Where do AI answers get their facts?

The three views

• Dandelion: sources only. The most central source is the hub, the 40 most-cited others are sub-hubs, and every remaining source hangs from the sub-hub it is most often cited alongside.
• Answers: queries (small squares) linked to every source their answers cite. Nothing is pinned; a source sits in the middle only if many queries pull it there.
• Orbit: the engine at the center and every source at a distance set by its rank. A summary, not evidence: the layout is imposed.

The numbers

• An answer is one engine response to one query in one run. Each query was asked three times per wave.
• Share of answers counts answers citing a source at least once.
• Lead over next ten divides the leader's answer count by the mean of ranks 2–11. The interval is a 90% bootstrap over queries.
• Shared center: the overall hub is a top-3 source in at least half of the 12 topics.
• Node view: a node is either a whole site (youtube.com) or one community inside it (a YouTube channel, a subreddit). An engine can lean on a platform without leaning on any one channel. Links whose URL doesn't name the account, such as most Instagram posts and Pinterest pins, are grouped as “instagram.com/(unknown account)” or “youtube.com/(unknown channel)” rather than split.

Source and caveats

Model: Adami, Ebadi and Nattagh-Najafi, “A dandelion structure of eigenvector preferential attachment networks”, Scientific Reports, 2024. Queries come from ORCAS (Craswell et al., 2020), used for non-commercial research.