How do AI engines choose which sources to cite?

Before an engine can cite you, it has to find you. Most answer engines run a retrieval step — they issue one or more searches (their own index, a partner search API, or a live web fetch) and pull back a few dozen candidate passages that look relevant to the query. If your page isn't in that candidate set, nothing else matters; you can't be cited from passages the model never saw.

Retrieval rewards the same things classic search does — crawlability, topical relevance, and authority — plus one thing that's specific to passage retrieval: chunk-level relevance. Engines don't retrieve whole pages, they retrieve passages. A page where the answer is buried in paragraph nine, wrapped in qualifiers, is a worse retrieval target than a page that states the answer cleanly near a descriptive heading.

Once the candidate passages are in hand, the model writes an answer and decides which sources to name. It doesn't cite everything it retrieved — it cites the few passages it actually leaned on. The deciding factor is whether your passage is the cleanest, most liftable answer to the question being asked.

Engines can't verify a claim the way a human fact-checker would, so they lean on a proxy: agreement across independent sources. A statistic, definition, or recommendation that shows up consistently across multiple credible pages is 'safe' to repeat. A claim that exists only on your site — with nothing corroborating it — is riskier, so the model is less likely to attribute its answer to you even if your wording is good.

This is why off-page signals still matter for GEO. Mentions, links, and consistent entity data across the web tell the engine that other sources treat you as authoritative. It's also why fabricated statistics backfire: the moment a claim can't be corroborated, it becomes a liability the model routes around.

When several passages are roughly equally relevant and trustworthy, secondary signals decide. Specificity is the biggest one — a page about 'how to contest a parking ticket in California' beats a generic 'parking tickets explained' page for the California query, because it answers the exact intent. Freshness breaks ties on anything time-sensitive (pricing, 'best X in 2026', recent changes). And clear authorship — a named, credentialed author and a real organization behind the page — gives the engine a reason to prefer you in domains where expertise matters.

The practical takeaway: write the answer first, make each key claim self-contained, ground every claim in something verifiable, and earn corroboration off-page. You're not gaming a ranking algorithm — you're making it easy and safe for a model to quote you. Pages built this way tend to win citations across engines at once, because they all reward the same clarity.