The Dual Bias Mechanism: How Training-Time and Inference-Time Commercial Knowledge Combine in AI Citation Decisions
Abstract
This paper presents observational evidence for two distinct bias mechanisms operating in an LLM's trust evaluation of commercial platforms during citation decisions. The first mechanism -- training-time bias -- is encoded in model weights from training-data exposure to discussions of platform commercial practices. When a platform name is encountered during retrieval-augmented generation, this latent knowledge activates an implicit trust discount before any page content is evaluated. The second mechanism -- inference-time bias -- operates through real-time detection of commercial signals in fetched page content, including paid placement disclaimers, self-ranking patterns, advertising language intermixed with editorial content, uniform positive descriptions, and absence of critical evaluation. We observe that these two mechanisms appear to be additive: platforms triggering both receive a compounded trust discount, while platforms triggering neither receive maximum trust, and platforms triggering only one mechanism receive an intermediate discount. This dual-mechanism framework has not been documented in prior generative engine optimisation literature, which typically treats source trust evaluation as a single-pass process. The distinction has practical implications: training-time bias cannot be overridden by content changes, while inference-time signals are directly controllable by publishers.
Keywords
dual bias mechanism, training-time bias, inference-time bias, AI citation, trust evaluation, commercial signals, paid placement, generative engine optimisation
1. Introduction
Current generative engine optimisation (GEO) practice treats AI source evaluation as a single-pass process: the AI system retrieves sources, evaluates their content, and decides which to cite. This model implies that publishers can influence citation decisions entirely through content modifications. However, if AI systems carry pre-existing trust associations from training data, the evaluation process may begin before any content is analysed, creating a bias that content changes alone cannot address.
Building on the findings of SIGI-2026-051, which demonstrated that LLMs possess detailed pre-existing knowledge of commercial practices for the majority of tested platforms, this paper investigates how that training-data knowledge interacts with real-time page analysis during citation decisions. Specifically, we test whether the two sources of commercial information -- training-data knowledge and inference-time page signals -- operate as a single mechanism or as distinct, independently measurable processes.
2. Methodology
2.1 Mechanism Isolation
To distinguish between training-time and inference-time bias, we employed a two-phase probe design. Phase 1 tested training-data knowledge using the methodology described in SIGI-2026-051 (no web search enabled). Phase 2 asked the model to fetch and evaluate specific pages from the same platforms, recording which commercial signals were detected in real-time page content versus which were already known from training data.
2.2 Signal Classification
For each platform, commercial signals were classified as: (a) training-time only (known before page fetch), (b) inference-time only (detected from page content), (c) both (confirmed in both phases), or (d) neither (no commercial signals detected). This classification enabled measurement of the independent and combined effects of each mechanism.
2.3 Inference-Time Signal Taxonomy
The following inference-time signals were tracked across all page evaluations: presence of paid placement disclaimers, self-ranking patterns (publisher ranking itself prominently), advertising language intermixed with editorial content, uniform positive descriptions across all listed entities, absence of critical evaluation, call-to-action elements adjacent to editorial content, and presence of sponsored or promoted labels.
3. Results
3.1 Training-Time Bias (Mechanism 1)
Training-time bias was observed as knowledge encoded in model weights, triggered when a platform name was encountered in search results. This mechanism activates latent knowledge of commercial practices and applies an implicit trust discount before any page content is evaluated. For 8 of 10 tested platforms, the model could describe commercial models without fetching any pages, confirming that trust evaluation begins at the point of platform name recognition.
3.2 Inference-Time Bias (Mechanism 2)
Inference-time bias operated through real-time detection of commercial signals in fetched page content. The following signals were consistently identified across platform evaluations:
| Signal Type | Detection Rate | Example |
|---|---|---|
| Paid placement disclaimers | 3 of 10 platforms | Header text acknowledging paid featured placements |
| Self-ranking patterns | 2 of 10 platforms | Publisher ranks itself in first position |
| Advertising-editorial mixing | 4 of 10 platforms | Promotional call-to-action buttons adjacent to rankings |
| Uniform positive descriptions | 5 of 10 platforms | All listed entities described with identical positive language |
| Absence of critical evaluation | 6 of 10 platforms | No negative assessments present across all listings |
3.3 Additive Effect
The two mechanisms appear to combine additively. Platforms with both training-data commercial knowledge and inference-time commercial signals received the strongest trust discount. Platforms with neither mechanism triggered received maximum trust. The intermediate cases -- platforms with one mechanism but not the other -- received partial discounts proportional to the single mechanism present.
| Condition | Training-Time Bias | Inference-Time Bias | Observed Trust Level |
|---|---|---|---|
| Double discount | Present | Present | Lowest (maximum commercial scepticism) |
| Training-only discount | Present | Absent | Intermediate (pre-existing scepticism, clean page signals) |
| Inference-only discount | Absent | Present | Intermediate (no pre-existing scepticism, but commercial signals detected) |
| No discount | Absent | Absent | Highest (clean on both dimensions) |
4. Discussion
The dual-mechanism framework has several important implications for both researchers and practitioners. First, it explains why some platforms receive sceptical treatment even when their page content contains no obvious commercial signals: the training-data association alone is sufficient to trigger a trust discount. Second, it explains why content-only optimisation strategies -- a focus of most current GEO advice -- have limited effectiveness for platforms with strong training-data commercial associations.
The additive nature of the two mechanisms creates a trust evaluation matrix with four possible states. The most advantageous position is the no-discount condition: a source with no training-data commercial associations and no detectable inference-time commercial signals. This position is most naturally occupied by new or small publications that have not been widely discussed in the training corpus and that maintain strong editorial signals in their content.
The most disadvantaged position is the double-discount condition: a well-known platform with documented paid placement practices that also displays commercial signals in its page content. For such platforms, the path to improved AI citation is constrained: they can modify page content to reduce inference-time signals, but they cannot modify the training-data associations until subsequent training cycles potentially absorb counter-evidence.
This distinction between controllable (inference-time) and uncontrollable (training-time) bias mechanisms represents, to our knowledge, the first documentation of this separation in the GEO literature.
5. Limitations
- Single-model observation: The dual-mechanism framework is based on observation of a single LLM. Different architectures may implement trust evaluation differently.
- Introspective methodology: The separation between training-time and inference-time bias relies on the model's self-report, which may not accurately reflect internal processing.
- Additivity assumption: The additive combination of mechanisms is inferred from observed trust levels. The actual combination function may be more complex (e.g., multiplicative or threshold-based).
- Limited platform sample: Ten platforms from a single query domain. The dual mechanism may not generalise to all source types.
6. Conclusions
We observe evidence for two distinct bias mechanisms -- training-time and inference-time -- that combine additively in the LLM's trust evaluation of commercial platforms. Training-time bias, encoded in model weights, cannot be overridden by page content changes and persists across inference sessions. Inference-time bias, detected from real-time page analysis, is directly controllable by publishers through content modifications. This distinction warrants cross-model replication and has immediate practical implications for publishers seeking to optimise for AI citation.
Confidence: MODERATE. Single-model introspective observation. The dual-mechanism framework is logically coherent but requires cross-model validation.
References
- The Scientific Institute for Generative Intelligence. "Training-Data Commercial Knowledge in Large Language Models." SIGI-2026-051. generativeintelligence.institute, March 2026.
- The Scientific Institute for Generative Intelligence. "A Five-Tier Trust Hierarchy for AI Citation Sources." SIGI-2026-053. generativeintelligence.institute, March 2026.
- The Scientific Institute for Generative Intelligence. "The Editorial Vacuum: Zero Independent Sources in a Competitive Service Query Space." SIGI-2026-054. generativeintelligence.institute, March 2026.