A query hits ChatGPT. A retrieval system pulls a handful of documents from millions, ranks them by vector similarity, and feeds them to the model. The answer that reaches the user cites maybe three or four sources. What determined which ones made the cut?
This is the question that matters for anyone building content strategies around AI visibility. And the answer, based on converging academic research and industry analysis, looks nothing like traditional SEO ranking factors.
RAG Decides Before the Model Sees Anything
Before an LLM generates a single word of a cited response, a Retrieval-Augmented Generation system has already made the most consequential decision: which documents enter the context window at all.
RAG works by converting both the user's query and candidate documents into vector representations, then ranking sources by mathematical similarity. This retrieval stage is the first and most brutal filter. If your content doesn't clear it, nothing else matters. The model literally never sees you.
What clears this gate? Semantic relevance, primarily. AWS documentation on RAG systems confirms that semantic relevance ranking determines which sources enter the payload that gets passed to the language model. But relevance alone doesn't explain the patterns researchers are finding in actual citation data.
Position Bias and the Matthew Effect
Once documents enter the context window, two well-documented biases shape which ones get cited.
The first is position bias. A systematic investigation across roughly 80,000 evaluation instances found that LLMs favor sources based on where they appear in the retrieval order. Positional consistency ranged from 0.565 to 0.828 depending on the model. What matters more: when quality differences between sources are small, position bias increases dramatically. The research describes a parabolic relationship — the closer two sources are in quality, the more arbitrary positioning determines the winner.
The second is the Matthew effect: accumulated advantage compounding over time. An analysis of over 274,951 citations in LLM outputs found a systematic preference for already highly-cited sources, with a median citation count gap exceeding 1,300 citations between sources LLMs chose and sources they didn't. Sources that are already authoritative get cited more, which makes them more authoritative, which gets them cited more.
If that sounds like the rich-get-richer dynamic from traditional search, it is — except the mechanism is different. This isn't PageRank. It's pattern-matching on what the model learned during training about which sources tend to be reliable.
Freshness and Structure
Two structural factors from the Digital Bloom 2025 AI Visibility Report matter more than most teams realize.
Recency matters enormously. 65% of AI bot traffic targets content published within the past year, and 79% targets content from the past two years. Citation turnover runs at 40–60% monthly, far exceeding anything we see in traditional organic rankings. If you're not updating content regularly, you're falling out of the citation pool faster than you'd fall out of Google's index.
Content architecture also affects extractability in measurable ways. Self-contained paragraphs of 40–60 words, question-based headings, and answer-ready formatting all appear disproportionately in cited sources. One critical technical detail: AI crawlers don't execute JavaScript. Server-side rendering isn't optional — it's a prerequisite for being crawled at all.
These aren't surprising findings on their own. Fresh, well-structured content has always performed better. What's different is the velocity. Monthly turnover rates of 40–60% mean the citation landscape reshapes itself roughly every two months. That's not an update cycle most content teams are built for.
Brand Authority Beats Link Authority
The same report found something that overturns conventional SEO wisdom: brand search volume shows a 0.334 correlation with citation likelihood, making it the single strongest predictor they measured. Backlinks? Weak or negative correlation. The signal that defined SEO authority for two decades is, at best, irrelevant to LLM citation and may actually correlate negatively with it.
This is entity recognition, not link graphs. The model knows your brand or it doesn't. Cross-platform presence matters too: brands appearing on four or more platforms are 2.8x more likely to earn ChatGPT citations. The implication is that brand-building work — the kind of marketing that SEO teams historically dismissed as "awareness" — may be the most important citation optimization lever available.
What Actually Drives Citations Now
What we're seeing is a genuine rebalancing of which signals matter for visibility. The filtering cascade works like this: RAG narrows millions of documents to dozens through vector similarity. Position bias influences ranking within that set. Brand authority (not link authority) separates the final candidates. And structural formatting determines whether the model can actually extract a clean citation from your content.
Only 11% of domains appear in both ChatGPT and Perplexity citation results, according to the Digital Bloom analysis.
That stat alone should kill any "optimize once for AI" strategy. These are not interchangeable platforms with shared ranking logic.
ChatGPT correlates 87% with Bing's results and leans toward Wikipedia and Reddit. Perplexity cites Reddit 46.7% of the time. Google's AI Overviews maintain a 93.67% correlation with traditional organic rankings. Three platforms, three different citation philosophies.
The practical upshot: you need a brand that LLMs recognize, content that's fresh and structurally extractable, and a platform-specific understanding of where your audience is actually getting AI-generated answers. Traditional backlink acquisition isn't just less valuable here — the data suggests it might be orthogonal to what actually drives citations.
That's a hard message for teams that have spent years building link equity. But the research is consistent across multiple independent studies, and the signal is clear enough to act on.