Publishing information doesn’t mean AI receives it.
We followed business information through OpenAI’s published browsing-tool code and real retrieval services.
Research notes
Resource collection
Research and evaluation
Design reproducible tests and report retrieval, citations and answer quality without overstating the evidence.
About this collection
A useful test starts with a question it can actually answer. These methods define controls, retained observations, scoring rubrics and limits before a team interprets the result. Research designs in this collection are proposals, not announcements of experiments that have already run.
Keep source delivery, verified retrieval, answer content and citations in separate records. Report inconclusive findings and explain what remains unknown. For accepted historical observations and their actual scope, visit the evidence and methodology pages; a proposed study cannot substitute for those records.
Research & evaluation
1–7 of 7 resources
How to Design a Reproducible AI Website Retrieval Experiment
Before testing whether AI understands a website, establish exactly what the experiment can observe. A reproducible study begins with controlled sources, fixed questions and retained evidence.
Research design
An Evaluation Rubric for AI Answers About a Business
An AI answer can sound persuasive while omitting the one condition that matters. A useful evaluation must score what is correct, what is missing and what should never have been asserted.
Evaluation guide
Does Markdown Improve AI Answer Quality? A Controlled Test Plan
Markdown may simplify website content for some machine-processing workflows. Whether that improves AI answers is a separate empirical question requiring a controlled experiment.
Research design
How to Document AI Citations Without Overclaiming the Results
A URL appearing in an AI answer is an observation. Whether that URL supplied the relevant fact—or caused the answer to change—is a different question.
Evaluation guide
When an AI Website Experiment Shows No Improvement: What to Publish
An experiment that fails to show improvement can still teach something useful. The important part is reporting what was tested, what happened and what remains unresolved.
Research reporting
The AI-Readable Web Research Agenda: 10 Questions We Still Need to Answer
The AI-readable web is developing faster than many of our assumptions can be tested. Instead of treating plausible mechanisms as proven outcomes, we should identify the questions that still need evidence.
Research design