SEOCompany.com Start your analysis
ServicesServices overviewNational & International SEOAI Search VisibilityEcommerce SEONews SEO for PublishersLocal SEO & Google Business ProfileReputation Management
InsightsInsights overviewAll articlesSEO GlossaryImportant SEO TermsLocal SEO articlesEcommerce articlesSEO tools
CompanyCompany overviewOur HistoryLocationsChoosing an SEO companyEditorial principlesOwnership & fundingContact
Pricing Start your analysis 800-724-0252
An archivist pulls route maps from wooden cubbies as runners pass outside.
  1. Home
  2. Insights
  3. SEO
  4. How AI Answers Pick Their Sources: Retrieval, Fan-Out and What Pages Control
SEO

How AI Answers Pick Their Sources: Retrieval, Fan-Out and What Pages Control

By Ken KnorrPublished Nov 19, 20264 min read
Short answer

Most AI search answers are built by retrieval: the system runs searches, pulls pages from an index, and writes an answer that cites some of them. Google says its AI features rely on its core ranking systems through retrieval-augmented generation and query fan-out. A page gets cited when it is indexable, retrievable for the sub-questions the engine actually searches, and the clearest answer to one of them. Nobody outside the AI labs can train a model on your brand.

Retrieval, not training

There are two separate ways your content can touch an AI model. Training happens months before a model ships, from datasets the lab controls. Retrieval happens at answer time, when the system searches the web or its own index and reads what it finds. SEO influences retrieval. It does not influence training, and anyone promising otherwise is overselling. We go through those claims in AI SEO claims, checked.

Google is explicit about this. Its guide to optimizing for generative AI features says those features are “rooted in our core Search ranking and quality systems” and describes retrieval-augmented generation, or grounding, as relying on core ranking systems to “retrieve relevant, up-to-date web pages from our Search index.” Our glossary entry on RAG goes deeper.

The other assistants keep the same split in their crawlers:

  • OpenAI uses OAI-SearchBot for ChatGPT search and GPTBot for potential training, and says each setting is independent, per its crawler documentation. Sites that block OAI-SearchBot won’t be shown in ChatGPT search answers, though they can still appear as navigational links.
  • Anthropic separates ClaudeBot (training), Claude-SearchBot (search quality) and Claude-User (fetches when a user asks), per its help center.
  • Perplexity documents its own search crawler and user-triggered fetcher on its crawler page.

The practical rule: you can block training crawlers and still allow search crawlers. Block the search crawler and you remove yourself from that assistant’s answers. Our AI crawlers entry lists the user agents.

Volunteers jog out in different directions from a central aid station.

Query fan-out: the engine asks its own questions

An AI answer rarely runs just your query. Google describes query fan-out as “a set of concurrent, related queries generated by the model.” Its example: for “how to fix a lawn that’s full of weeds,” the system might also search “best herbicides for lawns,” “remove weeds without chemicals” and “how to prevent weeds in lawn.” With Gemini 4 Hydrogen, Google says fan-out now runs simultaneous queries across all the relevant subtopics and links to a broader range of sites.

This changes what “ranking” means. You aren’t only competing for the head term. You’re competing to be the best result for the sub-questions the engine generates, and those vary from answer to answer. See query fan-out for more.

A coach hands a runner a folded note at a checkpoint in the rain.

What the newest citation research found

In October 2026, Mahe Chen and Nick Koudas at the University of Toronto published “What Drives AI Citations?”, an analysis of 158,099 answers with 468,037 citations from ChatGPT, Gemini and Perplexity. Three findings matter for site owners:

  1. Retrieval explains most of it. Every page ChatGPT cited was one its own search had retrieved. Which pages that search returned explained far more of the variation than page properties: 32% against 2.8%.
  2. Within a website, one content factor stood out. Pages whose title and paragraphs matched the questions people ask were cited more. FAQ blocks, structured data, summary boxes and page speed appeared to help only when comparing different websites, which suggests they mark stronger sites rather than acting as levers on their own.
  3. Brand memory matters too. ChatGPT wrote brands it already knew into its searches, preferred websites it knew, and sometimes named brands with no supporting page.

The engine’s rewritten subqueries “rarely coincide with Google’s or Bing’s results,” the authors found. So ranking well in Google helps, but it doesn’t guarantee a ChatGPT citation, and the reverse is true too.

Compare that with the original 2023 GEO paper, which tested tactics such as adding citations, quotations and statistics on a lab benchmark. Its ideas still travel widely, but the 2026 data from live engines points to retrieval and relevance over formatting tricks.

What a page can influence

  • Be retrievable. Google says a page must be indexed and eligible to show with a snippet, and the site must be included in Search generative AI features in Search Console. For other assistants, allow their search crawlers.
  • Match real questions in plain language. Write titles and paragraphs that state the question and answer it directly. That is the one within-site factor the Toronto data supports.
  • Cover the sub-questions a real buyer has. Do it on the page where a reader expects it, not in hundreds of thin variants. Google warns that mass-producing pages for fan-out queries violates its scaled content abuse policy.
  • Say something others don’t. Google’s guide favors “non-commodity” content: first-hand experience, original data, a clear point of view.
  • Build the brand off your site. Since models favor brands they know, genuine coverage, reviews and accurate listings matter. Inauthentic “mentions” don’t, per Google.
One water cup is set apart and spotlit on a long aid-station table.

Where answer boxes fit

You’ll notice our articles open with a short answer. We do that because readers want the answer first, not because it’s a proven citation lever. The Toronto data found summary boxes didn’t move citations between pages of the same site. Use them for people. If an engine likes them too, that’s a bonus.

What you can’t control

You don’t choose the model, the subqueries, the user’s location or history, or how many sources an answer shows. Answers also change from run to run, which is why single screenshots prove little. Our guide to measuring AI search visibility covers how to track this honestly.

All of this rewards the same steady work that has always built search visibility: crawlable pages, clear answers, original substance and a real reputation. It’s a marathon, not a sprint. Our AI Search Visibility service is built around exactly that.

The long run starts with one call

SEO is a marathon, not a sprint.

Tell us the searches you need to win. We'll tell you honestly whether they're winnable, what it takes and how long it will take. We don't guarantee rankings or AI citations. Firms that do are worth avoiding.

Start your analysis →800-724-0252Mon–Fri · 9–5 ET