llms.txt: Does It Actually Work for GEO? (Evidence-Based, 2026 Update)

llms.txt does not work for GEO. Google's May 2026 AI Search Optimization Guide (John Mueller) explicitly confirmed that llms.txt carries no special weight for AI citation selection. A SERanking study of 300,000 domains found no measurable correlation between llms.txt presence and AI citation rates. A 48-day server log study (wislr.com, Feb–March 2026) recorded 12,099 AI bot requests and zero requests to /llms.txt. The accumulating evidence is unequivocal: llms.txt does not improve AI search visibility.

The scale of what you are missing is now measurable. Google AI Overviews serve an estimated ~13 billion impressions every month globally (Nico Digital, 2026), and AI search across the major engines processes 3.5B+ queries every week (Axis Intelligence, June 2026). Within that discovery surface, third-party publishers earn 6.5× more AI citations than owned domains (Axis Intelligence, 2026) — earned media, not a self-described summary file, is what wins visibility. A file you write about yourself is the weakest possible citation signal.

This guide walks through what llms.txt was supposed to do, what every major study says in 2026, Google's official position (multiple times on record), and where to invest your optimization effort instead. The BlueJar three-layer technical GEO stack framework (Feb 2026) provides a useful lens: llms.txt sits in the "context layer" — the layer with the least proven citation impact — while robots.txt (crawl layer) and content quality (entity layer) carry the real weight.

What is llms.txt

llms.txt is a proposed standard for a Markdown-formatted file hosted at /llms.txt on the root of a domain. The intent was to give AI models a curated, machine-readable summary of a site's content — the way robots.txt tells crawlers what they may access, llms.txt would tell models what the site is about.

The proposal gained traction in mid-2024 among SEO practitioners searching for a GEO equivalent of robots.txt. Several major sites published llms.txt files, and tooling emerged to generate them automatically. The hypothesis: AI engines would fetch llms.txt and use it to improve retrieval accuracy. By 2026, the evidence against this hypothesis is decisive.

Google's official position (multiple statements)

Google has stated its position on llms.txt multiple times across two years:

Source: Google Search Central — John Mueller AI Search Guide (May 2026). Gary Illyes at Search Central Live Deep Dive (July 2025). John Mueller social post (June 2025).

"llms.txt is a content description helper, not a citation signal. AI does not give it special weight. The best way to get cited by AI is to write content that provides genuine information gain."

The SERanking study: 300,000 domains

SERanking ran the largest empirical test of llms.txt effectiveness in 2025. They analyzed 300,000 domains, comparing AI citation rates between sites that published llms.txt and sites that did not, controlling for content quality, domain authority, and topic.

"We found no measurable correlation between llms.txt presence and AI citation rates across the 300,000 domains studied. The signal was indistinguishable from noise. In fact, removing llms.txt from our XGBoost predictive model improved accuracy — meaning it added noise, not signal."

Only 10.13% of domains had an llms.txt file. Adoption was flat across traffic tiers — 9.88% for small sites vs. 8.27% for sites with 100K+ monthly visits. High-authority sites actually adopted it less. The finding is consistent with how AI search engines actually work: models like ChatGPT Search, Perplexity, and Google AI Overviews retrieve and cite from their indexed corpus, not from a site-provided summary file.

The wislr.com server log study (2026)

The most recent empirical evidence comes from a 48-day server log study (February–March 2026) by wislr.com. They recorded 12,099 AI bot requests from multiple crawlers across their server logs and analyzed which files those crawlers requested.

Source: wislr.com server log study (Feb–March 2026), cited in BlueJar technical GEO analysis.

Caveat: Some site operators (e.g., SEO Ray Martinez) report OpenAI polling their /llms.txt file every ~15 minutes, suggesting that crawling behavior varies by site and crawler. However, even when /llms.txt is fetched, no study has demonstrated a citation lift from its presence. The wislr.com data shows that for this site, across 48 days and 12,099 bot requests, no AI crawler attempted to fetch the file.

Which AI engines support llms.txt

Sources: Google Search Central (May 2026, June 2025). Gary Illyes at Search Central Live (July 2025). OpenAI, Perplexity, Anthropic documentation.

The three-layer technical GEO stack

BlueJar's 2026 analysis provides a useful framework for understanding where llms.txt fits — and does not fit — in a GEO strategy. Technical GEO has three non-interchangeable layers:

1. 1.
   Crawl layer (robots.txt) — Access control. Gates whether AI engines can fetch your pages at all. Highest impact. Block a retrieval bot and you are invisible regardless of content quality.
2. 2.
   Entity layer (Schema / structured data) — Labels who you are and what a page contains. Helps indexing and knowledge graph building. Medium impact. Does not drive direct citations but aids discovery.
3. 3.
   Context layer (llms.txt) — A curated summary of your site for language models. Lowest proven impact. No study has demonstrated measurable citation lift from its presence.

The BlueJar analysis concluded: "Get the crawl layer right first. Then invest in content quality and structured data. llms.txt is an optional signal with no proven citation effect — treat it accordingly."

Why llms.txt fails the GEO test

Three structural reasons explain why llms.txt cannot deliver the visibility lift its proponents claimed:

1. 1.
   AI engines cite from indexed content, not summaries

   Citation requires verifiable text the re-ranker can attribute. A self-authored summary at /llms.txt is not citable — AI engines would be quoting the site about itself. Google's May 2026 guide confirmed this: citation selection is based on indexed content quality, not self-described summaries.

2. 2.
   Self-description is trivially gameable

   A site could claim anything in its llms.txt. AI engines cannot trust self-description as a ranking signal. They rely on independent retrieval from indexed content. Gary Illyes compared it to the keywords meta tag — a signal that was abandoned precisely because it was gameable.

3. 3.
   No engine has announced consumption as a citation signal

   Google (×3 statements), OpenAI, Perplexity, and Microsoft have not announced that they consume llms.txt as a retrieval signal. Even Anthropic, which published its own llms.txt, has not stated that Claude reads external llms.txt files for citation selection.

Where to invest instead (2026 update)

The Princeton GEO study quantified the lift from strategies that actually work. Google's May 2026 guide adds one more: Information Gain — content that provides original value, data, or insights not already available on the web.

The window for cheap tricks is closed. AI referral traffic grew 357% year-over-year in 2025 (theworlddata.com), and AI Overviews now appear on up to 48% of queries (Digital Applied, March 2026). The click cost is now measurable too: AI Overviews cut position-1 organic click-through rate by 58% as of December 2025 (Ahrefs, Feb 2026) — brands that get cited win the remaining clicks; those that don't get none. With that much discovery happening inside AI answers, the only durable lever is content that earns citation on its own merits — not a self-described summary file that no engine reads.

Sources: Authoritas — AI Overview citation-factor study (2026). Aggarwal et al., KDD 2024 (Princeton GEO-bench). BrightEdge — AI Overview citation rates (2026). Google AI Search Guide (May 2026). SERanking 2025 llms.txt study. wislr.com server log study (Feb–March 2026). BlueJar technical GEO stack analysis (Feb 2026).

If you still want to publish llms.txt

llms.txt is not harmful — it simply does not help measurably. If you have implemented every higher-impact strategy and have spare time, publishing an llms.txt costs minimal effort. The risk is opportunity cost: time spent on llms.txt is time not spent on strategies with proven lift, or — more importantly in 2026 — on creating content with genuine Information Gain.

If you publish one, keep it accurate, concise, and aligned with your visible site content. Do not use llms.txt to make claims that do not appear in your indexed content — at best it is ignored, at worst it could be flagged as inconsistent if AI engines ever do consume it. Google's May 2026 guide categorizes it as a "content description helper" — treat it as such.

Frequently asked questions