Not sure where to start? Take a guided tour.
AI Site Grader
AI Query Scanner
llms.txt GeneratorWhat is the best way to read the null result?
Read p=0.85 as a non-significant test in the top 5,000 cited domains. The full sample returns p<0.001 with reported r=-0.065. Neither establishes that the causal effect is exactly zero.
Different cohorts, different results.
- Top 5,000 cited domains
- Subset of the top 5,000 cited domains; not the full-sample test.
- p = 0.85
- Full sample
- A small reported association (r = −0.065). Full 37,894-domain sample. The public JSON rounds this value to zero.
- p < 0.001
Mann-Whitney U tests · 14 March 2026.
Keep the two cohorts separate and distinguish statistical significance from causal evidence.
The evidence.
Source
The llms.txt Effect
| Measure | Result |
|---|---|
| Top-5,000 Mann-Whitney p-value Subset of the top 5,000 cited domains; not the full-sample test. | 0.85 |
| Full-sample Mann-Whitney p-value Full 37,894-domain sample. The public JSON rounds this value to zero. | <0.001 |
| Full-sample reported effect size r Small reported magnitude; not an estimate of causal citation lift. | -0.065 |
| Adoption rate Domains with llms.txt in the study. | 13.3% |
| Domains scanned AI-cited domains scanned for llms.txt. | 37,894 |
Population and method
Built from HTTP scans of 37,894 AI-cited domains, linked to 337,362 citations and 882 citation snapshots in the Trakkr corpus.
Where the evidence stops
This is an observational study. It measures correlation with citation outcomes, not a controlled experiment.
Adoption is uneven by sector, so raw averages can hide category concentration in SaaS and developer tooling.
The top-5,000 test returns p=0.85. The full scanned sample returns p<0.001 with reported r=-0.065. Neither establishes a causal effect. The public aggregates do not include confidence intervals or document the cohort for citation means and medians well enough to recompute the tests.
What to do with this
- Treat llms.txt as an optional housekeeping file, not a primary citation-growth lever.
- Prioritize answer quality, source coverage, and page structure before spending disproportionate effort on llms.txt.
- If you do publish llms.txt, measure discovery and crawl behavior directly instead of assuming it improved citation performance.
This answer matters because it turns a study finding into an operating rule teams can use when they decide what to publish, refresh, or measure next.
Questions about this answer.
Cite this answer.
Source study: llms.txt Effect“Read p=0.85 as a non-significant test in the top 5,000 cited domains. The full sample returns p<0.001 with reported r=-0.065. Neither establishes that the causal effect is exactly zero.”
Mack Grenfell · What is the best way to read the null result?
The llms.txt Effect · Trakkr Research · Updated 4 September 2026