Does llms.txt Help AI Citations? The Honest 2026 Answer

llms.txt is the most hyped AI-visibility file of 2026. It has real search demand, its own Lighthouse audit check, and a steady stream of blog posts calling it essential for getting cited by AI. It also has a problem: the one search engine that has said anything official about it has said, in plain language, that it does nothing. If you are spending time on llms.txt before the signals that actually drive citations, you are optimizing the wrong file. For the full picture of what does make a page citable, see What Is AI Citability, the pillar post for this cluster.

This post is the honest version: what llms.txt is, what the evidence actually supports, and where it belongs on your priority list (near the bottom).

What llms.txt Actually Is#

llms.txt is a proposed plain-text Markdown file at the root of your domain, yoursite.com/llms.txt, that lists your most important pages with short descriptions. The idea is to hand a large language model a curated map of your site so it does not have to infer structure from your navigation. A companion file, llms-full.txt, inlines the entire text of those pages into one document.

It is a community proposal, not a standard any major AI vendor has committed to reading. That distinction is the whole story. A convention only works if the consumers on the other end agree to honor it, and so far the largest consumer has publicly declined.

It is also routinely confused with robots.txt, which is a different file that does a different and far more important job: controlling whether AI crawlers are allowed to read your content at all. Getting robots.txt right is load-bearing. Getting llms.txt right is, at best, a courtesy.

What Google Actually Said#

On June 15, 2026, Google updated its guidance on AI features and websites. The relevant line is unambiguous: you do not need to create machine-readable files, AI text files, markup, or Markdown to appear in Google Search, including its generative AI capabilities, because Google Search itself does not use them.

Google's John Mueller went further, comparing llms.txt to the old keywords meta tag: a self-declared signal that Google abandoned years ago precisely because anyone could stuff it with whatever they wanted, so it carried no trust. An index file that a site writes about itself has the same structural weakness. The engine has no reason to trust your self-description over what it can read directly from your pages.

There is a wrinkle worth being honest about. llms.txt files did briefly appear on some Google-owned properties, which fueled speculation that Google secretly uses them. The mundane explanation is that Google's internal CMS added support for generating the file and some teams did not remove it. Presence on a Google domain is not endorsement by Google Search.

But Google Is Not Every AI Engine#

Here is where honesty cuts both ways. Google saying it does not use llms.txt is not the same as every answer engine ignoring it. ChatGPT, Perplexity, and Claude each run their own retrieval pipelines, and none of them has published a clear statement that they read or ignore llms.txt.

So the accurate position is not "llms.txt is useless." It is "llms.txt has no proven citation benefit on any engine, and an explicit non-benefit on the largest one." That is a weaker claim than the hype and a stronger claim than the file deserves. Absence of published evidence is not proof it does nothing, but it is a reason not to build a strategy on it.

Why It Is So Popular Anyway#

Three forces keep llms.txt trending despite the evidence:

• Search demand. The term draws roughly 5,400 searches a month, which means a steady audience of marketers hunting for the next easy win, and a steady supply of posts written to meet that demand.
• The Lighthouse effect. Chrome's Lighthouse auditing tool added an llms.txt check. Many people read a Lighthouse flag as a Google ranking endorsement. It is not. Lighthouse is a developer auditing tool, separate from Search ranking systems.
• Implementation is trivial. You can write an llms.txt in ten minutes. Easy things spread faster than effective things, because trying them costs almost nothing.

None of those three forces is evidence that the file earns citations. They are evidence that it is convenient to talk about.

What Actually Drives Citations Instead#

If you have time budgeted for AI visibility, spend it here first, in this order:

1. Crawl access. Confirm GPTBot, ClaudeBot, PerplexityBot, and Google-Extended are allowed in robots.txt. A blocked crawler zeroes out everything downstream. This is the single most important file, and it is not llms.txt.
2. Server-side rendering. Your content must be in the raw HTML. Most AI indexing pipelines do not execute your JavaScript, so client-only content is invisible to them.
3. Structured data. FAQPage, HowTo, and Article schema help engines classify your content as an answerable, authoritative source.
4. Answer-first structure. Lead with the direct answer. Engines favor the page that most cleanly resolves the query.

Every one of these has measured correlation with citation in real tests. llms.txt has none. That is the entire priority argument in one sentence.

The Lazy, Honest Verdict#

Publish an llms.txt if you want to. It is cheap, harmless, and takes ten minutes, and there is a non-zero chance some future crawler chooses to honor it. Just do three things: do it last, do not expect it to move your numbers, and if you do publish it, measure your citation rate before and after so you are working from data rather than hope. For how to run that before-and-after measurement, see How to Measure Your AI Citation Rate.

For the traditional-search and answer-engine context around why self-declared signals keep failing while earned ones keep working, see the companion write-up on answer engine optimization on chudi.dev.

Related Reading#

• the five signals that actually drive citation: the pillar post for this cluster
• what AI crawlers see and what they miss: the retrievability layer llms.txt is often confused with
• the structured-data subset with measured citation impact: which schema types actually feed AI retrieval
• the before-and-after measurement protocol: how to test any change, including this one