How an AI assistant picks a source, in short
An AI assistant picks a source by searching the web, reading the pages it finds, keeping the passages that answer the question best, and citing the page each sentence of its answer came from. The method is called Retrieval-Augmented Generation (RAG). The model does not cite from memory. It cites what it retrieved a moment ago, so a page it never retrieved cannot be cited, however good it is.
That gives you five checkpoints to pass, in order: the assistant turns the question into searches, retrieval finds candidate pages, a reranker keeps the best passages, trust checks decide what is safe to repeat, and the citation goes to the passage that supplied the claim. Fail an early checkpoint and the later ones never run. This post walks through each one. For the list of factors you can act on page by page, see the AI search ranking factors you control.
Step 1: The assistant turns the question into searches
Your page does not compete for the question the user typed. It competes for the searches the assistant runs on the user's behalf. A question such as "which CRM suits a 10-person agency" is usually rewritten into one or more shorter searches, for example "best CRM for small agencies" and "CRM pricing for small teams". Google describes this as query fan-out in AI Mode, and ChatGPT shows the searches it ran when it browses.
Two things follow. First, the follow-up questions matter as much as the head question, because each one can become its own search. Second, the wording of those searches is closer to classic search phrasing than to the conversational prompt, so the pages that rank for them in a search index are the first candidates.
Step 2: Retrieval finds candidate pages
Retrieval pulls a shortlist of pages and passages whose meaning matches the search, not just its exact words. The search and the stored text are both turned into vectors, numerical representations of meaning, and the system looks for the closest matches. A page about "reducing customer churn" can be retrieved for "lowering subscription cancellations" without containing that phrase.
What this means for your pages:
- You have to be in the index first. Each assistant retrieves from a search index or its own crawl. If the crawler is blocked, or your text only appears after JavaScript runs, there is nothing to match. Our guide to how AI crawlers find and index pages covers this step.
- Cover the topic in the words your readers use. Semantic matching rewards natural language, related concepts, and the follow-up questions around a topic. Forcing one exact phrase does not help.
- Topical focus helps retrieval. A site that covers one subject in depth gives the system many closely related passages to match. A generalist blog that touches the subject once gives it one.
Step 3: A reranker keeps the passages that answer best
The model can only read a limited amount of text per answer, so a second ranking step cuts the shortlist down to the passages that answer most directly. This limit is the context window. Retrieval is fast and broad. The reranker is slower and stricter: it scores each passage against the question and keeps the top few.
Passages that survive this cut tend to state the answer plainly and early. A paragraph that circles the point for several sentences before answering scores lower than a competitor's paragraph that opens with the answer, even when both contain the same fact.
Step 5: The citation goes to the passage that supplied the claim
As the model writes its answer, it attaches each citation to the source passage that most directly supported that sentence. Pages are cited passage by passage, not as a whole. The page with the clearest, most self-contained statement of a fact is the one that gets the link.
Passages that are hard to cite:
- Answers that need several paragraphs of context before the actual response appears
- Key claims buried in the middle of long paragraphs
- Lists with no sentence explaining what the list is
- Tables or charts with no text stating what they show
- Text inside tabs, accordions, or widgets that only load after JavaScript runs
Passages that are easy to cite:
- A direct answer in the first sentence of each section
- Specific facts: numbers, dates, names, and units, each with its source
- Discrete question-and-answer pairs
- Definitions that stand on their own
- All important text present in the HTML the server sends
This is also why an assistant sometimes restates your point inaccurately. It does not copy text. It writes new sentences and attributes them. If your original passage was ambiguous or contradicted itself, the rewrite can drift. Clear, single-claim sentences survive the rewrite best.
How the major assistants differ
The pipeline is the same everywhere, but each assistant retrieves from a different place, so the same page can be cited by one and missing from another. Vendors publish few details, so treat these as working observations rather than rules.
- ChatGPT: Search draws on third-party search indexes, with Bing the most commonly cited partner, plus OpenAI's own crawler. Pages missing from Bing are at a disadvantage.
- Perplexity: Runs its own crawler and index and searches live for most questions. It cites more sources per answer than most assistants and often surfaces recent or specialist pages.
- Gemini and Google AI Overviews: Both draw on Google's index. Google states that a page needs to be indexed and eligible to show a snippet in Search to appear as a supporting link, so classic Google SEO carries over directly.
- Claude: Uses web search when it needs current information. Anthropic publishes little about how results are ranked, so the general principles above are the best guide.
Answers also change from one run to the next, even for the same question on the same day. A single check tells you little. Repeated checks over time tell you whether you are actually being chosen.
What you can change, in order
Work through the checkpoints in the order the assistant runs them: be retrievable, match the searches, answer first, then build trust. Improving trust signals on a page that is never retrieved changes nothing.
- Be retrievable. Allow the AI crawlers you want citing you, get indexed in Bing and Google, and make sure your main text is in the server-rendered HTML.
- Match the searches. Write for the head question and its follow-ups, using your readers' words. Check which searches the assistant actually runs for your topic.
- Answer first. Open every section with the answer, then support it. Keep each section understandable on its own.
- Make every claim attributable. Name the author, link primary sources, add Organization, Person, and Article markup, and keep update dates honest.
- Stay focused. Build depth on the topics you want to be cited for before branching out.
For a full, step-by-step plan that applies these levers to a whole site, read how to get your business into AI search answers.
How to see this pipeline for your own pages
The quickest way to learn why you are or are not cited is to ask the assistants the questions your buyers ask and inspect what they did. Note which searches ran, which pages were read, and which were cited, then repeat the same questions on a schedule, because single answers vary.
Raechal does this for you: ChatGPT is asked live, the way buyers use it, and each answer comes with the searches it ran, the pages it read, and why you were or were not cited. Plans cover ChatGPT, Google AI Overviews, Gemini, Perplexity, and Claude, from one engine on Free to all five on Agency. For the metrics worth tracking and how to read them over time, see how to measure your brand's visibility in AI answers.
Frequently asked questions
Do all five steps matter equally?
No. Being retrieved is a prerequisite: if your page never enters the shortlist, nothing else counts. Once you are retrieved, how directly you answer and how easy you are to attribute decide most of the outcome.
Can a new, low-authority site be cited?
Yes, especially on niche questions where few strong sources exist. A focused, well-structured page that answers a specific question directly can be cited from a young domain if it is the best available passage for that search.
Why does an AI assistant cite a competitor for my topic?
Usually because the competitor's page was retrieved for the searches the assistant ran and stated the answer more directly. Check which searches ran, whether your page ranks for them, and whether your answer sits in the first sentence of the relevant section.
Why did an AI assistant restate my content incorrectly?
Assistants write new sentences rather than copying yours. If the source passage was ambiguous, relied on earlier context, or contradicted itself, the rewrite can drift. Short sections with one clear claim each are restated most accurately.
Does schema markup make an assistant cite me?
Not on its own. Structured data makes your entities, authors, and dates explicit, which helps attribution, but no assistant publishes a fixed boost for it. It supports a page that is already retrievable and clearly written.
