How AI engines choose which sources to cite

How does an AI engine actually build an answer?

In two stages. First retrieval: the engine expands your question into multiple searches, pulls candidate pages from a live index, and scores individual passages for how directly they answer. Then synthesis: the model composes a response from those passages plus its trained knowledge, citing the sources it leaned on. Each stage filters differently, and each is optimizable.

The practitioner vocabulary for the first stage — query fan-out, passage-level retrieval — matters because it kills a common misconception: engines don't select 'good websites', they select useful passages. A mediocre site with one perfectly direct answer paragraph can out-cite an authoritative site whose relevant page never gets to the point. The architecture behind this is RAG; the discipline tying the final answer back to retrieved sources is grounding. The second stage, synthesis, is where recommendations come from — and it draws on a wider pool than the current retrieval, which is why reputation work moves it and page tweaks alone don't.

Watch the two stages in one worked example. Ask an engine for the best accounting firm for e-commerce businesses: fan-out generates searches about e-commerce accounting specialists, marketplace tax handling, and comparison roundups; retrieval pulls directory pages, two 'best firms' articles and a firm's own services page whose opening paragraph directly answers who they serve; synthesis then weighs those against what the model already associates with the category and produces three names with reasons. Every firm named survived both filters. A firm blocked at retrieval — or absent from the roundup layer synthesis leans on — never had a turn.

What does the research say actually moves citations?

The strongest evidence is the Princeton GEO paper (Aggarwal et al., KDD 2024): across roughly 10,000 queries, adding quotations, statistics and source citations to content lifted its visibility in generative answers by 22–41%. Keyword stuffing and fluency polish underperformed. Engines cite what looks like evidence — the single most actionable finding in the field.

Freshness also registers: visible date markers and current-year references correlate with roughly 30% better citation rates in a meta-analysis of citation studies (Medium, 23-study roundup) — we treat that number as directional given its softer sourcing, but the mechanism is intuitive: an engine choosing between two equivalent passages prefers the one that signals currency. Stale dates are cheap to fix and quietly expensive to ignore.

Equally instructive is what the research says doesn't work. The Princeton experiments found keyword stuffing actively underperformed, and fluency-only rewrites — making prose smoother without adding substance — moved little. The pattern across studies is consistent: tactics that made content look optimized lost to tactics that made it more useful as evidence. If your instinct from a decade of SEO is to sprinkle target phrases, the data says to spend those hours adding a sourced statistic instead.

Do all engines pick the same sources?

Emphatically not — and this is the finding that should most change your strategy. Only 11% of domains cited by ChatGPT are also cited by Perplexity (AuthorityTech). Each engine runs its own index, retrieval logic and source preferences. Visibility in one tells you almost nothing about another.

The engines differ in style as well as selection. Perplexity is citation-dense — around 22 citations per response, tying claims to sources in 78% of complex questions versus ChatGPT's 62% (QuickSEO) — which makes it the easiest surface to earn a first citation on and a useful early indicator that on-site fixes are landing. Google AI Overviews build from Googlebot's normal index, so classic crawlability is the lever there. The practical consequence: measure per engine, always. A blended 'AI visibility score' averages away exactly the differences you need to act on.

Which third-party sites do engines lean on?

Analysis of 680M+ citations finds Reddit, Wikipedia-class references, YouTube and LinkedIn dominating most-cited domain lists across engines (Everything-PR), joined by vertical authorities — G2 for software, Avvo for legal, Healthgrades for medical. Engines synthesizing a recommendation poll the places where people compare things in public.

The uncomfortable implication: a meaningful share of your AI visibility is decided on pages you don't own. The constructive version: those pages are enumerable. Run your market's buying prompts, log which sources the engines cite, and you have a finite, prioritized list of where you need earned presence — reviews, community participation, directory depth. Earned is the operative word; seeded mentions get removed by moderators and discounted by engines, and buying your way onto the list reliably backfires.

What does this mean for your site, concretely?

Four moves, in order of speed. Open the gates: confirm AI crawlers can fetch you. Make passages, not pages: every money page opens with a self-contained answer to a real buyer question. Add the evidence layer: statistics, named sources, honest specifics — the Princeton-validated lift. Then build the footprint: consistent entity facts and earned presence on the third-party sources your market's answers already cite. The first three are weeks-scale on-site work; the fourth runs months and compounds — the AEO and GEO split, respectively.