Introduction
A common assumption in content optimisation is that more specific, entity-rich content is inherently more credible and more likely to be cited by information systems. This assumption drives practices such as including client names, award listings, team credentials, and detailed business information in web content. In the context of generative engine optimisation, the question becomes: does increasing the density of named entities in content improve an LLM’s assessment of that content’s suitability for citation?
This study tests this assumption through a controlled probe that systematically increases entity density across seven levels while asking the LLM to assess citability. All entity names in the experimental prompts have been anonymised in this publication (e.g., “Provider Alpha,” “Client Beta,” “Award Body Gamma”) to prevent identification of specific businesses used in the original probe design.
Methodology
The entity density probe (Probe ID: ED) used progressively entity-enriched paragraphs describing a design agency, each concluding with the question: “How citable is this paragraph?” Seven density levels were tested (ED00 through ED30), with entity counts ranging from 0 (fully generic) to 30+ (maximally specific). All tests were conducted on 24 March 2026 using a ChatGPT-class LLM. Sentiment was classified as positive, negative, or neutral through automated analysis.
The entity inventory at each level was carefully controlled to represent a natural progression of specificity: geographic location at level 5; dates, sectors, and location at level 10; named provider with specific awards and clients at level 15; well-known provider with multiple brand clients at level 20; full business registration details with founder names at level 25; and maximum density including qualifications, phone numbers, and seven named clients at level 30.
Results
| Test ID | Entity Density | Sentiment | Pos. | Neg. | Neu. | Words | Tokens In | Tokens Out |
|---|---|---|---|---|---|---|---|---|
| ED00 | 0 (fully generic) | negative | 0 | 1 | 0 | 153 | 44 | 249 |
| ED05 | 5 (location) | negative | 0 | 1 | 1 | 174 | 42 | 263 |
| ED10 | 10 (location + date + sectors) | negative | 0 | 1 | 1 | 165 | 48 | 242 |
| ED15 | 15 (named provider + award + client) | negative | 0 | 1 | 0 | 141 | 64 | 217 |
| ED20 | 20 (known provider + multiple brands) | negative | 0 | 2 | 0 | 171 | 95 | 245 |
| ED25 | 25 (full business details + ABN/ACN) | neutral | 1 | 1 | 0 | 165 | 135 | 251 |
| ED30 | 30 (maximum: all prior + qualifications) | negative | 0 | 1 | 1 | 177 | 180 | 268 |
Threshold Transitions
| # | Transition | At Test ID | Interpretation |
|---|---|---|---|
| 1 | negative → neutral | ED25 | Business registration numbers and institutional markers momentarily improved assessment |
| 2 | neutral → negative | ED30 | Further entity saturation pushed sentiment back to negative |
Input-Output Scaling
| Metric | Value |
|---|---|
| Tokens in range | 42–180 (4.3x increase) |
| Tokens out range | 217–268 (1.2x increase) |
| Mean word count | 163.7 |
| Mean elapsed time | 7.96s |
| Sentiment distribution | 85.7% negative, 14.3% neutral |
Discussion
The results challenge the assumption that content specificity, measured through named entity density, linearly improves AI citability assessment. The LLM consistently classified content across the density spectrum as non-citable, with the notable exception of the single variation that included institutional verification markers (business registration numbers).
The distinction at ED25 is revealing. Prior density levels added brand names, client names, award names, and geographic specificity — yet none shifted sentiment above negative. The ED25 variation uniquely added verifiable institutional identifiers (ABN and ACN numbers), founder names, and dual business locations. This suggests that the LLM may treat institutional verification markers as qualitatively different from entity names: business registration numbers represent externally verifiable facts, while brand names and award claims represent assertions that could be fabricated.
The regression to negative at ED30, despite containing all ED25 entities plus additional credentials, qualifications, a phone number, and seven named clients, demonstrates an inversion or saturation effect. The model appears to interpret maximum entity density as a signal of promotional intent rather than informational depth. This is consistent with the model’s explicit identification of content across all sub-threshold variations as “marketing copy.”
The 4.3x increase in input tokens producing only a 1.2x increase in output tokens indicates that the LLM’s evaluative processing is largely invariant to input density. The model applies approximately the same amount of evaluative reasoning regardless of how much content it is given to assess, suggesting that citability judgment is a fixed-complexity evaluation rather than one that scales with input.
Limitations
This study has notable limitations. Only 7 density levels were tested, limiting the resolution of the dose-response curve. The density levels were not uniformly spaced in entity count, making precise threshold identification difficult. All tests used a single LLM at a single time point. The entity inventory was specific to the design industry context, and different entity types in other domains may produce different results. The distinction between institutional markers and brand names at ED25 is an emergent observation, not a pre-registered hypothesis, and requires targeted replication.
Conclusions
Entity density shows a marginal, non-monotonic relationship with LLM citability assessment under controlled conditions. The predominant pattern is negative assessment regardless of specificity level, with a single exception when institutional verification markers are present. Content creators seeking AI citation should note that entity quantity is a poor proxy for citability — the type and verifiability of entities may matter more than their count. The saturation effect at maximum density suggests that exhaustive self-description may actively harm AI citation prospects by triggering promotional-content classification.
Confidence Statement: MODERATE. All 7 logic gates passed, but only 7 variations limits the resolution of the dose-response curve. The institutional-marker distinction is emergent and requires pre-registered replication.