AI Search Research
Observing how AI-mediated search systems actually behave.
AI Search Lab · Experiments · Empirical Research
AI Search is developing faster than most stable measurement frameworks around it.
Many claims about visibility, retrieval and selection are therefore based on individual observations, vendor data or interpretations of systems we cannot fully inspect.
My approach is deliberately narrower.
I use small, controlled experiments to investigate specific questions about how AI-mediated search systems:
- introduce commercial categories
- retrieve and represent information
- construct Candidate Sets
- evaluate competing entities
- select brands and sources
- explain recommendations
- respond to changes in context and information
The objective is not to reverse-engineer proprietary systems.
It is to identify observable patterns, separate them from interpretation and turn them into better research questions.
The AI Search Lab
The AI Search Lab is the empirical layer of my work on Search Systems and AI Visibility.
The research follows a simple principle:
SMALL EXPERIMENT
↓
OBSERVATION
↓
INFERENCE
↓
HYPOTHESIS
↓
NEXT EXPERIMENT
Each Research Note addresses a deliberately limited question.
Observations document what happened.
Inferences describe what those observations may reasonably suggest.
Hypotheses identify what still needs to be tested.
This distinction matters.
AI systems are dynamic, partially opaque and highly dependent on context.
A repeatable observation can be useful without becoming a universal rule.
Research Notes
Research Note 002
The Mechanics of AI Generated Selection
Candidate Stability · Selection Evidence · AI Result Graph
Why can highly stable Candidate Sets still produce unstable winners?
Experiments 004 and 005 examine Candidate Set stability, search depth, evidence activation and the reconstructed weighting mechanisms behind AI-generated provider selection.
24 documented responses · 173 visible fan-outs · 1,696 raw source retrievals · 264 citations
Research Note 001
From Educational Questions to Brand Selection
Commercial Emergence · Candidate Sets · Brand Selection
When does an educational AI Search journey become commercial?
Three connected experiments examined how responses changed as a user moved from understanding a problem toward evaluating concrete providers.
Across 46 documented responses, the experiments observed a progression from commercial category emergence to concrete Candidate Sets and eventually justified Brand Selection.
The resulting working model:
Problem Relevance
↓
Category Association
↓
Candidate Set
↓
Selection Evidence
↓
Brand Selection
The experiments also suggest that AI Visibility should be measured at different decision layers rather than treated as one binary state.
What I Am Investigating
The research program focuses on questions that sit between traditional search measurement and increasingly complex AI-mediated decision systems.
Current areas include:
Retrieval
Which information becomes available for a particular task — and through which retrieval environment?
Representation
What version of a source or entity is actually available to the system?
Candidate Formation
Which entities, sources or information objects enter the active decision space?
How does retrieved information perform against competing alternatives?
Selection
When does an entity move from being available to being prioritized?
Which sources and evidence support generated claims and recommendations?
How does machine selection translate into exposure and potentially influence human decisions?
These questions connect directly to the conceptual models developed in the Concepts Library.
→ Explore the Concepts Library
Research Principles
This work is intentionally conservative in what it claims.
Observation Before Explanation
Document the visible behavior before trying to explain the mechanism behind it.
Replication Before Generalization
Repeated observations are more useful than isolated screenshots.
They still do not automatically establish universal behavior.
Mechanism Before Tactic
The objective is to understand where a visibility outcome may originate before proposing an optimization tactic.
Failure States Matter
Not being selected can result from different problems:
representation,
retrieval,
evaluation,
selection
or grounding.
These should not be collapsed into one visibility metric.
Limitations Stay Visible
Model behavior, retrieval infrastructure and search results can change.
Every Research Note therefore documents its experimental context, sample and limitations.
From Research to Search Architecture
The Research Notes are not intended to exist in isolation.
They connect three layers of my work:
CONCEPTS
How can Search Systems and AI Visibility be understood?
↓
RESEARCH
What can we actually observe under controlled conditions?
↓
CASE STUDIES
What happens when these principles meet real websites, organizations and technical systems?
Research sits between theory and application.
It tests assumptions from the conceptual framework and generates questions that can later be examined in real search environments.
Ongoing Research
Research Note 001 established an initial question:
When does commercial relevance become Brand Selection?
The next experiments will move further into the decision architecture.
Questions include:
- Which Selection Criteria remain stable across repeated runs?
- Which sources support Candidate-Set and Selection statements?
- How stable are Candidate Sets over time?
- How do target-group and business-model characteristics change selection?
- Can controlled changes to source information alter Candidate-Set or Selection outcomes?
- How do these patterns differ across markets and AI systems?
The objective is not to produce a growing list of AI ranking factors.
It is to progressively improve the model of how information becomes available, evaluated and selected.
Research, Not Certainty
AI Search research has an unusual constraint.
We can observe outputs.
We can control parts of the input.
We can sometimes observe retrieval behavior.
But we cannot treat every internal mechanism as known.
That makes epistemic discipline important.
A useful experiment does not need to explain everything.
It needs to make the next uncertainty smaller.
Follow the Research
New Research Notes will be published as individual experiments produce findings worth documenting.
If you're working on AI Visibility, Search Systems or the transition from traditional search toward AI-mediated discovery, the Concepts Library provides the theoretical framework behind this work.
For applied examples: