The Signal Access Matrix: Determining Which Factors LLMs Can Actually Evaluate at Inference Time
Abstract
A persistent problem in generative engine optimisation research is the attribution of causal influence to signals that large language models cannot access at the point where citation decisions are made. This paper presents the Signal Access Matrix, a systematic framework for evaluating whether claimed GEO factors are accessible to LLMs through any of four channels: training data, search results, page content at inference, or derived inference. We map commonly claimed signals across these channels and classify each for causal plausibility. The analysis reveals that several factors commonly discussed in GEO literature -- including site speed, domain age as a direct signal, and certain technical SEO metrics -- fail the basic access test: the LLM has no pathway to evaluate them during response generation. Conversely, factors such as brand mentions, content depth, star ratings, and entity density pass with high plausibility because they are present across multiple access channels. The matrix further distinguishes training-time signals (baked into model weights, uncontrollable after model release) from inference-time signals (present in content processed during generation, controllable today). This distinction has immediate practical implications: of the highest-impact trust signals identified in companion research, 9 of 10 are primarily inference-time, meaning they can be influenced through content modification.
Keywords
signal access, inference time, training time, causal plausibility, LLM citation, generative engine optimisation, mechanism validation, trust signals, content accessibility, GEO methodology
1. Introduction
The seventh gate of the Logic-First Research Methodology poses a deceptively simple question: does the LLM actually have access to this signal at inference time? If the answer is no, then regardless of any observed correlation between the signal and citation outcomes, the signal cannot be a direct cause of those outcomes.
This question is more discriminating than it appears. Many factors discussed in GEO literature are borrowed from traditional search engine optimisation, where the ranking system has access to a wide array of signals through dedicated crawling infrastructure. An LLM operating in a retrieval-augmented generation pipeline has a fundamentally different information access profile. It does not have direct access to server response times, WHOIS records, or backlink profiles during response generation. Its information inputs are limited to what appears in its training data, what is returned by its search subsystem, and what content it can extract from pages at inference time.
The Signal Access Matrix formalises this constraint into a practical evaluation tool. For each claimed signal, the matrix asks four questions: Is this signal present in training data? Is it visible in search results? Is it available in page content at inference? And given these answers, is the claimed causal pathway plausible?
2. The Four Access Channels
2.1 Training Data Presence
Signals embedded in the model's pre-training corpus influence behaviour through learned associations. Brand names, domain reputations, and common factual associations are examples of training-time signals. These signals are frozen at the training cutoff date and cannot be modified by the site owner after model release. New entities absent from training data have zero training-time signal regardless of their current quality.
2.2 Search Result Visibility
In retrieval-augmented generation, search results provide a real-time information channel. Signals visible in search snippets, titles, and URLs can influence citation decisions. However, many page-level signals (schema markup, full content structure, internal linking) are typically not visible in search result snippets.
2.3 Page Content at Inference
When an LLM retrieves and processes a full page, all text-based content becomes available. This includes headings, paragraph text, statistics, entity names, and sometimes structured data. However, signals that exist only in non-text formats (server headers, JavaScript-rendered content, image metadata) may not be accessible depending on the crawler implementation.
2.4 Derived Inference
Some signals are not directly present but can be inferred from available information. For example, content quality is not a discrete signal but is derived from the LLM's evaluation of specificity, coherence, and information density in the available text. Derived signals require that the underlying information be present through one of the three direct channels.
3. The Signal Access Matrix
| Signal | In Training Data | In Search Results | In Page Content | Causal Plausibility |
|---|---|---|---|---|
| Brand mentions / entity name | Yes (if established) | Yes | Yes | Highly plausible |
| Star ratings | Yes (if aggregated) | Yes (snippets) | Yes (if displayed) | Highly plausible |
| Content depth / word count | Yes (via training) | Partial (snippet) | Yes | Highly plausible |
| Entity density | Yes (via training) | Partial (snippet) | Yes | Highly plausible |
| Statistics and specific numbers | Yes | Partial | Yes | Highly plausible |
| Schema markup | Possibly (if in crawled HTML) | Rarely visible | If crawler reads HTML | Plausible |
| robots.txt directives | Not typically | Not visible | Determines access | Plausible (binary gate) |
| Heading structure (H2 format) | No | No | Yes | Plausible (inference-time only) |
| Author attribution | If established author | Partial | Yes | Plausible |
| Domain age (as direct signal) | In WHOIS (not typically crawled) | Not in snippets | Not in page content | Implausible as direct signal |
| Site speed / server response | Not in training data | Not in search results | Not perceptible to LLM | Implausible |
| Backlink profile | Indirect (via authority) | Not visible | Not visible | Implausible as direct signal |
| URL structure / cleanliness | Visible in URLs | Visible in URLs | Visible | Low plausibility |
| No-paid-placement declaration | No | No | Yes (if stated) | Plausible (inference-time) |
| Methodology section | No | No | Yes | Plausible (inference-time) |
4. Training-Time Versus Inference-Time Classification
The distinction between training-time and inference-time signals has significant practical implications. Training-time signals are embedded in model weights and cannot be changed by content modification after the model's training cutoff. Inference-time signals are present in the content the model processes during response generation and can be influenced today.
Analysis of the 77 trust signals evaluated in companion research reveals that 9 of the top 10 highest-impact signals are primarily inference-time. These include content structure patterns (direct answers in first sentences, question-format headings), independence declarations (no-paid-placement statements), specificity markers (statistics, entity density), and methodological transparency (methodology sections). Only one top-10 signal -- proprietary data originating from an established source -- has a significant training-time component.
This finding suggests that content optimisation for AI citation is primarily an inference-time activity. The implication is that newly published content can compete with established sources by excelling on inference-time signals, even without the training-data advantage that established brands possess. This is consistent with the controlled probe finding that domain age has zero effect on LLM sentiment under isolated conditions (see SIGI-2026-011).
5. Implications for GEO Research Claims
The Signal Access Matrix provides a necessary condition test for any claimed GEO factor. If a factor does not pass the access test -- if the LLM cannot perceive it through any available channel -- then claims about its causal influence are unsupported regardless of observed correlations. This does not mean the factor is unimportant in a broader context; site speed may influence user behaviour, search engine rankings, or other intermediary outcomes. But it cannot be a direct cause of LLM citation decisions if the LLM has no access to the information.
Researchers claiming causal influence for a particular signal should be required to specify the access channel through which the LLM encounters it. If no plausible access channel can be identified, the claim should be reclassified as operating through an indirect pathway (e.g., site speed influences search ranking, which influences search result inclusion, which influences citation) and the additional causal steps should be made explicit.
6. Conclusions
We present a signal access matrix that tests whether claimed GEO factors are accessible to LLMs at the inference point where citation decisions are made. The matrix maps each signal across four access channels and classifies causal plausibility accordingly. Several commonly discussed factors -- including site speed, domain age as a direct signal, and backlink profiles -- fail the access test, while content-level factors such as entity density, statistics, and independence declarations pass with high plausibility. The training-time versus inference-time distinction reveals that the majority of high-impact signals are controllable through content modification, providing a practical focus for optimisation efforts.
Confidence: Framework paper. The access classifications are based on architectural analysis of LLM retrieval-augmented generation pipelines and are subject to revision as LLM capabilities evolve.
References
- The Scientific Institute for Generative Intelligence. "Domain Age as a Signal in LLM Evaluation: A Controlled Null-Result Experiment." SIGI-2026-011. generativeintelligence.institute, March 2026.
- The Scientific Institute for Generative Intelligence. "The Claim Audit Checklist: A Pre-Publication Quality Gate for GEO Research." SIGI-2026-071. generativeintelligence.institute, March 2026.
- The Scientific Institute for Generative Intelligence. "Necessary, Sufficient, and Contributory: Applying the INUS Framework to GEO Factor Analysis." SIGI-2026-073. generativeintelligence.institute, March 2026.