Do citations from AI assistants depend more on authority or relevance?

The short answer is: Relevance determines which specific source the AI cites, but Authority determines if the source is available to be cited in the first place.

If forced to pick which one matters more to the AI model itself, the answer is Relevance. However, because AI search engines operate in a two-stage pipeline, Authority acts as a strict gatekeeper.

Here is a breakdown of how authority and relevance interact when an AI assistant (like Perplexity, ChatGPT with Search, or Gemini) selects a citation.

1. Why Relevance Dominates the AI's Selection

AI models do not search the web or process text the way humans do; they use vector embeddings and semantic search. They evaluate content based on how closely the text’s meaning matches the user’s prompt.

• Contextual Precision: An AI’s primary job is to directly answer the user’s specific query. A highly authoritative website (like Harvard.edu) that only broadly touches on a topic will be ignored in favor of a lesser-known specialized blog that answers the user's exact, niche question.
• Semantic Similarity: LLMs rank snippets of text based on "cosine similarity"—a mathematical measure of how closely related two pieces of text are. The model selects citations that best fill the gap in its generated sentence, making relevance the ultimate deciding factor during synthesis.
• Information Extraction (RAG): Modern AI assistants use Retrieval-Augmented Generation (RAG). The LLM reads the top 10–20 retrieved search snippets and picks the ones that contain the exact facts needed to construct its answer.

2. Why Authority Acts as the "Gatekeeper"

While the AI model prioritizes relevance, it can only cite sources that make it into its context window. How do sources get there? Through traditional search engines (like Bing or Google).

• Search Engine Gatekeeping: Most AI assistants rely on search APIs (primarily Bing) to gather real-time data. Search engines heavily weigh Domain Authority, backlinks, page speed, and brand trust. If a relevant page sits on Page 4 of Bing because it lacks authority, the AI will never see it, and therefore will never cite it.
• RLHF and Safety Alignment: AI models are fine-tuned using Reinforcement Learning from Human Feedback (RLHF) to avoid spreading misinformation or "hallucinations." Developers program the systems to favor high-trust domains (e.g., Wikipedia, government databases, major news outlets, peer-reviewed journals) when resolving factual queries.
• Consensus Bias: AI assistants look for "consensus facts." If a high-authority site states a fact and ten low-authority sites contradict it, the AI is trained to trust the high-authority consensus.

The Two-Stage Citation Pipeline

To understand which matters more, it helps to look at the process as a two-stage funnel:

[ User Query ]
      │
      ▼
┌────────────────────────────────────────────────────────┐
│ STAGE 1: Retrieval (Search Engine API)                  │
│ Filter: Heavy bias toward AUTHORITY & Traditional SEO   │
│ Result: Top 10–20 web results                          │
└────────────────────────────────────────────────────────┘
      │
      ▼
┌────────────────────────────────────────────────────────┐
│ STAGE 2: Synthesis & Citation (LLM / RAG)              │
│ Filter: Heavy bias toward RELEVANCE & Direct Answers   │
│ Result: 3–5 cited sources in the final AI response     │
└────────────────────────────────────────────────────────┘

1. Stage 1 (Authority-Driven): The search engine grabs the top results based on traditional ranking algorithms (where Authority and overall SEO dominate).
2. Stage 2 (Relevance-Driven): The AI reads those top results and selects the sources that best answer the prompt (where Relevance dominates).

Summary: The Verdict

• Authority gets you into the room; Relevance gets you cited.
• If your content is highly authoritative but irrelevant to the specific prompt, the AI will discard it.
• If your content is hyper-relevant but lacks baseline authority/SEO, the search engine feeding the AI will never retrieve it.

For broad or simple queries ("What is photosynthesis?"), Authority wins because thousands of sites have relevant answers, so the AI defaults to major educational or scientific sites.

For complex, long-tail, or niche queries ("How to fix error code 0x80070005 in Windows 11 after the May update"), Relevance wins, and the AI will happily cite an obscure forum or tech blog if it provides the exact, working solution.