Who Is Actually Accessing llms.txt?
Observation
We deployed llms.txt and llms-full.txt on seo.codequest.work. GA4 shows some access, but cannot distinguish between AI crawlers and humans. GA4 relies on JavaScript measurement, and most bots don't execute JS, placing them outside GA4's measurement scope entirely.
Anomaly
Large-scale studies from organizations with CDN/server log access paint a very different picture from the 'works for all AI search' expectation. Mintlify's 25-company CDN analysis shows nearly all llms.txt access comes from ChatGPT (OAI-SearchBot) alone, with llms-full.txt receiving roughly 5.6x more requests than llms.txt. Longato's 1,000-domain 30-day CDN log audit found zero llms.txt requests from PerplexityBot. Zero from ClaudeBot as well. The premise that llms.txt 'works across all AI search' collapses at the data level.
This exposes GA4's limitation. GA4 applies User-Agent-based bot filtering, making it unable to accurately capture AI crawler access. Verifying llms.txt effectiveness requires CDN log-level resolution. Relying on GA4's 'llms.txt has traffic' is a judgment built on a measurement blind spot.
Why We Still Include It as a Check Item
What We Observed on Our Own Site
After deploying llms.txt on seo.codequest.work, we confirmed actual access from ChatGPT. The same phenomenon reported in Mintlify's 25-company CDN analysis is occurring on our own site. Deploy llms.txt, and at minimum, ChatGPT will come read it. This is an observed fact.
Design Decision: 'Check It Before It's Proven'
SEO_CHECK includes llms.txt as a technical SEO check item — 1 point on Basic, up to 3 on Pro. Why include it despite no correlation in a 300K domain study? The reasoning is clear: if it could become an AI search evaluation factor depending on future developments, include it as a check item now. 'Wait until the effect is confirmed' means users are always behind. When deployment cost is near zero, getting users to prepare proactively is the more valuable approach.
What the Modest Score Conveys
Rich Results eligibility gets 20 points because its CTR improvement effect is proven. llms.txt gets 1-3 points because access is confirmed but impact on citations or rankings isn't. The weight difference reflects confidence levels. If AI vendors officially declare support or large-scale studies confirm effectiveness, we'll revise the score. Including it proactively as a check item and weighting it according to current evidence are not contradictory.
The Pro plan's 3-point breakdown: 1 for presence, +1 for title and links completeness, +1 for .well-known path support. We evaluate implementation quality, not just presence. If you're going to prepare, prepare properly — a design principle shared across all SEO_CHECK items.
Is There Any Trace of AI Citations?
Discovery
GSC provides no means to isolate AI citation traffic as of April 2026. Google AI Overviews impressions are visible, but traffic from ChatGPT or Perplexity citations falls outside GSC's measurement scope. Measuring llms.txt effectiveness through GSC faces a fundamental architectural impossibility.
External Research Data
SE Ranking's 300K domain study found no statistical correlation between llms.txt presence and AI citation frequency. ALLMO's analysis of 94,614 cited URLs found only 1 llms.txt page cited (0.001%). OtterlyAI's 90-day experiment showed just 84 llms.txt accesses out of 62,100 AI bot requests (0.1%). No data currently supports the claim 'deploying llms.txt gets you cited in AI search.'
GA4 can't identify bots, GSC can't isolate AI citations, and without CDN logs there's no way to accurately measure llms.txt effectiveness. Discussing the ROI of an unmeasurable tactic has inherent limitations.
Hypothesis 1: llms.txt Value Is Developer Tool Integration, Not GEO
Observation
The systems actually using llms.txt daily are coding assistants like Claude Code, Cursor, and Windsurf. Developers explicitly specify URLs to load documentation on command. This is fundamentally different from AI search engines autonomously fetching llms.txt.
Hypothesis
The practical value of llms.txt lies in 'document structuring for developer tools,' not 'AI search optimization.' While ROI as a GEO tactic remains unsupported by data, it functions as a documentation delivery mechanism within the developer ecosystem.
Verification
In dev5310's experiment, PerplexityBot's llms.txt requests were also zero. Perplexity crawls content directly without referencing llms.txt. Same for ClaudeBot. Meanwhile, Claude Code and Cursor will read it when instructed by users. 'Autonomously fetched' and 'loaded on command' are fundamentally different use cases.
Hypothesis 2: Why ChatGPT Heavily Accesses llms-full.txt
Observation
Mintlify's CDN data shows llms-full.txt receives roughly 5.6x more access than llms.txt. ChatGPT accounts for the vast majority of llms-full.txt traffic. Access from other AI systems is negligible.
Hypothesis
LLMs tend to prefer full-text ingestion over RAG-style partial retrieval. Choosing the full text (llms-full.txt) over the index (llms.txt) aligns with this tendency. However, 'accessing' and 'using in responses' are separate issues. Whether access translates to citations remains unverified, and SE Ranking's 300K domain study found no correlation.
Analysis
The fact that only ChatGPT reads llms.txt suggests OpenAI may be utilizing this specification in some form. However, no data proves 'reading = reflected in citations.' The no-correlation result across 300K domains suggests that even if ChatGPT reads llms.txt, it may not contribute to response quality or citation probability.
Hypothesis 3: What 'Zero Official Support' Really Means
Observation
Zero AI vendors have officially declared support for llms.txt. OpenAI, Anthropic, Google, Meta — none have issued formal support statements. Google's John Mueller explicitly stated 'No AI system currently uses llms.txt.'
Contradiction
Yet Google's own documentation sites have llms.txt deployed. Mueller's comment contradicts Google's implementation. This contradiction could reflect 'future-proofing' or 'internal CMS auto-generation,' but either way, the state is 'no official support, yet the file is deployed.'
Hypothesis
AI vendors may be avoiding the legal and normative risks of declaring 'official support.' Declaring support would invite arguments like 'if you read llms.txt, you should also respect robots.txt disallow.' By separating de facto reading from official support, they may be preserving flexibility in crawl scope.
Decision Criteria from This Analysis
The deployment cost of llms.txt is extremely low (one text file). It won't hurt, but its priority as a GEO tactic is low. Equating 'deployed' with 'AI search optimized' is an overestimation.
To measure effectiveness, track User-Agent-specific access in CDN logs. GA4 and GSC cannot capture this. Without CDN logs, accept that there's no way to quantitatively evaluate llms.txt impact.
Content quality, structured data, and E-E-A-T likely contribute more to AI citations than llms.txt. 'Improving communication accuracy with Google' also benefits AI citations — the same conclusion as our structured data analysis.
What GEO Really Is
GEO (Generative Engine Optimization) is a concept proposed by researchers from Princeton, Georgia Tech, and IIT Delhi in 2023, referring to content optimization for all AI-powered search engines — not just Google, but ChatGPT, Perplexity, Gemini, and beyond.
Yet what the data shows is that no silver bullet specific to GEO currently exists. llms.txt effectiveness is unproven. AI search engines behave differently from each other. No unified optimization method can be neatly packaged as 'GEO tactics.'
SEO_CHECK's Position
At this point, GEO is essentially good SEO. AI search engines crawl site content directly, not llms.txt. Content expertise, semantic clarity through structured data, E-E-A-T establishment — improving communication accuracy with Google simultaneously makes your content more citable by AI.
However, AI search is evolving rapidly. Today's 'no proven effect' could become tomorrow's 'essential practice.' That's why we believe the right approach is to do SEO with GEO in mind. This is exactly why SEO_CHECK includes llms.txt as a check item even before its effectiveness is proven — prepare now for what might change tomorrow, especially when the implementation cost is near zero.