0.319 (0.292–0.345). These averages include 867, 652 and 337 eligible questions out of 1,200 per provider. Higher consistency does not establish greater accuracy.
Different domains led the rankings
GitHub led OpenAI, arXiv led Claude and Reddit led Gemini by citation occurrences. These post-hoc rankings include repeated annotations and are exploratory; they do not estimate traffic or the effect of a publishing strategy.
Source types in the reviewed subset
The reviewed domains account for OpenAI 5,998 of 11,738 citation occurrences (51.1%), Claude 4,294 of 18,327 citation occurrences (23.4%), Gemini 7,511 of 15,734 citation occurrences (47.7%). Repeated native annotations are retained. Review coverage differs by provider; category shares within this selected subset do not describe all cited websites.
YouTube citations and cached metadata
Gemini cited YouTube in 187 of its 2,400 answers, OpenAI in one, and Claude in none. YouTube was 3.299% of Gemini’s citation occurrences (95% interval 2.659%–4.024%); the answer-presence metric has a different denominator. An observed zero from Claude does not establish that it never cites YouTube, and a citation does not establish watching a video.
AI Citation Explorer
Look up a domain’s rank, citation share and number of answers citing it. The Explorer also shows top-50 tables for each provider and question group.
Open the Explorer at full size. The standalone file makes no external requests.
Method in five sentences
We asked the same designed panel of 1,200 English questions twice through the official OpenAI, Claude and Gemini APIs, with no conversation history or added system prompt. The returned models were gpt-6.1-sol, claude-sonnet-5-5 and gemini-3.8-flash, using native search and their default reasoning, temperature and automatic search choices. Collection used OpenAI Flex, Claude Messages Batches and standard Gemini Interactions, with documented output and tool limits. We counted answer-linked native citations and compared registrable-domain sets, retaining zero-citation completed answers but excluding empty unions from Jaccard averages. Intervals use 10,000 question-cluster bootstrap samples within the four planned strata, retaining both runs and providers; collection took 13.814 hours including queues and interruptions.
Limitations
These official API results can differ from consumer ChatGPT, Claude and Gemini apps. The English question panel was designed for this study, not sampled from user demand, and two observations in one collection window cannot describe all future behavior. Changing search indexes, batch scheduling and gaps of up to 348.7 minutes between matched archives may explain part of the difference. Native annotations do not independently verify that a source supports an answer.
Human review covers the selected top 100 of 4,631 domains; the other 4,531 remain provisional. Category comparisons describe only this frequently cited subset. The remaining domains are pooled as unreviewed, so this review does not establish category preferences across the web.
Public metadata is cached for all 239 unique cited video IDs. Views and other mutable fields describe retrieval after collection. Captions are availability flags; neither they nor a citation establish that a provider watched a video. Missing Google response/request IDs limit provider-side reconciliation, and invoices remain unreconciled.
Downloads
Data · Domain tables · Question panel · Statistics · Methodology · Limitations · Data dictionary · Validation · Checksums · Evidence ZIP
Cite this release
Pyke, Curtis. Kingy AI Citation Index (KACI), Wave One. Version v2026.10.05. Published October 5, 2026. Kingy.ai. DOI: 10.5281/zenodo.23170591. Zenodo evidence record. CC BY 4.0 attribution applies to Kingy-authored analysis; third-party material retains its applicable rights and terms.
Disclosure
Curtis Pyke commissioned and funded KACI for Kingy.ai. Codex assisted with collection, analysis and drafting. Kingy.ai appears in eight Claude answers and three Gemini answers, with no OpenAI citations observed. It is ranked by the same counting rules as every other domain. The confirmed Kingy AI YouTube channel appears in one Gemini answer, with none observed from OpenAI or Claude.