Three independent studies tracking 515M+ AI bot requests found llms.txt read rates near zero. Google, OpenAI, and Anthropic have all spoken. Here is what actually drives AI citations in 2026.
When Jeremy Howard proposed llms.txt in September 2024, the idea was elegant: a structured Markdown file at the root of your site that tells AI models what content matters. Twenty months later, three independent studies have reached the same conclusion: AI crawlers do not read it.
| Study | Sample | Finding |
|---|---|---|
| Limy.ai (May 2026) | 515M AI bot requests, 90-day window | 408 requests hit /llms.txt — statistically negligible |
| SE Ranking (Nov 2025) | 300,000 domains | No correlation between llms.txt and AI citation frequency; ML model improved when llms.txt removed as variable |
| OtterlyAI (2026) | 90-day crawler logs, 100+ sites | File requested in 0.1% of AI crawler visits |
| Semrush crawl test | Controlled, 3 months | Zero visits from GPTBot, ClaudeBot, or PerplexityBot |
The SE Ranking data is particularly damning. Their machine learning model for predicting AI citation frequency got noisier when llms.txt presence was included as a feature. The file is not just ineffective — it is an active distraction from the signals that actually predict citations.
| Platform | Position | Reads llms.txt? |
|---|---|---|
| Gary Illyes: "We do not support it and are not planning to." John Mueller compared it to the deprecated keywords meta tag | No | |
| OpenAI | Official crawler docs only acknowledge robots.txt. Complete silence on llms.txt | Occasional probes, no commitment |
| Anthropic | Most supportive — lists llms.txt in official docs as recommended | Some ClaudeBot reads, no citation commitment |
| Perplexity | Published own llms.txt + llms-full.txt; most explicit endorsement | Browse mode reads it |
There is a critical distinction: publishing your own llms.txt is a supply-side action. Whether your crawler reads other sites' llms.txt is a demand-side question. These are different things.
In May 2026, Shopify shipped llms.txt generation to millions of stores — without a single blog post or announcement. The SEO community interpreted this as validation. The reality is simpler: Shopify did it because it costs nothing and takes 15 minutes. It is a no-regret move, not a strategic bet.
One narrow, empirically validated use case: developer documentation for AI coding assistants. Cursor, Claude Code, GitHub Copilot, and Windsurf actively fetch llms.txt from documentation sites to generate accurate API integration code. Mintlify, Fern, GitBook, and Vercel Docs all default-generate it. For e-commerce stores, blogs, and content sites — zero evidence of citation impact.
| Signal | Evidence Strength | Cost |
|---|---|---|
| Complete Product/Article Schema (JSON-LD) | Strong — directly parsed by all major LLM crawlers | Medium |
| robots.txt AI crawler access | Strong — GPTBot, ClaudeBot, PerplexityBot all respect it | Low |
| Structured, citation-friendly content | Strong — clear claims, dated facts, concise answer blocks | High |
| Third-party review platforms | Moderate — G2, Capterra for SaaS; Reddit for DTC | High |
| llms.txt | None detected | Low |
Yoast's 2026 predictions team captured it best: "Structured data shifts from ranking enhancer to retrieval qualifier." Keep your llms.txt if you already have it — it costs nothing. But if you are allocating resources between schema markup, content quality, and llms.txt generation, spend on the first two.