LLMS.txt Does Not Improve AI Visibility. 300,000 Domains Make That Clear.

The SEO community spent the last year implementing llms.txt files with the conviction that they'd help AI models understand and cite their content more often. SE Ranking just analyzed 300,000 domains and found that the file has zero measurable impact on AI citations. Not a small impact. Not "it depends." Zero.

This is going to sting for anyone who prioritized it.

The study, and why it matters more than the usual correlation piece

SE Ranking's research team used XGBoost regression, Spearman correlation, and SHAP analysis across 300,000 domains to determine whether having an llms.txt file made a site more likely to be cited by AI models. The methodology here is more rigorous than most SEO research. Machine learning plus statistical analysis, not just a simple "sites with X get more traffic" correlation that usually passes for evidence in this space.

The core finding: removing llms.txt from their predictive model actually improved the model's accuracy. The file wasn't helping the model understand citation behavior. It was adding noise. That's not a neutral result. It's an actively negative one.

Adoption was also lower than the hype would suggest. Only 10.13% of the 300,000 domains had implemented llms.txt at all. And the adoption rate was basically identical across traffic tiers: low-traffic sites (0–100 visits) sat at 9.88% adoption, while high-traffic sites (100,000+ visits) came in at 8.27%. If llms.txt were actually working, you'd expect the sophisticated sites with dedicated SEO teams to be adopting it at meaningfully higher rates. They aren't. In some accounts I've worked with, the SEO team pushed hard to prioritize llms.txt implementation over other technical debt, and in retrospect that was wasted sprint capacity.

Why the hype outran the evidence

I think this happened because the concept was elegant. Robots.txt tells crawlers how to behave. Llms.txt tells AI models what your site is about. It's a clean analogy, and clean analogies are dangerous because they make unproven ideas feel like established best practices.

The reality is messier. AI models don't consume llms.txt files the way crawlers consume robots.txt. There's no confirmed mechanism by which ChatGPT, Gemini, or Perplexity read an llms.txt file and use it to decide whether to cite a domain. The proposal came from community discussion and some platform acknowledgments, but "we support reading this file" is very different from "this file influences our citation decisions."

From what I've seen, the teams that went all-in on llms.txt were often the same teams that chased AMP pages, structured data for every possible schema type, and every other technical SEO trend that promised outsized returns. Not because those people are bad at their jobs. The pattern is that technical SEO has a bias toward implementation-heavy solutions because they feel productive. Writing an llms.txt file is a concrete deliverable you can put in a Jira ticket. Improving content depth across 200 pages is not.

What actually correlates with AI citations

We covered the emerging research on AI citation signals a few weeks ago, and the findings from that analysis align with SE Ranking's conclusion here. The signals that correlate with AI model citations look nothing like traditional ranking factors.

What the data points toward: content depth, topical authority (not domain authority — topical), freshness of information, and clear factual claims with sourcing. These are harder to game and harder to package into a technical implementation checklist, which is probably why they get less attention in SEO communities on Reddit and elsewhere.

The uncomfortable part of this finding is that it suggests AI optimization, at least in the form of adding files and metadata, doesn't work. What works is the same advice that has worked for a decade: write substantive content that a knowledgeable human would find useful. The AI models are, in this case, surprisingly good at identifying it on their own without a special file telling them where to look.

The real cost isn't the file itself

An llms.txt file takes maybe twenty minutes to create. That's not what concerns me.

I'm concerned about the opportunity cost in organizations where llms.txt became a priority item. I've seen SEO roadmaps where "implement llms.txt" sat above content audits, page speed improvements, and internal linking restructuring. In one account we advise (B2B SaaS, around 40,000 organic sessions per month), the team spent two weeks building an automated llms.txt generation pipeline that pulled from their CMS. Impressive engineering. Two developers, a product manager, and a sprint's worth of QA. And according to this data, completely useless for its intended purpose.

That same two weeks could have been spent on the content depth improvements that actually correlate with AI citations. Honestly, that math doesn't feel great when you're the one who recommended the work.

And this is the thing that bothers me most about the SEO hype cycle in general. We keep optimizing for signals before we've confirmed they're signals. The industry did it with AMP, did it with passage indexing structured data, and now we've done it with llms.txt. The cost each time isn't the implementation itself. It's everything else that didn't get built.

A narrower version might still earn its place

I don't want to be completely dismissive — the concept of machine-readable site descriptions isn't inherently bad. If major AI platforms formally adopt llms.txt as a standard input to their retrieval systems, with documented impact on citation behavior, the math changes entirely. Bing's webmaster team has hinted at interest. OpenAI has acknowledged the proposal exists.

But right now, based on 300,000 domains of evidence, it isn't doing what people hoped. In SEO, the difference between "this might work eventually" and "this works now" is the difference between a roadmap item and a priority. I'd estimate that by Q4 2026, fewer than 3% of SEO consultancies will still list llms.txt as a priority recommendation for clients. Right now, it belongs on the backlog, not the sprint.

Close the file and open Search Console instead

If you've already built llms.txt, leave it. It isn't hurting anything. But if it's sitting on your Q2 roadmap as a priority, move it down.

Instead, pull up your Google Search Console data for the last 90 days. Filter by queries where your pages appear and AI Overviews are also present. Those are the pages where AI models are actively summarizing your topic right now. The question to ask about each one: is your content the most comprehensive, most current, and most clearly sourced piece on that subject? If it isn't, fixing that will do more for your AI visibility this quarter than any file you could add to your root directory.

The sites that show up in AI answers won't be the ones with the best technical implementation. They'll be the ones whose content is so thorough that the models can't summarize the topic without referencing it. LLMS.txt was a shortcut, and 300,000 domains of data says it didn't lead anywhere.