Does an llms.txt file predict higher AI visibility?

Hypothesis

Sites that publish an llms.txt file achieve a materially higher SAVI (AI Visibility Index) than sites without one.

Method

Segment the 850k corpus by llms.txt presence; compute mean SAVI for each segment with a 95% confidence interval; report the between-group delta and whether it is significant. Pre-registered before the query is run - the result publishes regardless of direction (including null).

Dataset

850k+ audit corpus (metric: no_llms_txt - currently 69.7% have none).

Metrics

Result

CONFIRMED IN DIRECTION, EFFECT WEAK. Mean GEO 49.14 with llms.txt vs 43.78 without, delta +5.36 (95% CI +5.17 to +5.55), n=29,692 vs 73,182. 1.37 of those points are definitional (26% of the gap) because the check itself is 8 of 105 AI-citability points at a 0.18 weight. Prevalence is 91.0% among AI-Ready sites but already 22.0% among sites scoring under 50, so it discriminates poorly. Category gaps are flat (+2.1 to +7.7) with no standout. Our reading is that this does not meet the registration's word 'materially'.

Decision

ANSWERED. Correlational; a positive result motivates a controlled follow-up (does ADDING llms.txt raise SAVI). No causal claim from this alone. Answered 2026-08-19 on the SAVI index panel wave 1 (102,873 domains, scorer geo-145-2026-07, population-weighted).

Lessons learned

• A signal that costs nothing and requires no judgement stops carrying information about judgement. llms.txt is increasingly emitted by templates.
• Detection is presence only, not contents or quality, and the score measures PUBLICATION not consumption. Whether any engine reads the file is untested and is the better question.
• Publish alongside EXP-0006: the contrast between a discriminating marker and a diluted one is the finding, and neither result is as strong on its own.

Confidence

HIGH