How RAG Systems Select the Sources AI Answers Cite

Retrieval-augmented generation (RAG) systems select sources in stages. A search index first decides which pages are eligible. Vector similarity then scores individual passages against a rewritten query, a reranker trims the shortlist, and the language model cites the passages it used to write the answer. Authority and recency act mostly in the early stages. Passage clarity decides the last one.

That staging is the most useful thing to understand about AI source selection. Each stage filters out candidates before the next one sees them, so a brilliant paragraph on a page that never makes the candidate pool has no chance. The sections below follow the pipeline in order, because that is the order in which you should diagnose a missing citation.

What happens between a question and a citation?

The assistant turns the user's question into searches, pulls back candidate pages, breaks them into passages, scores those passages, and hands the best few to the model as context. The model writes from that context and attaches citations to the sources it drew on.

The exact implementation differs by product and is not fully published, but the general shape is consistent across ChatGPT, Perplexity, Gemini, Claude, Microsoft Copilot, and Google AI Overviews:

Query rewriting: The assistant rewrites the prompt into one or more search queries. A question like "is Notion good for a small law firm" might become separate searches about Notion security, client confidentiality, and legal templates. Google calls this query fan-out in its description of AI Mode.
First-stage retrieval: Those queries go to a search index, which returns a limited set of candidate URLs for each one.
Chunking: The retrieved pages are split into passages, often a few paragraphs each, sometimes guided by headings.
Scoring: Each passage is compared with the query, usually through a mix of vector similarity and keyword matching.
Reranking: A second model rereads the top passages alongside the query and reorders them by how well they answer it.
Generation and attribution: The language model writes the answer from the surviving passages and links the ones it relied on.

The practical consequence is a diagnostic order. If your page never appears, check eligibility first (stages one and two), then passage quality (stages three through five), then how quotable your content is (stage six). Answer engine optimization (AEO) work that starts by rewriting copy often fixes the wrong stage.

How does a RAG system decide which pages are eligible?

Mostly by borrowing a search engine's judgment. The candidate pool comes from a search index, so a page has to be crawlable, indexed, and ranking reasonably well for the rewritten queries before any AI-specific scoring happens.

Which index depends on the product. Google AI Overviews and AI Mode build answers from Google's index, and Gemini grounds its answers in Google Search. Microsoft Copilot draws on Bing. ChatGPT, Perplexity, and Claude search the live web through a mix of third-party search providers and their own crawlers, and none of them fully discloses that mix.

Classic search engine optimization (SEO) fundamentals therefore carry straight into AEO. Because systems usually pull only the top handful of results per query, a page sitting deep in the results for "payroll software for restaurants" rarely enters the pool, however well its passages are written. Fan-out helps a little here. A page that ranks poorly for the user's exact phrasing can still be retrieved if it ranks well for one of the narrower subqueries the assistant generates.

Crawler access is the other gate. OpenAI documents separate crawlers for different jobs: OAI-SearchBot fetches pages for ChatGPT search results, while GPTBot collects content that may be used for training. Perplexity uses PerplexityBot. If your robots.txt or your content delivery network blocks the search crawler for a given assistant, your pages drop out of that assistant's live answers entirely.

How does vector search score a passage?

Vector search converts both the query and each passage into embeddings, which are long lists of numbers that represent meaning, then measures how close they sit to each other. Passages whose meaning is close to the query score high even when they share few exact words.

That is the strength of vector search. A query about "keeping client files private in Notion" can match a passage about "workspace permissions and data encryption" because the concepts overlap. It is also the weakness. Embeddings blur exact terms like product names, model numbers, and integration names, which is why most production systems run a hybrid that combines vector similarity with keyword ranking such as BM25. A passage that states the concept and also names the specific product tends to do well on both.

Scoring happens at the passage level. A long guide is effectively many separate candidates, and each chunk competes on its own. Take a question like "does Asana integrate with Salesforce." A help article whose opening paragraph says plainly that Asana connects to Salesforce, and how, gives the retriever a tight match. An integrations page that shows a grid of partner logos gives it almost nothing, because logos in images carry no text to embed.

Chunking creates a second trap. When a page is split, a paragraph that begins "It also supports this for enterprise plans" loses its subject. The embedding of that chunk no longer knows what "it" or "this" refers to, so it matches nothing well. Passages that name their subject in the first sentence survive the split intact.

Where do content authority signals come in?

Authority enters at three points: the search ranking that shapes the candidate pool, quality filters or rerankers that some systems apply to sources, and the model's own familiarity with a brand from training. None of these weights are published, so treat any precise claim about them with suspicion.

The first point is the strongest and the most familiar. Links, topical depth, and site reputation all influence how a search engine ranks a page, and that ranking decides whether the page is retrieved at all. Content authority in this sense is inherited from SEO.

Rerankers and source filters are less visible. Some systems appear to favor sources that read as independent for evaluative questions. For a prompt like "best CRM for a nonprofit," the cited pages are often review roundups, comparison articles, and forum threads rather than the vendors' own sites. The practical lesson is that your own page may lose the citation to a third-party page that mentions you, and that is still a win if the mention is accurate and favorable.

The third point is training, which belongs to generative engine optimization (GEO) rather than retrieval. A model that learned during training that HubSpot and Salesforce are CRM vendors will recognize those names in retrieved passages and may weight them more readily when composing an answer. A brand the model has never encountered has to be fully explained by the retrieved text itself. This is one place where GEO and AEO meet: consistent mentions across Wikipedia, Reddit, and industry publications make your retrieved pages easier for the model to use.

How much does recency affect source selection?

Recency matters a lot for time-sensitive questions and very little for stable ones. Questions about pricing, current product lineups, regulations, or "best" lists tend to pull fresher sources, while questions about definitions and established processes are answered from whatever explains the concept best, old or new.

Freshness can enter in several ways. Query rewriting may add terms like "latest" or the current year to time-sensitive questions. The underlying search index already applies its own freshness signals. Some systems also read publication or update dates from the page and prefer newer candidates when several pass the relevance bar.

The common mistake is treating recency as a date stamp. Updating the "last modified" line on an unchanged comparison page gives the model nothing new to cite, and the stale passage inside still competes on its old wording. If a page covers something that changes, such as feature sets or plan structures, the content itself needs to change, and the date shown on the page should reflect that real edit.

Which content quality indicators help a passage get cited?

The passages that get cited are the ones a model can lift with confidence: specific, self-contained, and stated plainly. At the final stage, the model attributes claims to passages that clearly support them, and vague marketing copy supports nothing in particular.

Signal	Why it helps at retrieval	What to check on your page
Direct answer up front	The first chunk under a heading matches the query closely	Does the first sentence under each heading answer that heading?
Self-contained passages	Chunks keep their meaning after the page is split	Do paragraphs name their subject instead of relying on "it" or "this"?
Named entities	Hybrid scoring rewards exact product and company names	Are products, integrations, and categories named in text?
Text in the HTML	Crawlers that skip scripts or images still see the content	Does key content appear when you view the page source?
Specific, checkable claims	The model can attribute a precise statement to a source	Could someone quote this sentence as a fact?
Consistent facts across pages	Conflicting passages give the model reason to pick another source	Do your pricing, features, and specs match everywhere?

Structured data deserves a careful note. Schema markup helps search engines understand and index a page, and that indexing is what AI search features retrieve from. It does not make a model cite a page, and nobody can promise that it will. Treat it as part of being eligible, alongside crawlability and clean HTML.

Mistakes that keep good pages out of AI answers

Most missing citations trace back to a handful of avoidable problems, and each one fails at a specific stage:

Blocking search crawlers by accident: Bot protection settings at providers like Cloudflare can challenge or block AI crawlers. The cost is total invisibility in that assistant's live answers, regardless of content quality.
Confusing training crawlers with search crawlers: Blocking GPTBot to keep content out of OpenAI's training is a separate decision from blocking OAI-SearchBot, which removes you from ChatGPT search results. Teams sometimes block both when they meant to block one.
Rendering key content only with JavaScript: Some AI crawlers fetch raw HTML and may not run scripts. Product details loaded client-side can be missing from what gets indexed and chunked.
Burying the answer: A pricing explanation that starts after several paragraphs of brand story lands in a chunk the retriever scores poorly for the pricing question.
Expecting schema to force citations: Markup supports indexing. Teams that add it and stop there are usually left wondering why nothing changed.
Judging from a single prompt: AI answers vary with phrasing, context, model version, and chance. One run that cites you, or one that does not, tells you very little.

How do you see which sources get picked for your topics?

Sample the answers repeatedly and record what gets cited. Because responses are non-deterministic, visibility is a frequency across many runs, so a single screenshot is never evidence of anything stable.

Start with the questions your buyers ask, phrased the way they would type them. Run each one several times in ChatGPT, Perplexity, Gemini, and Google AI Mode, and note every cited URL. Then open the cited pages and find the specific passage the answer draws on. That passage is your real competition, and comparing it with the matching section of your own page usually shows which stage you are losing at.

Two other data sources fill in the picture. Your server logs show whether crawlers like OAI-SearchBot and PerplexityBot are fetching your pages at all. Your analytics show referral visits from domains such as chatgpt.com and perplexity.ai, which come from people who clicked a cited link. You cannot tag content so that AI answers report back to you, so those referrals and your own sampling are the evidence available. If you want this sampling run continuously across assistants, nonBot AI tracks where a brand appears in AI answers and which pages those answers cite.

Where to start this week

Pick five questions a buyer asks before choosing a product in your category. Run each one at least three times in ChatGPT, Perplexity, and Google AI Mode, and list every cited URL in a spreadsheet. Check whether your own pages appear, and if they do not, confirm in your robots.txt and server logs that the search crawlers for those assistants can reach you.

Then take the single most frequently cited competitor page, find the passage the answers draw from, and rewrite the first two sentences under the matching heading on your own page so they answer the question as directly as that passage does.

FAQ

Does ranking well on Google guarantee a citation in AI Overviews?

No. Ranking makes a page eligible because AI Overviews build answers from Google's index, but the system still scores passages and chooses which ones to use. A high-ranking page with its answer buried deep in the text can lose the citation to a lower-ranking page that states it plainly.

Can I block AI training and still appear in AI search answers?

For some vendors, yes. OpenAI separates GPTBot, which gathers training content, from OAI-SearchBot, which fetches pages for ChatGPT search, so you can block one and allow the other. Other vendors handle this differently, so read each one's crawler documentation before editing robots.txt.

Do longer articles get cited more often?

Length does not help directly, because retrieval scores passages rather than whole pages. A longer page can match more of the subqueries generated by fan-out, but only if each section opens with a clear, self-contained answer. Padding adds chunks that compete with nothing.

Why does the same prompt cite different sources each time?

Several parts of the pipeline vary between runs. The rewritten queries can differ, search results shift as indexes update, and the model's generation involves sampling. That is why citation frequency across many runs is a more honest measure than any single answer.