Before an AI answers a question with sources, it searches on its own. Three terms describe this step: Query Fan-Out, RAG and Grounding. Once you understand them, you know why some pages get cited in AI answers and others don’t.
Here you get the whole process from prompt to source, each term explained on its own, and what it means for your content.
Key takeaways
- Query fan-out: The AI splits a question into several sub-searches and runs them at the same time. Google uses this in AI Mode.
- RAG: Before answering, the language model fetches current content from a search. The term comes from a 2020 study.
- Grounding: The answer is based on these retrieved sources. Only then do you get citations.
- At Google, the sources come from the regular ranking systems. If you are not found in search, you are not cited either.
- For your content, this means: Cover topics completely instead of serving individual keywords.
On this page:
- The process: from prompt to source
- Query fan-out: one question, many searches
- RAG: what the AI retrieves
- Grounding: when an answer gets sources
- What to do with it
- FAQ
- Sources
The process: from prompt to source
When you ask an AI assistant a question with a search component, four steps run in the background. The terms sound technical, the logic behind them is simple: search first, then read, then answer.
PROMPT "Domain ending for my startup?" │ ▼QUERY FAN-OUT ├─ ai domain cost ├─ de or com └─ domain ending startup │ ▼RAG · RETRIEVAL matching pages from search │ ▼GROUNDING answer is based on sources │ ▼ANSWER text with citations
In short: only what was retrieved beforehand makes it into the answer. And what gets retrieved is what the sub-searches find.
Query fan-out: one question, many searches
A question to an AI is often longer and more layered than a classic search query. That is why the AI does not search for the whole sentence. It splits it into sub-questions and searches for each one. Google described this in May 2025 for AI Mode: the question is broken down into subtopics, and many searches run for them at the same time. For its deeper research feature Deep Search, Google speaks of “hundreds of searches”.
75 PROMPTS └▶ 3,180 sub-searches MOST FREQUENT SUB-SEARCH site:placetel.de cloud telefonanlage preisePeec AI, 28 Sep to 27 Oct 2026
Google also explains fan-out with an example in its guide to AI features from 2026: the question of how to fix a lawn full of weeds turns into searches on weed killers and on prevention.
For a founder asking about the right domain ending, it could look like this: the AI searches in parallel for the cost of an .ai domain, for the comparison of .de and .com, and for startups’ experiences. Each of these sub-searches has its own results.
The AI doesn’t search for your question. It searches for its sub-questions.
In short: you are not found for the one question someone asks, but for the sub-searches the AI turns it into. If you only cover half of a topic, you drop out of the remaining sub-searches.
RAG: what the AI retrieves
On its own, a language model only knows its training data, i.e. texts up to a certain cut-off date. For current or very specific questions, that is not enough. Retrieval augmented generation, RAG for short, solves this: the model retrieves matching documents before answering and writes its answer on that basis.
The term was coined in 2020 by a team led by Patrick Lewis in the study Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Today the principle is in almost every AI search, from ChatGPT to Perplexity.
Where the documents come from decides your chances. At Google the answer is clear: according to the guide, AI search draws on the core ranking systems of Google Search. What is not found there is not available to the AI.
GOOGLE AI MODE · AI OVERVIEWS Google index and ranking CHATGPT SEARCH own crawler: OAI-SearchBot CONSEQUENCE not findable, not retrieved
Grounding: when an answer gets sources
Grounding means: the AI bases its answer on the retrieved sources rather than only on what it learned. Only then does it show links to these sources, and only then are there citations.
Not every answer is grounded. For general questions the model can answer from its training, it often doesn’t search at all. For current questions, comparisons, prices or local providers, it almost always does.
FROM TRAINING learned up to the cut-off → mention without a link FROM RETRIEVAL live from search → mention with a citation
Watch out: a mention from training knowledge and a citation from retrieval are two different things. For the first, what has been written about you over the years counts. For the second, what counts is whether your page is found today.
What to do with it
- Topics instead of keywords: Cover the sub-questions of a topic, on one page or in a cluster of pages that link to each other.
- Findable first, then citable: Classic SEO remains the foundation. Pages that are not indexed or rank poorly don’t show up in retrieval.
- Clear sections: Subheadings that answer a sub-question help people and machines. According to Google, you don’t need to chop your texts into small snippets.
- Make sub-searches visible: Some tracking tools show which searches an AI assistant runs for a prompt. That is the best template for missing content.
From my practice
TOPIC ├─ sub-question A → section ├─ sub-question B → section └─ sub-question C → sectionevery sub-search finds an answer
At Placetel I looked at the sub-searches from 30 days in Peec: 75 prompts produced 3,180 different sub-searches. The most frequent was “site:placetel.de cloud telefonanlage preise” (cloud phone system prices). So the AI checks prices directly on the providers’ sites, and just as much at competitors like easybell, NFON and sipgate. Since then I deliberately keep pricing pages crawlable and make sure they show “from” prices as text, even where there used to be only a form or a demo request.
FAQ
What is query fan-out?
Query fan-out means: the AI splits a question into several sub-searches, runs them at the same time and combines the results into one answer. Google uses the method in AI Mode.
What does RAG mean?
RAG stands for retrieval augmented generation. Before answering, the language model retrieves matching content from a search or database and bases its answer on it.
What is the difference between RAG and grounding?
RAG describes retrieving the content. Grounding describes the answer building on this content and referring to it. In practice the two terms are often used interchangeably.
Why am I mentioned but not linked?
Then the mention probably comes from the model’s training knowledge. Only what was retrieved at the moment of the answer gets linked.
Next in the handbook: An AI can only retrieve what its crawlers reach. Which AI bots exist and how to control them is shown in the next chapter: Allow or block AI crawlers.
Sources
- Google: AI Mode in Google Search: updates from Google I/O 2025, May 2025
- Google Search Central: AI features and your website, July 2026
- Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, 2020
