How LLMs Decide What To Cite: The Attribution Pipeline That Powers AI Search
Every time you ask ChatGPT, Perplexity, or Gemini a question and get an answer with inline citations, a complex pipeline of retrieval, scoring, and attribution decisions has already fired behind the scenes. The citations that appear in AI-generated answers are not random, and they are not generated from scratch. They are the output of a structured, multi-stage decision process — one that is simultaneously powerful and deeply flawed.
This article unpacks exactly how large language models (LLMs) decide what to cite, why citation accuracy remains a persistent problem, and what it means for brands and content creators who want to appear in AI-generated answers.
The Two Knowledge Pathways
Before any citation can appear, an LLM must first decide where its knowledge comes from. There are two fundamentally different pathways:
1. Parametric Knowledge (Training Data) When a model generates an answer without web search enabled, it relies entirely on what it "learned" during pre-training. This knowledge is static — fixed at the model's training cut-off date. Entities mentioned frequently across authoritative training sources develop stronger neural representations and are more likely to be recalled. According to the 2025 AI Visibility Report, approximately 60% of ChatGPT queries are answered from parametric knowledge alone. However, any URLs or citations produced from this pathway are generated from memory, not dynamically retrieved — they may be entirely fabricated.
2. Retrieved Knowledge (RAG Systems) When web browsing or search is enabled, the model uses Retrieval-Augmented Generation (RAG) . This is the only pathway that produces genuine, verifiable citations. The model queries external search indexes or document stores at runtime, retrieves candidate sources, evaluates them, and then generates an answer grounded in those sources.
As Flavio Longato, LLM Optimization Strategist at Adobe, puts it: "If there is no retrieval layer, there is usually no citation. The model can generate fluent text from its training data, but it cannot point to a current, verifiable source."
The RAG Pipeline: How Sources Are Retrieved
When a user submits a query to a RAG-enabled LLM, the system executes a four-step pipeline:
Step 1: Query Encoding & Fan-Out
The user's natural language query is converted into a dense vector embedding. But many modern systems go further — they expand a single query into multiple sub-queries through a process called prompt fan-out. For example, a query like "What is the best PDF editing software?" might be expanded into:
- Best PDF editing software
- Convert PDF to Word comparison
- PDF editor features review
- PDF tools pricing 2026
This ensures the system captures subtopics, entities, and variations in user intent.
Step 2: Hybrid Retrieval
Retrieval systems use a hybrid approach combining two methods:
- Semantic Search (Dense Vectors): Matches the semantic meaning of the query to indexed content
- Keyword Matching (BM25): Matches exact terms for precision
Research shows hybrid retrieval delivers a 48% improvement over using either method alone. Results from both approaches are combined using Reciprocal Rank Fusion (RRF) to produce a unified ranking.
Step 3: Reranking
The initial set of candidates is reranked using a more sophisticated relevance model. This step improves NDCG@10 by up to 28% and ensures the most relevant sources float to the top. The reranker evaluates both semantic relevance and surface-level keyword alignment.
Step 4: Source Evaluation & Selection
Once results are surfaced, the system evaluates each candidate across multiple dimensions:
This is the critical insight: Citations are rarely chosen from a single top ranking position. They are selected based on aggregate visibility across multiple fan-out queries. A site ranking ninth but covering every sub-intent thoroughly may outperform a site ranking first for one narrow query.
G-Cite vs. P-Cite: Two Citation Paradigms
Recent academic research has formalized two paradigms for how citations are actually generated:
Generation-Time Citation (G-Cite)
The model produces answer text and citation markers together in a single left-to-right pass. It decides locally, token-by-token, which sources to cite based on what has been written so far and any retrieved evidence. G-Cite tends to produce higher precision but lower coverage of relevant citations.
Post-hoc Citation (P-Cite)
The model first drafts a complete answer, then runs a second pass to add or verify citations against available evidence. P-Cite achieves higher coverage with competitive correctness, but adds latency. A comprehensive evaluation published at ACL 2025 found that retrieval augmentation is the fundamental driver of quality in both paradigms — the largest gains come from investing in the retrieval infrastructure, not the citation strategy itself.
Recommendation for high-stakes applications: Use a P-Cite-first approach, reserving G-Cite for precision-critical settings like strict claim verification.
The Citation Hallucination Problem
Understanding how LLMs should cite is only half the picture. The uncomfortable reality is that LLMs frequently cite sources that don't exist or misattribute claims to real sources that don't support them.
How Bad Is It?
- GPT-3.5 fabricated 55% of its citations in a study of 84 generated literature reviews. GPT-4 improved this to 18% fabrication — a major step, but still unacceptable for professional use.
- Across 10 commercially deployed LLMs tested across four academic domains, citation hallucination rates spanned a fivefold range from 11.4% to 56.8%.
- Deep research agents hallucinate URLs at higher rates than search-augmented LLMs. Per-query, Gemini 2.5 Pro Deep Research generates 113 URLs per query, with 3–13% hallucinated.
- A 2026 study found that citation hallucination is prompt-induced, not intrinsic — no model spontaneously generates citations when unprompted. Every fabricated citation is a product of being asked to cite.
Geographic and Domain Bias
Citation hallucination is not evenly distributed. A cross-country analysis found that DOI hallucination rates exceeded 80% in lower-income countries and increased sharply for publications from the 2020s. Models also show systematic bias: GPT-4 exhibits a strong preference for highly cited papers, amplifying the Matthew effect in citations. The median citation count of LLM-generated references was 1,326 higher than ground truth references.
Why Internal Mechanisms Matter
Research on model internals (FACTUM, 2025) reveals that correct citations are characterized by higher contextual attention alignment, stronger use of retrieval heads, and higher parametric force — all working in concert. When these internal mechanisms decouple, attributional drift occurs: the longer the model writes, the more likely it is to forget which statement came from which source.
Platform-by-Platform: Citation Sources Diverge
Different AI platforms draw from fundamentally different source ecosystems:
Cross-platform overlap is shockingly low. Only ~11% of domains are cited by both ChatGPT and Perplexity. If your brand is visible on one platform, there is a 9 in 10 chance it is invisible on the other.
What Predicts LLM Citations? (And What Doesn't)
A 7,000-citation analysis of 1,600 URLs revealed the following correlation strengths:
The #1 predictor of being cited by an LLM is how often people search for your brand by name. Not backlinks. Not domain rating. Brand search volume.
What GEO Research Tells Us
Princeton University's foundational GEO study (10,000 queries, 9 source types) found that specific content modifications drive citation rates significantly:
- Cite external sources: +115.1% visibility for sites ranked #5
- Add quotations: +37% improvement on Perplexity
- Add statistics: +22% improvement
- Fluency optimization: 15–30% boost
- Keyword stuffing: Negative impact
Crucially, lower-ranked sites benefit more from GEO optimization than sites already at the top — meaning citation visibility is more contestable than traditional search rankings.
The Attribution Crisis: Ecosystem Exploitation
A 2025 study introduced the concept of "ecosystem exploitation" — the gap between relevant web content consumed by an LLM when answering a query and the sources actually cited in its response. The findings are striking:
- Sonar (Perplexity): Citation gaps in 99.3% of queries, with 3.12 missing citations per query
- Gemini: Citation gaps in 70.3% of queries, with 3.04 missing citations
- GPT-4o: Only 10.5% probability of any citation gaps — but likely because OpenAI filters its search logs aggressively, disclosing only URLs it ultimately cites
This means LLMs routinely consume more web content than they attribute, raising questions about fair use, intellectual property, and the reliability of citations as a complete record of evidence used.
What This Means for Brands and Content Creators
If you want your content cited by AI search engines, the research points to a clear set of actionable strategies:
1. Build Brand Search Volume Brand search volume is the #1 predictor of LLM citations. Invest in brand awareness campaigns, PR, and off-site visibility. People need to search for your brand name before AI systems will cite it.
2. Cover the Full Query Ecosystem Depth beats position. A site ranking ninth but covering every sub-intent of a topic thoroughly may outperform a #1 site for one narrow query. Map the full cluster of related questions and answer each one explicitly.
3. Structure Content for Extractability AI engines score passages independently. Every paragraph (40–60 words) should be a self-contained citable unit. Lead with the answer, use question-format H2s, and include statistics, quotations, and outbound citations.
4. Use Structured Data Schema.org JSON-LD (Article, Organization, Person, FAQPage) helps AI engines parse authorship and topic without guessing. Valid schema correlates with 1.6–1.9x higher citation rates in YMYL topics.
5. Monitor Cross-Platform Gap With only 11% overlap across platforms, you cannot optimize for one engine and expect visibility everywhere. Use tools like Foxcite to track citation presence across ChatGPT, Claude, Gemini, Perplexity, and Grok simultaneously.
6. Prioritize Recency 65% of AI bot hits target content published within the past year. Freshness is a powerful signal. Stale content gets ignored.
The Road Ahead
LLM citation mechanisms are evolving rapidly. Post-hoc citation methods, retrieval-augmented validation frameworks (like CiteGuard, which approaches human-level accuracy at 68.1% on the CiteME benchmark), and agentic self-correction tools are all pushing toward more reliable attribution.
But the fundamental architecture remains: LLMs cite what they can retrieve, what they can verify across multiple queries, and what they can attribute to a trustworthy source. Brands that understand this pipeline — and optimize for its mechanics — will earn the citations that drive visibility in the answer engine era.
For those ready to measure and improve their AI citation footprint, Foxcite provides continuous, multi-engine tracking across ChatGPT, Claude, Gemini, Grok, and Perplexity — delivering citation gap intelligence, automated audit logs, and client-ready reports built for the new search landscape.
Written for Foxcite — AI Search Intelligence for the Answer Engine Era.
Foxcite Content Team
Contributor