By Mehdi HassanHow do AI search engines choose which sources to cite?
AI search engines choose sources by retrieving pages that answer the parts of a question, then writing an answer grounded in the passages they trust most and citing some of them. The process differs by engine, and none publish full details, but Google, OpenAI and Perplexity describe the same broad pattern: search first, then summarise from what was found.
This guide breaks that pattern into four stages and maps each to something your marketing site can do. It sits within our answer engine optimization guide, which covers the full strategy around it.
Stage | What the engine does | What your site controls |
|---|---|---|
1. Access | Crawls and indexes pages it's allowed to fetch | Robots rules, CDN bot settings, server-rendered HTML |
2. Fan-out | Splits the question into sub-queries | Pages that answer specific sub-questions |
3. Retrieval | Finds the passages that best match each sub-query | Answer-first sections, clear headings, tables |
4. Grounding and citation | Writes the answer from trusted passages and cites some | Accuracy, authorship, freshness, brand reputation |
Stage 1: Can the engine access your page?
An engine can only cite pages it can fetch and read. Google's AI features draw on Google's own index, crawled by Googlebot. ChatGPT search uses OpenAI's OAI-SearchBot along with search partners, and Perplexity uses PerplexityBot plus its own retrieval. If any of these is blocked, you're absent from that engine's answers no matter how good the page is.
- Check robots.txt for rules that block the search crawlers you want. Blocking a training crawler is a separate choice from blocking a search crawler.
- Check your CDN or firewall. Bot protection can quietly block AI crawlers by default.
- Server-render the content. Many AI crawlers don't run JavaScript, so text that appears only after scripts run may not exist to them.
- Keep URLs stable. If you migrate, redirect old URLs, because engines may still hold them; see migrating without losing AI citations.
Stage 2: What is query fan-out?
Query fan-out is when an engine turns one question into several narrower searches before it answers. Google has described AI Mode working this way, and other engines behave similarly. A question like "best CRM for a 20-person agency" might become searches for CRM pricing for small teams, agency-specific features, integrations and reviews.
This is why AI citations don't follow rankings neatly. Conductor reports 57% of AI Overview citations come from outside the top 10, and BrightEdge puts the overlap with the top 10 near 17%. A page that ranks poorly for the broad query can still be the best answer to one sub-query, and get cited for it.
Hypothetical example: one question, five sub-queries
For example, suppose a buyer asks an AI engine: "Should a seed-stage SaaS startup build its marketing site in Webflow or Next.js?"
The engine might search for Webflow pricing for startups, Next.js marketing site cost, Webflow CMS limits, which is easier for non-developers to edit, and migrating from Webflow to Next.js later. Five different pages could each be the best answer to one of those, and the final response might cite three of them. A cost page with a clear table could be cited even though it never mentions the original question. This is illustrative; engines don't publish their sub-queries.
Stage 3: Which passages get retrieved?
Engines retrieve passages that answer a sub-query directly and make sense on their own. A passage that names its subject, gives a specific answer in its first sentence and sits under a heading that matches the question is easy to match. One that starts with background and refers to "the above" is hard to use.
Position matters too. Reported analyses of ChatGPT citations put about 44% of them in the first 30% of a page. Put the main answer near the top, then open each section with its own answer. Use real HTML tables for comparisons and lists for steps. The how-to version of this is in how to structure content for AI search.
Engines also differ in how many passages they use. Semrush's AI Visibility Index, from 126 million prompts, found ChatGPT cites about 15 sources per answer against about 3 for Gemini. More citations per answer means more room for smaller sites, but also that each one gets less attention. See ChatGPT search vs Google for more on how they differ.
Stage 4: How do engines decide which sources to trust?
When several passages say similar things, engines lean on trust: is the source accurate, current and recognised? The signals are much like Google's long-standing quality signals, plus a heavier weight on what the rest of the web says about you.
- Reputation: in Ahrefs' study of 75,000 brands, web mentions correlated 0.664 with AI Overview visibility, against 0.218 for backlinks.
- Freshness: AirOps reports over 70% of cited pages were updated within 12 months.
- Authorship and evidence: named authors, sources and specific numbers make claims safer to repeat.
- Consistency: if your site, profiles and third-party mentions describe you the same way, there's less for the engine to doubt.
Structured data is often sold as a trust signal for AI. Ahrefs' controlled 2026 test found no citation lift from schema, so treat it as hygiene; our article on schema markup and AI citations has the details.
What can a marketing site actually control?
You control access, structure, accuracy and the facts you publish about yourself. You partly control reputation, through what you give others to mention. You don't control the model, the sub-queries it generates or which of several good sources it picks on a given day.
Open the doors
Allow the search crawlers you want, check CDN settings and make sure key pages are server-rendered.
Answer the sub-questions
Build pages for the specific questions buyers ask on the way to a decision: pricing, comparisons, integrations, security, use cases.
Write passages that stand alone
Answer first, name the subject, keep sections focused and use tables and steps where they fit.
Give engines reasons to trust you
Named authors, real sources, visible dates and consistent company facts across the web.
Measure what happens
Track AI referrals and run a monthly prompt audit; our guide to measuring AI search visibility shows how.
The complete, prioritised list is the AEO checklist. And because AI answers often send no click, attribution needs its own approach, covered in B2B attribution when buyers research in ChatGPT.
