LLM search cites a different set of domains than Google, media type can outweigh content (in the scenario studied), and a correct answer does not guarantee faithful attribution. Three papers examined.
"AI cites us" sounds like success, but the citation can be biased or fail to support the claim. AI search has its own map of domains — carrying SEO assumptions across is risky. Without understanding bias it is easy to build a metric that misleads your own team.
Zhang et al. compare source coverage and citation bias in LLM search against classic systems: some domains appear only in LLM search, and source reliability differs [10]. The conclusion: the AI channel needs its own source map rather than an extrapolation from organic.
Dai et al. show political bias in LLM-generated citations: in the scenario studied, the choice of media source could influence citation behaviour more strongly than the content of the specific text [21]. This is not a general law across all domains but a result within one study of political citation bias. For GEO it is a signal: the authority and type of a domain are a possible distinct risk factor in citation.
Wallat et al. separate correctness from faithfulness: an answer can be right while the link fails to support the specific claim [39]. So the fact of a citation does not by itself mean the source was used honestly in the answer.
Wrong reading: "we are objectively visible". Correct: check absorption and faithfulness [39].
Wrong reading: "the content is better". Correct: possible bias by media type, within the study's scope [21].
Wrong reading: "AI will be the same". Correct: a different map of domains [10].
Coverage and citation bias in LLM search. arxiv.org/abs/2512.09483
Media source outweighs content (political bias). aclanthology.org/2025.emnlp-main.872/
Correctness ≠ faithfulness of attribution. dl.acm.org/doi/10.1145/3731120.3744592
Partly. The citation may come without absorption, or fail to support your claim; a separate faithfulness layer is needed [39].
Because LLM search has different coverage and citation bias — some domains appear only there [10].
A systematic skew in source selection: in the political-bias scenario studied, the type and authority of a medium could influence citation more than content — not to be carried across as a general law for all domains [21].