llms.txt is a dud, the 2026 skeptics say — they're half right
Several well-read 2026 posts argued that llms.txt is low-yield. They measured the right thing and reached a fair conclusion. The conclusion just doesn't apply to how llms.txt was meant to be deployed in the first place.
llms.txt is a dud, the 2026 skeptics say
The line is making the rounds: "We shipped llms.txt on a portfolio of 50 sites. Six months in, zero measurable lift. The file is a dud." Mintlify wrote a version of it. A Medium post titled "the llms.txt is dead" got picked up. Anthropic publishes an llms.txt for its own docs but doesn't crawl yours. Hacker News has had three threads.
If you're considering shipping the file, the skepticism is the loudest signal in your feed right now. And it's not wrong about what it measured.
It's wrong about what was measured.
This post separates the two questions that have been collapsed into one: "does llms.txt work on its own" (no, mostly not) and "does llms.txt belong in your AI-readiness stack" (yes, when shipped with the rest of the layers it was always meant to sit beside).
What the skeptics actually measured
Every "llms.txt is a dud" post in 2026 is measuring the same configuration:
- Site ships only
/llms.txt. - No
agents.json, noagent-instructions.md, no per-page Schema.org JSON-LD, no<link rel="alternate">discovery pointers in the homepage<head>. - They measure citation lift in ChatGPT, Claude, Perplexity, and Gemini AI Mode versus a control of similar sites with no AI-readiness layer at all.
Result, consistently, across the published measurements: lift is either statistically zero or in the single-digit-percent range — small enough to disappear in the noise of normal SEO variance.
That measurement is honest. The conclusion that follows from it ("ship llms.txt, get little") is honest. The headline that gets attached ("llms.txt is dead") is not — because the measurement isn't of llms.txt's value, it's of llms.txt's value deployed alone, in a configuration the spec never claimed would work.
What llms.txt actually is, and isn't
Read llmstxt.org's own framing closely. The file is positioned as a curated reading list: "here are the pages on this site worth reading, in priority order, with a one-line summary of each." That's it. That's the entire scope.
It doesn't claim to:
- Tell agents what they can do on your site (that's
agents.json). - Tell agents how to behave once they decide to do something
(that's
agent-instructions.md). - Make your pages cite-worthy on their own (that's content, Schema.org, and conventional SEO).
- Get you indexed if you weren't (that's
sitemap.xmlandrobots.txt).
llms.txt solves one narrow problem: when an agent lands on your domain with a small context budget, which 3–20 of your URLs should it actually read?
Shipped alone, against a site with no other AI-readiness signals, that narrow problem is invisible. The agent is going to fail at the next step (figuring out what action surface your site exposes, how to invoke it, how to behave once invoked) and that failure swamps any reading-priority lift.
The skeptics' measurement is correct. It's the measurement of "we fixed the reading layer but left the acting and behaving layers broken" — and the answer to that is, predictably, that the reading fix isn't where the conversion gain lives.
Why the layer that does show citation lift is the stack
The 2026 posts that report measurable citation lift from AI-readiness files all describe the same configuration shape:
| Layer | The file | What it answers for the agent |
|---|---|---|
| Reading | llms.txt | "What on this site is worth my limited reading budget?" |
| Acting | agents.json | "What can I do on this site? Search? Buy? Book?" |
| Behaving | agent-instructions.md | "How should I do those things — what's allowed, what's the brand voice, what's the canonical answer?" |
| Discovery | <link rel="alternate"> in <head> | "Where is each of the above? Don't make me probe paths." |
| Structure | Schema.org JSON-LD per page | "What is this specific page about, in machine-friendly form?" |
That's the stack. Drop any one layer and citation lift collapses toward the skeptics' numbers. Ship all five and lift is real enough to read out of the noise.
The 11 sites we have data for that shipped the full stack between February and April 2026 averaged a 17 → 72 jump in our internal agent-readiness score, with the citation-lift signal correlated to score. (Detail in our before/after post.) Sites that shipped llms.txt alone in the same window averaged a 17 → 28 jump. Different deployment, different outcome — exactly the direction the skeptics' measurement predicts.
What this means if you're a site owner reading the headlines
Three concrete reads:
If you've been sitting on shipping
llms.txtbecause the skeptics said it's dead — they're right that the file alone is low-yield. They're wrong that the file is dead. Ship it as part of the four-file stack, not on its own, and the measurement changes.
If you've already shipped only
llms.txtand seen nothing — that's the expected result. You haven't disproved the file's value; you've measured the cost of partial deployment. Add the other layers and the lift shows up.
If you're publishing the skepticism — please name the configuration you measured. "We shipped llms.txt alone and saw no lift" is a useful data point. "llms.txt is dead" is not, because the file was never sold as a standalone fix.
The 2026 minimum, restated
For a site that does any transaction — sells, books, captures
leads — the minimum AI-readiness surface is the five-layer stack
above. The narrow case where llms.txt-alone is enough is a
pure-content publisher with no transactional surface and no
behavior guidance worth giving — and even then, ship agents.json
for the search endpoint at minimum.
If learning each spec isn't your afternoon, the BridgeToAgent kit generates the first four layers (everything except the per-page Schema.org density, which depends on your CMS) in under two minutes from your real DOM. $49 once, no subscription.
The skeptics are right that the file alone won't move the needle. They're not telling you to wait. They're telling you not to stop at one file. Take their advice.
Related
- What free
llms.txtgenerators don't tell you — and how to spot a bad file before it lowers your Lighthouse score — the practical follow-up: six failure modes free generators produce and the 30-second spot-checks for each - agents.json vs WebMCP vs llms.txt — what every site needs in 2026
- llms.txt vs robots.txt → what the difference actually is
- WebMCP shipped in Chrome 146 — what your site needs now
- Run the free readiness audit →