1. Introduction and Scope
This paper catalogues 155 distinct empirical observations generated through systematic probing of AI search pipeline behaviour. Unlike surveys of published literature or theoretical frameworks, these observations were obtained through direct experimentation: queries were submitted to AI platforms under controlled conditions, and the resulting entity selections, source citations, confidence signals, and linguistic patterns were recorded and classified.
The catalog is structured for reference use. Each category summarises the observation type, count, and evidence level. Detailed explorations of individual findings are published separately (SIGI-2026-084 through SIGI-2026-088). Researchers and practitioners can use this catalog to identify relevant findings by category and evidence strength.
2. Methodology Overview
Four experimental methods were employed across the research programme:
Controlled pipeline probes (n=129): Structured queries designed to isolate specific pipeline behaviours, including recommendation architecture variations, framing effects, position stability tests, absence detection, source authority hierarchies, contradiction resolution, and entity emergence thresholds. Each probe type was administered in 7-19 variations.
Focused observations (n=10): Extended single-session investigations of specific phenomena identified during controlled probing, such as the circular citation economy and the self-authored position lock pattern.
Contextual queries (n=35): Queries designed to test how contextual modifiers (geographic, temporal, budgetary, quality-related) alter entity selection and source composition.
Pure numerical tests (n=87): Queries using bare numeric formats (integers, percentages, currency amounts, ratings, years) to investigate how AI systems process and route numeric inputs.
3. Catalog Summary by Category
| Category | Observations | Evidence Level | Key Theme |
|---|---|---|---|
| A: Pipeline Behaviour & Entity Routing | ~25 | 3-4 | 8-stage pipeline model, query interpretation, search augmentation patterns |
| B: Trust Signal Hierarchy | ~20 | 3-4 | Signal ranking, threshold effects, contradiction rules |
| C: Content Structure & Formatting | ~20 | 2-3 | BLUF effectiveness, heading impact, entity density requirements |
| D: Competitive Intelligence | ~20 | 2-3 | Market concentration, platform dominance, entry barriers |
| E: Position Dynamics | ~15 | 3-4 | Position lock patterns, instability factors, framing effects |
| F: Methodological Insights | ~15 | 3-4 | Self-observer technique, probe design, classification validity |
| G: Entity Recognition Thresholds | ~25 | 3-4 | Emergence curve, suppression mechanisms, verification requirements |
| H: Special Numeric Format Findings | ~15 | 3 | Format capture, temporal routing, cultural entity override |
4. Selected High-Confidence Findings
4.1 Pipeline Behaviour
AI search pipelines operate through at least 8 discrete processing stages, from query interpretation through final presentation. Current optimisation practices address only 2-3 of these stages, leaving the majority of the pipeline unoptimised. Price anchors in queries activate entirely different entity pools than quality anchors, and the search augmentation stage issues 2-4 reformulated queries that serve primarily as confirmation mechanisms rather than discovery tools.
4.2 Trust Signal Hierarchy
Press coverage represents the single largest trust signal jump observed (+27 confidence points in the entity emergence experiment). The minimum viable trust stack for AI recommendation inclusion consists of one verified business registration signal plus two independent third-party validation signals. Entities below this threshold are suppressed regardless of on-page content quality.
4.3 Entity Recognition Thresholds
Entity emergence follows a non-linear curve. Entities are invisible to AI recommendations until a critical mass of trust signals is accumulated, then lock in rapidly. The confidence curve progresses from 0% (name only) through 5% (website exists), 15% (business registration), 22% (verified portfolio), 35% (third-party reviews), and jumps to 62% with trade press coverage. The threshold for reliable AI inclusion sits at approximately 62%.
4.4 Suppression Mechanisms
Seven distinct suppression mechanisms were identified: training data absence, external discourse absence, informationally closed systems, template homogeneity (greater than 70% identical text across pages), missing visual evidence, promotional language patterns, and zero competitor acknowledgment. Each mechanism independently prevents AI citation regardless of other signal strength.
4.5 Numeric Format Findings
Bare numeric inputs in AI queries are systematically captured by culturally prominent entities. Percentage symbols provide near-zero content-type signalling, with pop culture entities overriding mathematical interpretation. Year values follow a predictable four-stage temporal routing model. The 4.9 versus 5.0 star distinction triggers different processing pathways: credentialing versus brand-naming. These findings are explored in detail in SIGI-2026-084 through SIGI-2026-086.
5. Evidence Quality and Limitations
The catalog spans evidence levels from Level 4 (controlled experiments with clear differential responses) to Level 2 (single observational data points). Approximately 40% of findings are rated Level 3-4 (controlled observations or controlled experiments), 35% at Level 3 (systematic observations with multiple data points), and 25% at Level 2 (individual observations or descriptive findings).
All probes were conducted on a single AI model at a single time point (March 2026). Cross-model validation has not been performed. The commercial entity sponsoring this research operates in the service industry verticals studied, creating a potential conflict of interest. Findings should be treated as preliminary observations requiring independent replication.
6. Conclusions
The 155-observation catalog demonstrates that systematic probing of AI search pipelines can generate structured empirical findings across multiple domains of inquiry. The catalog reveals consistent patterns in how AI systems select, rank, and present entities, while also identifying areas of instability and unpredictability. Individual findings provide the foundation for targeted investigation, while the catalog as a whole maps the current frontier of empirical knowledge about AI recommendation behaviour.