A Taxonomy of 77 Trust Signals in AI Citation Decisions: An Introspective Framework from a Single Large Language Model

What Is the Purpose of a Trust Signal Taxonomy for AI Citation?

As large language models increasingly serve as intermediaries between users and information sources, understanding which signals influence citation decisions has become a matter of practical significance for web publishers. The emerging field of Generative Engine Optimisation (GEO) lacks a comprehensive signal framework analogous to the ranking factor taxonomies that guided traditional search engine optimisation for two decades. This paper attempts to address that gap through the most granular approach available given current methodological constraints: structured introspective elicitation from a single LLM.

We observed that when prompted through a systematic elicitation protocol, the model identified 77 discrete signals across 10 categories, each scored on a 1-to-10 scale for estimated citation impact. This taxonomy represents the model's self-reported assessment of its own decision-making process. We emphasise that LLMs are known to confabulate explanations for their own behaviour, and these findings should be treated as informed hypotheses generating testable predictions rather than validated conclusions about actual model behaviour.

How Were the 77 Trust Signals Identified and Scored?

The elicitation was conducted across three structured sessions with a single large language model. Each session presented the model with specific content categories and asked it to identify, score, and explain signals that influence its source selection when generating cited responses. Session one covered domain, infrastructure, crawl, and schema signals (Signals 1-27). Session two covered content structure, authorship, provenance, and initial commercial signals (Signals 28-50). Session three covered social proof, ecosystem, uniqueness, and remaining commercial signals (Signals 51-77), along with a final consolidated ranking of all 77 signals.

Each signal was scored on four dimensions: impact (1-10 scale), direction (positive, negative, or neutral), processing layer (training-time, inference-time, or both), and confidence level (high, medium, or low based on the model's self-assessed consistency of response). We observed that the model maintained internal consistency across sessions, with signals referenced in later sessions receiving scores consistent with their initial assessments. However, this consistency may reflect the model's coherence in narrative construction rather than accurate self-knowledge.

What Are the 10 Highest-Impact Trust Signals?

The model's self-reported top 10 signals span four categories: Content Uniqueness, Commercial Independence, Content Structure, and Authorship/Provenance. We observed that no Social Proof, Schema, or Domain signals appear in the top 10, suggesting these categories may function as supporting infrastructure rather than primary citation drivers.

How Are the 77 Signals Distributed Across Categories?

We observed that the 10 categories exhibit substantial variation in average signal scores. Content Uniqueness emerged as the highest-scoring category with an average of 8.3 across 5 signals, while Crawl Configuration was the lowest at 2.7 across 7 signals. This suggests that the model assigns disproportionate weight to content-level signals over technical infrastructure signals.

What Negative Signals Were Identified?

We observed that the model identified 8 signals with negative directional polarity, meaning their presence is reported to actively reduce citation probability rather than simply failing to contribute positively. The highest-scored negative signals include JavaScript dependency (7.5), sponsored content labels (7.5), and syndicated content on low-authority domains (7.5). The model reported that negative signals function asymmetrically: a single strong negative may outweigh multiple moderate positive signals. This asymmetry claim remains unquantified and represents a hypothesis requiring experimental testing.

What Is the Training-Time Versus Inference-Time Distribution?

Each of the 77 signals was classified by the model as operating at the training-time layer (fixed knowledge baked into model weights), the inference-time layer (evaluated when processing a retrieved document during RAG), or both. We observed that 9 of the 10 highest-impact signals were classified as inference-time, with only domain age having a significant training-time component among the top tier. This suggests that if the model's self-report is accurate, most high-impact citation signals are directly controllable through page-level content and markup decisions rather than requiring long-term brand building or domain authority accumulation.

What Methodology Was Employed in This Research?

The study employed structured introspective elicitation, a method in which a large language model is prompted to identify, score, and explain the signals it uses when making citation decisions. Three sessions were conducted with a single LLM, each covering a distinct signal category set. The model was asked to provide impact scores (1-10), directional polarity (positive, negative, neutral), processing layer (training-time, inference-time, or both), confidence level (high, medium, low), and a prose explanation for each signal. Results were consolidated into a ranked taxonomy across the three sessions.

This methodology has significant limitations that are detailed in the Limitations section below. Introspective self-report is the weakest form of evidence in behavioural science, and LLMs have no guaranteed introspective access to their own computational processes. The methodology was chosen because no non-introspective alternative currently exists for generating comprehensive signal inventories at this granularity.

What Results Emerged from the Complete Framework?

Across the 77 signals, scores ranged from 1.0 (HTTPS baseline, crawl-delay configuration, sitemap priority values) to 9.5 (proprietary data and original research). The distribution was right-skewed: 10 signals scored 8.0 or above, 17 scored between 6.5 and 7.5, 22 scored between 4.5 and 6.0, and 28 scored below 4.5. This distribution pattern is consistent with the model's own claim of a Pareto-like concentration of citation impact in a small number of signals.

We observed that the model consistently valued specificity and verifiability over general quality claims. Signals that enable external verification (named authors, specific statistics, methodology disclosures, platform-verified reviews) scored systematically higher than signals that assert quality without verification pathways (generic testimonials, self-reported awards, design quality claims). This pattern, if confirmed behaviourally, would suggest that LLM citation decisions may approximate an evidence-based credibility assessment rather than a popularity or authority metric.

Discussion: What Do These Findings Suggest About AI Citation Behaviour?

The taxonomy suggests several patterns that, if validated through behavioural testing, would have significant practical implications. The dominance of inference-time signals in the top tier suggests that citation behaviour may be substantially controllable through content-level optimisation. The high scores assigned to commercial independence signals suggest that LLMs may apply trust discounts to commercially conflicted sources more aggressively than traditional search engines. The concentration of impact in a small number of signals suggests that a focused optimisation strategy targeting 10-15 signals may yield greater returns than comprehensive optimisation across all 77.

We also observed that several signals widely prioritised in current GEO practice scored low in the taxonomy: site speed (1.5), URL structure cleanliness (2.0), crawl-delay configuration (1.0), and sitemap priority values (1.0). This suggests that current GEO guidance may overweight technical infrastructure signals relative to their actual citation impact, though this assessment relies on the model's self-report and requires independent confirmation.

What Are the Limitations of This Research?

This study has several significant limitations. First, introspective self-report is evidence level 1-2, the weakest tier in the SIGI evidence hierarchy. LLMs confabulate explanations for their own behaviour, and the model's reported signal weightings may not correspond to its actual computational processes. Second, the study reflects a single model at a single time point. Other models (and future versions of the same model) may weight signals differently. Third, the scoring methodology lacks external validation. The 1-10 scores are ordinal rankings reflecting the model's stated preferences, not measured effect sizes from controlled experiments. Fourth, the elicitation prompt structure may have influenced the signals identified. A differently structured elicitation might produce a different taxonomy. Fifth, the Pareto estimate (top 10 signals accounting for approximately 80% of impact) is the model's own approximation, not a measured distribution.

Conclusions

This taxonomy of 77 trust signals represents the most granular mapping currently available of factors that a single LLM self-reports as influencing its citation decisions. The framework identifies a concentrated set of high-impact signals dominated by content uniqueness, commercial independence, and content structure categories. These findings should be treated as a hypothesis-generating framework that prioritises signals for subsequent behavioural testing, not as validated conclusions about AI citation mechanics. Independent experimental validation using controlled probe methodologies is required to confirm whether the model's self-reported signal weightings correspond to its actual citation behaviour.