This repository contains data, scripts, and findings from the Agent-Adoption Correlation Study — Respectarium's empirical investigation of how website agent-readiness practices correlate with LLM (large language model) visibility outcomes.
Each tagged release is a frozen study artifact. Future studies (Q2, Q3, …) ship as new tags in this same repo.
Current release: study-2026-04 — first cross-sectional study (2026-04-26).
Live preprint: respectarium.com/research/correlation-2026-04
Plain-language interpretation: respectarium.com/research/correlation-2026-04/interpretation
Companion specification: respectarium.com/spec/agent-adoption/v1 · github.com/respectarium/agent-adoption-spec
The relationship between agent-readiness signals (measured by three independent scanners — Respectarium, Cloudflare, and Fern) and LLM visibility outcomes (Respectarium's leaderboard rank across Claude, GPT, and Gemini), on a sample of 908 brands across 50 B2B SaaS categories.
The three scanners produced 72 agent-readiness predictors in total (66 individual scanner checks plus 5 aggregate metrics and 1 derived signal). After variance and pre-registered data-quality filters excluded 22 predictors, 50 were evaluated against five LLM-visibility outcomes.
Pre-registered analytical thresholds were committed before any results were viewed (see methodology.md).
- 2 of 50 signals graduate to PROMOTE_SCORED status under pre-registered thresholds:
cloudflare.level— Cloudflare's aggregate readiness levelrespectarium.markdown-negotiation—Accept: text/markdowncontent-negotiation support
- Effect sizes are small to medium (Cohen's d ≤ 0.65). No "large" effects (d > 0.8) anywhere in the study.
- Claude vs GPT structurally reverse on 4 signals — the same agent-readiness check predicts higher Claude listing AND lower GPT listing (or vice versa), at FDR-significant levels in both directions.
- 20 of 66 per-check signals have <5% adoption in the brand universe — the agent-readiness specification ecosystem is in its infancy as of 2026.
- Cross-scanner divergence is real: same-named checks across different scanners can correlate at ρ ≈ 0.03 (essentially uncorrelated). Implementations differ enough that "redirect-behavior" measured by one scanner is fundamentally not the same as "redirect-behavior" measured by another.
Full ranked findings: results/SUMMARY.md.
git clone https://github.com/respectarium/agent-adoption-research
cd agent-adoption-research
git checkout study-2026-04
cd scripts
npm install
npm run all
# Outputs land in ../results/, identical to the committed canonical run.See REPRODUCIBILITY.md for the complete reproducibility protocol.
Agent-Adoption Correlation Study (study-2026-04). Respectarium, 2026-04-26. Available at: https://respectarium.com/research/correlation-2026-04 Source + data: https://github.com/respectarium/agent-adoption-research/releases/tag/study-2026-04
Open an Issue or Discussion. Replication attempts, methodological critique, and proposed analytical improvements are all welcomed.
agent-adoption-research/
├── README.md ← you are here
├── REPRODUCIBILITY.md ← step-by-step re-run protocol
├── methodology.md ← pre-registered analytical thresholds + COI disclosure
├── CONTRIBUTING.md ← how to propose changes
├── CHANGELOG.md ← per-tag release notes
├── LICENSE.md ← CC-BY 4.0 (data + content) + MIT (scripts)
├── data/
│ ├── merged.json ← canonical merged dataset (908 rows)
│ ├── merged.csv ← flat CSV view of the same data
│ └── data-dictionary.md ← column definitions + encoding rules
├── scripts/
│ ├── 00-data-quality.ts ← coverage, variance filter, bot-protection census
│ ├── 01-univariate.ts ← Spearman + Pearson + FDR + Bonferroni
│ ├── 02-per-llm.ts ← drop-null + binary recode per LLM
│ ├── 03-per-category.ts ← per-eligible-category breakdown
│ ├── 04-multivariate.ts ← OLS with category fixed effects
│ ├── 05-bot-protection.ts ← bot_protected as standalone + confounder
│ ├── 06-redundancy.ts ← pairwise correlations + connected-components clustering
│ ├── 07-scanner-h2h.ts ← scanner head-to-head metrics
│ ├── 08-group-comparison.ts ← Welch's t-test on pass/fail groups
│ ├── 09-distributions.ts ← outcome shape + per-category means
│ ├── 10-verdicts.ts ← consolidated verdicts + signals.csv
│ ├── lib/ ← typed helpers (predictors, outcomes, inference, regression, util, types)
│ ├── package.json ← stats deps only (jstat, ml-matrix)
│ └── tsconfig.json
├── results/ ← canonical outputs from running scripts/ on data/merged.json
│ ├── 00-data-quality.{json,md}
│ ├── … (all 11 scripts' outputs)
│ ├── SUMMARY.md
│ └── signals.csv ← publication-ready ranked signal table
├── governance/
│ └── maintainers.md
├── .github/ISSUE_TEMPLATE/
└── .gitignore
This release supports two distinct reproducibility levels:
Analytical reproducibility — fully supported. Given data/merged.json and scripts/, you can re-run all 11 stats scripts and produce numerically identical outputs to the committed results/ (the analyses are deterministic — no randomness).
Pipeline reproducibility — partial by design. The data-collection layer (running scanners against ~908 brands) is not published in this repo. The Respectarium scanner is closed-source; the specification it implements is published openly at respectarium.com/spec/agent-adoption/v1. Independent implementations against the same spec are encouraged. Cross-scanner divergence (documented as a primary finding in this study) means independent results should be comparable, not numerically identical.
For details, see REPRODUCIBILITY.md.
Respectarium operates one of the three scanners evaluated in this study. To mitigate:
- Pre-registered analytical thresholds (see
methodology.md) - Findings unfavorable to the Respectarium scanner are reported transparently. In study-2026-04, the Respectarium scanner's
scoreaggregate has zero predictive power for LLM-visibility outcomes (ρ ≈ 0.016, FDR-adjusted p = 0.69) — this finding is published rather than concealed. - Per-scanner results reported throughout, allowing readers to examine each scanner's signals independently.
- Cross-scanner divergence on shared check names (ρ ≈ 0.03 on three pairs of same-named checks) is documented as a primary finding rather than a footnote.
Full disclosure: methodology.md.
- Content + data: CC-BY 4.0. Free use, modification, redistribution with attribution.
- Scripts: MIT. Free use including in commercial applications.
Quarterly re-runs publish as new tags in this repo (study-2026-07, study-2026-10, etc). The main branch tracks the latest study; past studies remain accessible via their tags forever.