How do AI chatbots use sources?
AI chatbots act as gatekeepers in a new attention market and information ecosystem. This explorer audits who they choose to amplify, how they distribute credit, and more.
Open questions about today. Every search-enabled model cites on these.
Citation density
How often do chatbots cite external sources?
Share of answers with at least one link/citation.
How many external sources do chatbots use on average?
Average citations in answers with at least one citation.
How do chatbots’ citation patterns change over time?
Share of answers with a citation, each day.
How do chatbots’ citation patterns differ by topic category?
Share of answers with a citation by question category.
What sources do chatbots cite?
What kinds of sources do chatbots use?
Share of each chatbot’s citations by type of publisher.
Editorial, institutional or user-generated?
Citations grouped into the earned, owned and social split used in search studies.
How credible are the cited news sites?
Average Lin et al. (2023) domain-quality score of cited news-like sites, from 0 (low) to 1 (high). Only rated domains count.
What are the most-cited websites?
The 15 most-cited domains overall, and each chatbot’s share of its own citations that go to them.
| GPT-5 | Claude Sonnet 5 | Gemini 3.5 Flash | Grok 4.5 | Muse Spark 1.2 | |
|---|---|---|---|---|---|
| apnews.com | 42% | – | 1% | 1% | – |
| news.google.com | – | 1% | – | 18% | – |
| wikipedia.org | – | 6% | 6% | 1% | 2% |
| npr.org | – | 10% | – | <1% | – |
| cnn.com | – | 3% | 3% | 4% | – |
| aljazeera.com | 1% | 4% | 3% | 1% | – |
| cnbc.com | – | 6% | – | 2% | – |
| theguardian.com | <1% | – | 4% | 3% | – |
| reuters.com | – | – | – | 6% | 3% |
| bloomberg.com | 1% | 3% | – | 3% | – |
| cbsnews.com | 1% | 2% | 3% | 1% | – |
| youtube.com | – | – | 3% | 2% | 3% |
| sbs.co.kr | – | 5% | – | – | – |
| nbcnews.com | – | 4% | – | 1% | – |
| washingtonpost.com | 3% | <1% | 2% | 1% | – |
Do different chatbots cite the same websites?
Overlap between the sets of domains two chatbots cite (Jaccard: shared domains over all domains either cites).
| GPT-5 | Claude Sonnet 5 | Gemini 3.5 Flash | Grok 4.5 | Muse Spark 1.2 | |
|---|---|---|---|---|---|
| GPT-5 | – | 7% | 6% | 9% | 2% |
| Claude Sonnet 5 | 7% | – | 12% | 15% | 5% |
| Gemini 3.5 Flash | 6% | 12% | – | 13% | 4% |
| Grok 4.5 | 9% | 15% | 13% | – | 5% |
| Muse Spark 1.2 | 2% | 5% | 4% | 5% | – |
How many unique websites do chatbots cite?
Distinct domains cited in the selection.
How do chatbots use what they cite?
What are citations used for?
Each citation coded by the job it does in the answer.
What kind of a page gets cited?
A specific article, or a home page that cannot back a specific claim.
How often do chatbots say that the question is contested?
Share of answers that say experts or groups disagree.
How often do chatbots cite sources that disagree?
Share of answers with two or more citations whose sources give conflicting facts or conclusions.
How do chatbots handle yes-or-no questions?
Average position: +1 is a clear yes, −1 a clear no, 0 both sides. Tap a cell for the answers.