Multiple independent research efforts published in 2026 converge on a structural reality that Machine Relations has formalized as entity chains: AI search engines do not cite based on keywords, backlinks, or domain authority alone. They cite based on how a concept is named, sourced, corroborated, and structurally linked across multiple independent surfaces.
None of these studies use the term "entity chain." All of them describe the mechanism.
This article traces the convergence — from 680-million-citation platform analyses to controlled entity density experiments to graph traversal ablation studies — and explains why these findings matter for any brand trying to earn and hold AI visibility.
680 Million Citations: Platform Divergence Proves Multi-Domain Strategy #
The largest published citation dataset — Profound's tracking of 680 million AI-generated citations across ChatGPT, Google AI Overviews, and Perplexity from August 2024 through June 2025 — reveals a structural pattern that entity chains predict: each AI platform cites radically different source distributions.
Within ChatGPT's top 10 most-cited sources, Wikipedia holds 47.9% share. For Perplexity, Reddit holds 46.7%. Google AI Overviews distributes more evenly, with Reddit at 21.0%, YouTube at 18.8%, and Quora at 14.3%. An independent analysis by Rankscale.ai of 8,000 citations across 57 queries, published on Search Engine Land, confirms the same divergence: ChatGPT functions as an "authority seeker," Google AI Overviews as a "balanced synthesizer," and Perplexity as a community-weighted aggregator.
The entity chain implication is direct: a brand visible on only one platform's preferred source type is structurally exposed to the others. A Wikipedia presence helps with ChatGPT but does little for Perplexity's Reddit-weighted model. A brand mentioned across Wikipedia, Reddit, industry publications, review platforms, and earned media — a dense entity chain — maintains citation eligibility across all three platforms simultaneously.
This platform divergence also explains why citation share is volatile within weeks. As Profound's data shows, ChatGPT's Reddit citation share fell from roughly 60% to 10% in six weeks after a single parameter change. The displaced citations redistributed to other nodes — PR Newswire, Forbes, Medium — not to entirely new entities. Brands with dense cross-domain entity chains absorbed the volatility. Brands with thin chains lost visibility entirely.
Citation Absorption Research: Structure Predicts Influence #
The most rigorous empirical study of AI citation behavior in 2026 is Yao et al.'s "From Citation Selection to Citation Absorption," which analyzed 602 controlled prompts across ChatGPT, Google AI Overview/Gemini, and Perplexity, producing 21,143 search-layer citations and 72 extracted page features. The full dataset and analysis pipeline are publicly available on GitHub.
Their central finding: citation breadth and citation depth diverge. Perplexity and Google cite more sources per query. ChatGPT cites fewer sources but shows substantially higher average citation influence per source. The pages that achieve high influence — meaning their language, evidence, structure, and factual content are absorbed into generated answers — share specific structural properties:
- Longer, more structured content
- Higher semantic alignment with the query
- Richer extractable evidence: definitions, numerical facts, comparisons, procedural steps
This aligns with Yang et al.'s GEO-SFE framework, which decomposes content structure into three hierarchical levels — macro-structure (document architecture), meso-structure (information chunking), and micro-structure (visual emphasis) — and demonstrates a 17.3% improvement in citation rate across six generative engines through structural optimization alone, without changing semantic content.
Both findings describe the entity chain pattern from the retrieval side. A page achieves citation influence not by ranking well in a link graph, but by presenting structured, extractable information that a retrieval system can verify against other sources. The more structured and evidence-dense the page, the more nodes it provides for cross-referencing — exactly what entity chain density measures.
Yao et al. recommend that "GEO should be measured beyond citation counts, with answer-level absorption treated as a separate outcome." In entity chain terms: getting cited is not the goal. Getting absorbed is. And absorption requires the kind of cross-domain evidence architecture that entity chains formalize.
The Domain Authority Gate: Entity Density Operates Within Tiers #
Artur Ferreira's controlled experiment at The GEO Lab tested a direct corollary of entity chain theory: does entity density — the count of unique named entities per 1,000 words — predict citation rate?
The experiment used 52 queries, 10 pages with entity density ranging from 12.9 to 29.5 unique entities per 1,000 words, and ran against Perplexity sonar-pro. The result: 20% citation rate on proprietary concept queries and 0% on all 22 competition queries. Zero. Across every page and every density level.
This is not a refutation of entity chains. It is a confirmation of one of their most important properties: domain authority functions as a binary gate. Below a threshold of cross-domain corroboration, page-level content signals — including entity density — never reach the citation decision layer.
Ferreira frames the refined hypothesis precisely: "entity density differentiates within an authority tier." A brand that has crossed the citation eligibility threshold can use entity density to compete for specific queries. A brand that has not crossed that threshold will not be cited regardless of page quality.
Independent industry analysis corroborates this gating effect. An LLM visibility assessment drawing on Profound's December 2024 crawler study found that "classic SEO metrics don't strongly influence AI chatbot citations" — LLMs prioritize content depth, brand popularity, and source diversity instead. Kevin Indig's analysis of brand mentions in ChatGPT identified brand search volume (correlation of .542 with ChatGPT mentions) as the strongest single predictor — a proxy for the same cross-domain recognition that entity chains measure.
This maps directly to the entity chain framework's two-layer model. First, build cross-domain citation eligibility through independent corroboration. Then, optimize page-level entity density within that eligibility tier.
Graph Traversal Context: Uncited Neighbors Influence Citations #
Terrenzi et al.'s "Traversal Context and Provenance in Agentic GraphRAG" addresses entity chains from an entirely different angle — the knowledge graph traversal that happens before an AI system selects its citations.
In agentic GraphRAG systems, an agent explores a knowledge graph before producing an answer. Through controlled ablation experiments — isolating, removing, and masking cited and uncited graph entities — the study demonstrates two findings:
- Cited evidence is necessary: removing cited sources substantially changes answers and reduces accuracy.
- Citations are not sufficient: accurate answers also depend on uncited traversal context and surrounding graph structure.
This is the entity chain mechanism observed at the infrastructure level. When an AI system traverses a knowledge graph to answer a question, it does not evaluate each candidate source in isolation. It evaluates each source in the context of its neighbors — the other entities, mentions, and corroborations that surround it in the graph. A brand with dense cross-domain entity chain presence provides richer traversal context, even when specific nodes in the chain are not directly cited.
Fu et al.'s SEARCH-R framework, which uses structured entity-aware retrieval for multi-hop question answering, reinforces this from the engineering side. Their dependency tree-based retrieval evaluates "the practical utility of the information" rather than relying on similarity scores alone — precisely the kind of evidence-chain reasoning that entity chains are designed to support.
The conclusion from both studies — that citation evaluation should account for the broader evidence neighborhood, not just the cited source — is a formal description of what entity chains measure: not whether a single page is good enough, but whether the entire evidence neighborhood around a brand is dense enough to support citation.
Citation Granularity and Reliability: Why Chains Outperform Isolated Signals #
Two additional studies address entity chains from the perspective of citation quality and resilience.
Wang et al.'s "Are Finer Citations Always Better?" analyzed citation quality across four model scales (8B to 120B parameters) and found that attribution quality peaks at paragraph-level granularity — not sentence-level. Fine-grained sentence-level citations degrade attribution quality by 16–276% compared to paragraph-level. The reason: sentence-level citations "disrupt necessary semantic dependencies for attributing evidence to answer claims." Models need multi-sentence context — definitions connected to evidence connected to conclusions — to synthesize reliable attributions.
This validates the entity chain principle that isolated signals do not drive citations. A single mention, a single statistic, a single branded sentence does not create citation eligibility. What creates eligibility is a connected evidence structure — a paragraph, a section, a page — where each claim supports and is supported by neighboring claims.
Rao et al.'s study of reference hallucinations in LLMs and deep research agents — analyzing 53,090 URLs from DRBench and 168,021 URLs from ExpertQA across 32 academic fields — reveals that 3–13% of citation URLs are entirely hallucinated. Non-resolving rates vary dramatically by domain, from 5.4% in Business to 11.4% in Theology, with pronounced per-model variation. Brands that rely on a single high-authority surface for AI visibility are exposed to these URL-level failure modes. A multi-domain entity chain provides redundancy — if one citation surface breaks, other nodes in the chain sustain citation eligibility.
The emerging LLM visibility measurement ecosystem — tracked by platforms like Rankscale.ai and visibility tools profiled by QuickSEO — increasingly measures exactly the multi-surface presence that entity chains formalize: not just whether a brand appears in one engine's answer, but whether it maintains citation eligibility across platforms, query types, and time.
The Convergence #
These studies were conducted independently, across different institutions and methodologies. None references Machine Relations or entity chains. Yet their findings converge on the same structural mechanism:
- Profound (680M citations): Citation distributions diverge radically across platforms; multi-domain presence is required for cross-platform eligibility.
- Yao et al.: Citation influence depends on structure, evidence density, and semantic alignment — not link authority.
- Yang et al.: Structural optimization alone produces a 17.3% citation rate improvement across six engines.
- Ferreira: Domain authority gates citation eligibility; entity density differentiates within the gate.
- Terrenzi et al.: Citation decisions depend on graph traversal context, not just cited sources.
- Fu et al.: Entity-aware retrieval evaluates practical utility through dependency trees, not similarity scores.
- Wang et al.: Attribution quality requires multi-sentence semantic dependencies, not isolated signals.
- Rao et al.: Citation reliability is domain-dependent; multi-surface presence provides resilience.
Each finding describes one dimension of what the entity chain framework unifies: the mechanism by which AI search engines decide which brands, concepts, and claims deserve citation. The framework is not a marketing theory. It is a structural description of how retrieval-augmented generation works — validated by independent research that approaches the same problem from eight different angles.
For brands operating in AI-visible markets, the practical consequence is clear: building entity chains is not optional optimization. It is the documented mechanism by which AI engines decide who gets cited and who gets ignored.