• Most AI search uses retrieval-augmented generation: search first, then write.
• A single prompt is expanded into many sub-queries, a process known as query fan-out.
• Content is retrieved as passages, so each section needs to make sense on its own.
• Retrieval uses both keyword matching and semantic similarity.
• You can only be cited if you were retrieved, and you can only be retrieved if you were indexed.

Language models on their own have two weaknesses: their knowledge stops at a training cut-off, and they can state things that are not true. AI search products address both by fetching current information and making the model write from it. That process is called retrieval-augmented generation (RAG), and the act of tying an answer to retrieved sources is called grounding.

The pipeline, step by step

1. Understanding the prompt

The system works out what the user wants, taking into account earlier messages in the conversation, location and sometimes personal context. It also decides whether a web search is needed at all. Simple or timeless questions may be answered from training data alone, with no sources.

2. Query fan-out

Instead of running one search, the system generates several. A prompt such as "best CRM for a ten person estate agency in the UK" might become searches for CRM comparisons, estate agent CRM features, UK pricing, reviews and integrations. Google describes this technique by name for AI Mode and AI Overviews, and other assistants do something similar. These sub-queries are typically short, keyword-like and different from what the user typed.

3. Retrieval

Each sub-query is run against an index. That might be Google's, Bing's or the provider's own. Retrieval commonly combines:

• Lexical search: matching the words in the query.
• Semantic search: comparing embeddings, which are numerical representations of meaning, so that related passages are found even without shared keywords.

Some systems also fetch pages live to read the full text.

4. Chunking and ranking

Pages are split into passages, often called chunks. The system scores them for relevance and quality and keeps a limited number. Only so much text fits into the model's working context, so competition at this stage is fierce.

5. Generation

The model writes an answer using the selected passages, blended with what it already knows. It decides which brands to mention and which claims to include.

6. Citation

Sources are attached to the claims they support. Not every retrieved page gets a visible citation, and a brand can be mentioned in the text without its website being linked.

What this means for your content

• Be indexed and rank for the sub-queries. Because fan-out produces many narrower searches, pages covering specific angles of a topic get more chances than one broad page.
• Write self-contained passages. A paragraph that depends on the three before it loses its meaning when lifted out. Name the subject explicitly instead of relying on "it" and "this".
• Put the answer first. A passage that opens with a clear statement scores better for relevance than one that warms up slowly.
• Use the vocabulary of the topic. Semantic search is forgiving, but precise terminology and related entities still help a passage match.
• Include facts worth quoting. Figures, definitions, steps and comparisons are what models reach for when supporting a claim.
• Keep content fresh. Many systems favour recently updated sources for anything time-sensitive.
• Serve it fast and in HTML. Live fetchers work to short timeouts and do not run JavaScript.

Why answers vary

Models generate text probabilistically, fan-out queries change from run to run, indexes update and personalisation applies. Two people asking the same question can receive different answers with different sources. This is why GEO measurement relies on repeated sampling and share-of-voice style metrics.

Mentions vs citations

• A mention is your brand named in the answer. It may come from training data or from third-party pages that were retrieved.
• A citation is a link to a source. It may or may not be your own site.

A brand can be recommended on the strength of a review site's content, with the review site getting the citation. Both outcomes are valuable and they call for different tactics.

Common misunderstandings

• "The model read my whole site." It read a few passages, selected in milliseconds.
• "Ranking number one guarantees a citation." It improves the odds. It does not guarantee anything, since sub-queries differ from the main query.
• "AI search does not use search engines." Almost all of it does.
• "I can tell the model what to say with hidden instructions." Prompt injection is treated as an attack and filtered.

How to investigate

• Some tools and browser extensions reveal the fan-out queries an assistant ran. Use them to find sub-topics you do not yet cover.
• Check which of your pages are cited for a set of prompts and compare them with your organic rankings for the related searches.
• Look in server logs for user-triggered AI fetches to see which pages are being pulled into live answers.

Add your title here

This is a paragraph. Writing in paragraphs lets visitors find what they are looking for quickly and easily. Make sure the title suits the content of this text.