How LLMs Choose Sources: What Decides Which Pages an AI Answer Uses
AI assistants don't rank pages the way Google does. Here's the actual pipeline, from query rewriting to retrieval to selection, and which parts of it you can influence.
When ChatGPT, Gemini or Perplexity answers a question with sources, it ran a pipeline to get there. The pipeline has four stages. Each stage filters the set of pages that can end up in the answer, and each one responds to different signals. If you want your pages chosen, you need to know which stage you're losing at.
This post walks the pipeline, then lists what's known from published research about how selection works, then says what you can do about it.
Two ways a model "knows" something
An LLM can draw on two sources of information when it writes.
Parametric knowledge is what the model absorbed during training. Ask ChatGPT who founded a company that's been around for twenty years and it answers from memory. Nothing is retrieved. There's no citation because there's no source, only weights.
Retrieved knowledge is what the model fetches at answer time. Modern assistants run a web search, read the top results and write from them. This is where citations come from and where most of GEO operates.
Most answers about brands, products and anything recent are retrieval-driven. OpenAI's search products use OAI-SearchBot to build an index and ChatGPT-User to fetch pages on demand. Anthropic documents Claude-SearchBot and Claude-User for the same two jobs. Perplexity has PerplexityBot and Perplexity-User. Google's AI features run on Google's index.
The practical consequence: a brand that's absent from training data can still show up in answers by being retrievable. A brand that's well known from training data can still be left out of a specific answer if retrieval returns pages that don't mention it.
The four stages
Stage 1: query interpretation and fan-out
The user's prompt is rarely searched as written. The system rewrites it into one or more search queries. Google describes this as query fan-out: a single question about "a good laptop for video editing under $1,500" becomes several searches covering GPUs, screen quality, specific models and prices.
What this means for you: the engine is searching sub-questions you never see. A page that answers one sub-question well can be retrieved for a prompt that doesn't mention your topic by name.
Stage 2: retrieval
Each rewritten query hits a search index and returns candidates. For Google's AI features, that's the same index and much of the same ranking as classic search. For ChatGPT and Perplexity, it's their own indexes plus third-party search in some cases.
This stage behaves most like SEO. Crawlability, indexation, relevance and site authority decide whether you're a candidate. If your page isn't in the index, nothing downstream can save it.
Stage 3: reading and selection
The system now has somewhere between a handful and a few dozen candidate pages. It reads them (or the parts it fetched) and picks passages that answer the question. This is the stage that behaves least like SEO.
The model is looking for text it can use. Direct statements. Numbers with context. Definitions. Comparisons laid out as comparisons. A page that buries the answer under 600 words of preamble loses to a page that states it in the second sentence, even if the first page ranks higher.
Stage 4: generation and attribution
The model writes the answer and attaches sources. Attribution behavior differs by engine. Perplexity numbers every claim. Google AI Overviews show a set of links beside the text. ChatGPT cites inline for some answers and not others. Some mentions carry no link: the model names a brand because the retrieved pages named it, and moves on.
So a brand can be in the answer without being cited, and a page can be cited without the brand being recommended. The two outcomes are tracked separately in BrandVisibility.ai: mentions in the AI rank tracker and URLs in the citation tracker.
What the research shows about selection
The Princeton team behind the term GEO tested nine content changes against a generative engine and measured which ones increased a source's presence in answers. In GEO: Generative Engine Optimization, the changes that helped most were adding statistics, adding quotations from sources and adding citations. Keyword stuffing didn't help. The best methods raised visibility by up to 40% on their benchmark, with results varying by domain.
Citation-pattern studies add a second lesson: the engines lean on a small number of domains. Profound's analysis of roughly 30 million citations across engines found that Wikipedia accounted for 47.9% of ChatGPT's citations and Reddit for 46.7% of Perplexity's. Semrush then tracked a sharp shift: between August and September 2025, Reddit went from appearing in close to 60% of ChatGPT responses to around 10%. The pattern can change in a month when a provider adjusts its retrieval.
Three conclusions:
- The engines reward pages that look like evidence: numbers, named sources, quotable sentences.
- Community and reference sites get used a lot, so being described accurately on them matters.
- Source mixes are unstable. A tactic that works on one engine this quarter may not next quarter, which is why we measure each engine separately. See tracking AI visibility across providers.
What you can influence at each stage
| Stage | What decides it | What you control |
|---|---|---|
| Interpretation | The engine's rewrite logic | Cover sub-questions explicitly, with their own headings |
| Retrieval | Index membership, relevance, authority | Crawler access, technical SEO, topical depth, links from trusted sites |
| Selection | Extractability, specificity, evidence | Direct answers up top, stats with sources, tables, dated updates |
| Attribution | Engine behavior, source mix | Being the origin of a fact or number; being described on the sites each engine favors |
The biggest lever most teams haven't pulled is at the selection stage. Rewriting the top of every important page so the first 60 words answer the question the heading asks is cheap and measurable. The second biggest is attribution: publishing original numbers (a benchmark, a survey, a price index) gives the engines a reason to cite you rather than paraphrase someone else.
A worked example
Suppose the prompt is "best email warmup tool for agencies."
- Fan-out produces searches like "email warmup tools," "email warmup for agencies," "email warmup pricing," "email warmup reviews."
- Retrieval returns a few vendor pages, two comparison articles, a G2 category page and a Reddit thread.
- Selection pulls the comparison article's table (clean, extractable), two sentences from the Reddit thread (specific, opinionated) and the pricing line from one vendor page (a number).
- Generation names four tools, cites the comparison article and the Reddit thread, and links one vendor.
If you're a fifth tool, you lost at retrieval (not in the comparison article) or at selection (your page had no quotable pricing line). The fix is specific to the stage. That's why sampling the answer and reading the cited sources is the first diagnostic step. The competitor tracking view shows which sources the engines used for the prompts where a competitor appears and you don't.
Frequently asked questions
Do LLMs use backlinks to pick sources?
Indirectly. Backlinks affect whether a page is retrieved, because the underlying indexes use authority signals. Once a page is in the candidate set, selection is about the text, not the links pointing at it.
Why does the same prompt give different sources each time?
Sampling in the model, differences in the rewritten queries and ongoing index updates. Treat one answer as a sample, not a result. Daily runs over a few weeks give you a stable visibility rate. The AI visibility rate and share of voice post explains how to compute it.
Can I get into answers without being in the training data?
Yes. Most brand and product answers are retrieval-driven. If your pages are indexed and extractable, and third parties describe you accurately, you can appear in answers for a brand the model has never "seen" in training.
Does schema markup influence selection?
It helps the engines understand what a page is and what entity it's about, which helps retrieval and entity resolution. It doesn't make weak text more quotable. Treat it as a prerequisite rather than a lever.
Which engine is easiest to influence?
Perplexity responds fastest to fresh, well-structured pages because it re-crawls often and cites heavily. Google AI Overviews track organic rankings most closely. ChatGPT has the most volatile source mix. A per-engine approach is in how to get cited by ChatGPT and how to rank in Perplexity.
