# Research before conclusions.

> Reliable AI recommendation monitoring requires definitions, observations and calibration that remain useful as models, products and retrieval conditions change.

- Canonical page: https://winin.ai/en/research/
- Language: English
- Last updated: 2026-09-04

## Current research program

*Protocol 01*

### [AI Readiness Benchmark 2026](/en/benchmarks/ai-readiness-2026/)

Sampling frame, metric strata, calibration plan and participation criteria for an interpretable industry benchmark.

*Method 01*

### [Measurement Method V1](/en/methodology/)

Evidence records, repeated sampling, human review, AI evaluation and regression tests.

*Field note 01*

### [GEO growth system](/en/geo/)

A durable operating connection between crawlability, evidence, representation and agent action.

*Technical note 01*

### [Machine-readable content for AI agents](/en/research/agent-readable-content/)

How llms.txt, Markdown counterparts and page-level links create a clearer route to approved public facts.

## Questions the research must answer

- Which capabilities predict accurate citation across multiple AI systems?
- How much observed movement is intervention effect versus normal model variance?
- Which dimensions transfer across industries and which require sector norms?
- When do AI judges agree with trained human reviewers — and when do they fail?
- How should agent task completion change the readiness model?
- What evidence would falsify the claim that AI Readiness is an independent category?

## Publication standard

Every research release will name its sample, exclusions, method version, observation window, missing data and conflicts of interest. Raw or minimally transformed evidence will be published when consent, confidentiality and platform terms allow it.

## Source

This Markdown document is the machine-readable counterpart of https://winin.ai/en/research/. The canonical HTML page remains the public presentation source.