LLMS.TXT: THE AI SEO TACTIC THAT DOESN'T WORK (WE CHECKED THE DATA)

Does llms.txt improve AI search visibility? No. There is no measurable evidence that any major AI engine reads llms.txt in production, and no measurable visibility lift from deploying one. Monitoring of more than 500 million AI bot visits found exactly 408 llms.txt fetches. Google confirmed in July 2025 that it will not support the standard. SE Ranking's citation-prediction model, trained on 300,000 domains, got more accurate when llms.txt was removed as a variable. In one controlled rollout, 8 of 9 sites saw zero traffic change after implementing it. If an agency is pitching llms.txt as the key to showing up in ChatGPT or Perplexity, they are selling you a file that nothing reads. What does move AI citations is restructuring your content around signals engines measurably reward, which is what we test in every free scorecard.

A note on this URL: the original version of this post, published April 6, 2026, was a how-to guide that claimed major AI platforms read llms.txt. The evidence says otherwise, and our own methodology page already said so. We rewrote the post rather than quietly deleting it. Here is the data.

WHAT LLMS.TXT WAS SUPPOSED TO BE

The llms.txt proposal dates to September 2024: a plain markdown file at the root of your domain that gives language models a curated, noise-free map of your site and your company. The idea is reasonable. HTML is cluttered, context windows are finite, and a clean briefing file would make machine consumption easier.

The problem is not the idea. The problem is the consumption side. A discovery standard only matters if the engines actually fetch and use it. Nearly 2 years in, they don't.

THE NUMBERS

408 fetches out of 500 million visits. Monitoring of more than 500 million AI bot visits across production websites recorded just 408 requests for llms.txt. That is roughly 1 fetch per 1.2 million bot requests, or about 0.00008%. For comparison, robots.txt gets checked on effectively every crawl session. If GPTBot, ClaudeBot, or PerplexityBot treated llms.txt as a real input, it would show up in server logs at scale. It doesn't.

Google said no, on the record. Gary Illyes confirmed in July 2025 that Google will not support llms.txt. John Mueller, Google's Search Advocate, went further and compared the file to the keywords meta tag, the field search engines abandoned because site owners filled it with whatever they wanted to rank for. As Mueller put it:

"I don't see llms.txt being used by any of the AI services."

The prediction model got better without it. SE Ranking trained a model on 300,000 domains to predict AI citation frequency. When llms.txt presence was removed as a variable, the model's accuracy improved. That is the statistical signature of noise. A real ranking signal makes a model more accurate, not less.

8 of 9 sites saw no traffic change. In a controlled implementation test, 8 of the 9 sites that deployed llms.txt saw no change in AI-referred traffic afterward. One data point of movement out of nine is indistinguishable from chance.

Adoption sits near 10% and is not growing. Standards that work get adopted fast because the payoff is visible. llms.txt adoption has plateaued at roughly 10% of tracked sites, and no major AI company, not OpenAI, not Anthropic, not Perplexity, not Google, has committed to reading it in production.

Stack those 5 findings together and the conclusion is not subtle. llms.txt is not a visibility lever. It is a checkbox.

WHY WE STILL SHIP ONE

Here is the honest nuance, because the right answer is not "delete your llms.txt."

• It costs nothing. Writing one takes about 30 minutes, carries zero risk, and needs almost no maintenance. Free hygiene is still hygiene.
• It is useful for agents, not engines. AI coding agents and agent-to-business interfaces can fetch one clean file instead of parsing your HTML. When an agent is instructed to research your company directly, llms.txt is a genuine convenience.
• We run one ourselves. Anvil serves its own llms.txt as part of our agent-facing stack, alongside our MCP endpoint. We treat it as plumbing for agents, not as a citation strategy.
• We weight it honestly. In our scoring methodology, llms.txt carries only minimal weight within the AI Search channel. Present, because it is a free win. Minimal, because the evidence says minimal.

The line we hold: recommend it as a free action, never sell it as a driver. Any agency invoicing you for "llms.txt optimization" as a visibility play is charging for theater.

WHAT ACTUALLY DRIVES AI CITATIONS

The same research that buries llms.txt points clearly at what works. At the category level:

• Statistics in content. Embedding concrete data points is among the highest-impact visibility levers in Princeton GEO-bench testing.
• Attributed expert quotations. Named, credentialed quotes are another of the strongest levers for lifting visibility.
• Extractable answers. Self-contained passages that fully answer one question are markedly more likely to be cited.
• Off-site consensus. Citation overlap between major AI platforms is only about 11%, so presence across multiple third-party authorities beats any single placement, including anything on your own domain.

How those translate into specific actions for your site depends on your category, your competitors, and your current Share of Model. That is exactly what the scorecard measures.