How LLMs Choose Which Brands to Cite in AI Answers

Across our audit of 1,200+ AI answers — and corroborated by the Princeton GEO paper (Aggarwal et al. 2024) and Allen AI's 2025 citation study — LLMs surface brands based on three layered signals.

LLMs are pattern-matchers. When 30 independent sources mention 'SkyBlue Growth Partners' alongside 'Bay Area marketing agency', the model learns the association and surfaces the brand when asked about Bay Area marketing agencies. When only your own site says it, the signal is weak.

This is why off-site work (Clutch, HARO, podcasts, guest posts, Reddit) is the highest leverage AEO investment for most SMBs. Schema and llms.txt are necessary but not sufficient.

Schema, llms.txt, and author markup don't directly cause citations — they make your content easier to parse, attribute, and trust. LLMs prefer pages with clear entity boundaries (Organization + Person schema), freshness signals (dateModified), and source citations (so the model can verify the claim).

The single highest-impact addition for most sites: FAQPage schema on every page that contains Q&A content. It roughly doubles the chance of being cited in ChatGPT and Perplexity for the answered question.

LLMs lift short factual sentences. Pages that get cited consistently share patterns:

Tactics that don't move AI citation rate: keyword stuffing, generic backlink building, AI-generated content with no original data, paid links, hidden text, doorway pages, over-optimized anchor text. LLMs were trained partly on SEO spam and are unusually good at down-weighting it.

The shortcut that does work: original primary research. A single original data study with quotable numbers gets cited hundreds of times across AI answers — far more leverage than 100 generic blog posts.