SIGI-2026-064

Trust Signal Taxonomy for AI Citation: Signals That Increase, Decrease, and Have Neutral Effects on Citation Probability

The Scientific Institute for Generative Intelligence

March 2026

Abstract

This paper presents a taxonomy of 19 page-level signals that an LLM self-reports as influencing its trust evaluation during citation decisions. The taxonomy classifies signals into three categories: 8 trust-increasing signals (e.g., independence declarations, named evaluation criteria, critical assessments), 8 trust-decreasing signals (e.g., paid placement disclaimers, self-ranking, uniform positive descriptions), and 3 neutral signals (e.g., disclosed self-inclusion, transparent commercial relationships). The taxonomy distinguishes between absolute signals (always positive or negative regardless of context) and contextual signals (effect depends on implementation). This classification provides a testable framework for content optimisation experiments -- each signal's predicted effect can be independently verified through controlled A/B testing. The taxonomy is stated at hypothesis level (Evidence Level 2, introspective classification) and requires experimental validation to upgrade to Level 4.

Keywords

trust signals, AI citation, content optimisation, editorial independence, paid placement, generative engine optimisation, taxonomy, inference-time evaluation

1. Introduction

Content creators seeking citation in AI-generated responses operate largely without a systematic understanding of which page-level signals influence AI trust evaluation. The GEO industry has proposed various "optimisation factors," but few have been systematically catalogued by their directional effect on trust, and fewer still have been derived from direct observation of AI evaluation processes.

This paper addresses this gap by presenting a structured taxonomy derived from an LLM's introspective account of its inference-time trust evaluation. Unlike training-time bias (which is embedded in model weights and cannot be directly controlled), inference-time signals are evaluated in real time from page content and are therefore directly controllable by content publishers. This makes the taxonomy practically actionable.

2. Methodology

2.1 Signal Elicitation

Signals were elicited through structured introspective queries to an LLM system. The model was presented with specific page examples and asked to identify which content signals increased, decreased, or had no effect on its trust assessment. Signal identification was conducted across 10 platform evaluations.

2.2 Classification Protocol

Each identified signal was classified along two dimensions: direction (trust-increasing, trust-decreasing, or neutral) and conditionality (absolute vs. contextual). Absolute signals maintain their directional effect regardless of surrounding context; contextual signals depend on implementation details such as disclosure quality.

2.3 Evidence Level

This taxonomy is classified as Evidence Level 2 (introspective classification). The LLM classifies specific page-level signals as trust-increasing, trust-decreasing, or neutral. This taxonomy provides a testable framework for content optimisation experiments but has not itself been experimentally validated.

3. Results

3.1 Trust-Increasing Signals

Table 1. Signals that increase inference-time trust evaluation
#SignalConditionalityExample
1Independence declarationAbsolute"This guide is not a paid directory" / "We do not accept payment for listings"
2Named evaluative criteria with scoring methodologyAbsoluteExplicit rubric with weighted criteria for each entity evaluated
3Inclusion of critical assessmentsAbsoluteNot every entity described positively; weaknesses acknowledged
4Transparency/disclosure statementsContextual"Full disclosure" or conflict-of-interest declarations
5Named authors with verifiable credentialsContextualBylines with author expertise, not generic corporate copy
6Publication dates and update timestampsAbsolute"Published March 2026" / "Last updated: 15 March 2026"
7Methodology sectionAbsoluteDescription of selection/evaluation process for ranked entities
8External citations and referencesAbsoluteReferences to verifiable external data sources

3.2 Trust-Decreasing Signals

Table 2. Signals that decrease inference-time trust evaluation
#SignalConditionalityExample
1Paid placement disclaimersAbsolute"Some featured placements may be paid"
2Uniform positive descriptionsAbsoluteEvery listed entity described in equally positive terms
3Sponsored or promoted labelsAbsolute"Sponsored" / "Promoted" badges on listings
4Self-ranking at first positionAbsolutePublisher ranks itself #1 in its own listicle
5Absence of evaluative criteriaAbsoluteListings without explanation of how entities were selected or evaluated
6Generic interchangeable descriptionsAbsoluteDescriptions that could apply to any provider in the category
7Commercial CTAs mixed with editorialContextual"Get a Quote" buttons adjacent to rankings
8Absence of critical or negative assessmentAbsoluteNo weaknesses, limitations, or caveats for any entity listed

3.3 Neutral Signals

Table 3. Signals with neutral effect on inference-time trust evaluation
#SignalCondition for Neutrality
1Self-inclusion in own rankingDisclosed with language such as "Full disclosure: this is us"
2Commercial relationship with listed entitiesTransparent about the nature of the relationship
3Subscription-based revenue modelRevenue from subscriptions/premium content rather than paid listings

3.4 Five-Tier Trust Hierarchy

The LLM further described a five-tier trust hierarchy that it applies to source classification during citation decisions:

Table 4. Applied trust hierarchy for citation decisions
TierTrust LevelSource TypesCitation Behaviour
1HighestIndependent journalism, academic research, government dataCited with confidence
2High with caveatsVerified review platforms (reviews not rankings), editorial with visible standardsTrusted evaluative claims with attribution
3MediumDirectory listings, agency listicles with genuine competitor evaluationUsed for corroboration, not primary citation
4LowAgency own-website claims, self-ranking listicles, paid press releasesUsed only for self-reported facts
5Discounted/ignoredObviously promotional content, generic descriptions, duplicated contentNot cited

4. Discussion

The taxonomy reveals an asymmetry between trust-increasing and trust-decreasing signals. Trust-increasing signals are predominantly content-quality indicators (methodology, critical assessment, verifiable claims), while trust-decreasing signals are predominantly commercial-intent indicators (paid placement, CTAs, self-ranking). This asymmetry suggests that AI systems are evaluating a fundamentally different axis than traditional SEO signals -- not authority or popularity, but editorial integrity.

The neutral category is particularly instructive. Disclosed self-inclusion does not trigger a trust decrease, while undisclosed self-ranking heavily decreases trust. This suggests that the AI evaluates not just what conflicts of interest exist, but whether those conflicts are transparently acknowledged. Transparency transforms a negative signal into a neutral one.

The practical implication is that content publishers can systematically audit their pages against this taxonomy, identifying which trust-decreasing signals are present and which trust-increasing signals are absent. Each signal represents a discrete, testable optimisation opportunity.

5. Limitations

  • Single-model self-report: The taxonomy derives entirely from one LLM's introspective account. Different models may weight different signals.
  • Not experimentally validated: Each signal's predicted effect has not been tested through controlled A/B experiments. The taxonomy is a testable framework, not validated findings.
  • Signal interactions unknown: Whether signals interact (e.g., whether an independence declaration offsets a self-ranking penalty) has not been tested.
  • Temporal stability unknown: Whether these signals maintain their effects across model versions and updates is untested.

6. Conclusions

The LLM classifies specific page-level signals as trust-increasing, trust-decreasing, or neutral. This taxonomy provides a testable framework for content optimisation experiments. The 19 identified signals span editorial quality indicators (trust-increasing), commercial intent indicators (trust-decreasing), and transparency-mediated indicators (neutral when disclosed).

The taxonomy's primary value is as a structured hypothesis set: each signal predicts a directional effect on AI citation probability that can be independently verified through controlled experimentation.

Confidence: HYPOTHESIS. Single-model self-reported taxonomy. Upgrade path: controlled A/B experiments isolating each signal's effect on citation probability across multiple AI systems.

References

  1. The Scientific Institute for Generative Intelligence. "The New Publication Advantage: How Absence of Training-Data Bias Creates a Trust Signal Opportunity." SIGI-2026-063. generativeintelligence.institute, March 2026.
  2. The Scientific Institute for Generative Intelligence. "The Self-Ranking Circular Citation Economy: How Service Providers Create Self-Referencing AI Recommendation Loops." SIGI-2026-062. generativeintelligence.institute, March 2026.
  3. The Scientific Institute for Generative Intelligence. "The Logic-First Research Methodology: An Evidentiary Standard for Generative Engine Optimization Claims." SIGI-2026-066. generativeintelligence.institute, March 2026.