How AI engines choose which sources to cite
Every grounded AI answer is built in two stages: the engine retrieves candidate pages, then the model writes an answer and cites the pages it used. Getting cited means winning both stages.
Stage 1: retrieval — the fan-out decides the candidate pool
When you ask ChatGPT “best CRM for agencies”, it doesn’t answer from memory alone. It fires one or more web searches — often reworded versions of your prompt plus related queries — and reads the results. This is the query fan-out.
The implication: the pages that can be cited are the pages that rank for the fan-out queries, not necessarily for your literal prompt. If the engine searches “agency crm reviews 2026” and your page doesn’t rank there, it was never a candidate.
You can see the fan-out directly: the monitor API returns searchQueries[] alongside the answer. Run your prompt set once and you’ll know exactly which searches feed your category’s answers.
Stage 2: selection — what the model prefers to cite
From the retrieved candidates, the model picks a small set to cite. Across the major engines, the patterns are consistent:
- Extractable statements win. Pages that make a specific, quotable claim — a number, a definition, a named recommendation — are easier to cite than pages that hedge.
- Structure helps. Headings that mirror the question, tables, and lists give the model clean chunks to lift.
- Topical fit beats domain authority. A niche blog post that directly answers the fan-out query often beats a famous domain’s generic page.
- Recency matters for time-sensitive prompts. Engines weight recent pages when the question implies it (“best X 2026”, pricing, news).
What varies by engine
Perplexity cites aggressively and shows its sources prominently. ChatGPT cites fewer pages and favors authoritative-looking explainers. Gemini leans on Google’s index and favors pages already strong in classic SERP features. The only reliable way to know your category’s pattern is to measure it per engine — the same prompt set run through each engine endpoint shows you exactly which sources each one picks.
The practical loop
- Run your prompts; collect
sources[]per engine. - Read the cited pages — note their format, depth, and what claim each makes.
- Publish or update pages that make a better, more specific claim for the same query.
- Re-run on a schedule. Citation share moves over weeks, not days.
Citation selection isn’t random, and it isn’t authority alone — it’s retrieval fit plus quotability. Both are measurable.