In August we wrote about llms.txt and other machine surfaces. Since then the evidence has grown a lot, and it points one way. Having an llms.txt doesn't get you cited by AI answer engines, and most files are never read. But an llms.txt that an agent finds through a link, alongside Markdown versions of your pages, does help the agents that use it.
This post sums up that evidence, adds our own numbers from 102 developer companies, and explains what we changed in the Agent Readiness Score (ARS).
llms.txt is a map that agents follow when you hand it to them. On its own, sitting at the root of your site, it does almost nothing: almost no one requests it, and it doesn't change whether AI answers cite you. Linked from your pages, next to Markdown versions of those pages, it helps coding agents find their way.
Does llms.txt Get You Cited?
Three independent studies say no.
- SE Ranking (15 October, vendor study) looked at nearly 300,000 domains. Only 10.13% served an llms.txt, and among the highest-traffic sites the share was lower, 8.27%. It found the file had no effect on how AI systems cited sites; removing it from their model made the model more accurate. Caveat: SE Ranking doesn't name the AI systems or the data behind its citation counts.
- CueScout (15 October) compared 537 B2B software vendors, half of which served a valid llms.txt. Across 80 Google AI Mode searches, 24.3% of vendors with the file were cited against 21.8% without. Within bands of similar site size the difference wasn't significant (p = 0.56). None of the 708 cited links pointed at an llms.txt.
- AgentReady.org's controlled experiment served one docs site in 19 configurations. Adding llms.txt, like adding JSON-LD or sitemaps, made no difference to whether agents retrieved a planted answer. Only a JavaScript-only page and a 403 to agent user-agents stopped retrieval.
Does Anyone Read It?
Mostly not. Ahrefs (15 June, vendor data) checked server logs for 137,210 domains. Of the 28% that served a valid llms.txt, 97% got no requests at all for it in May 2026. About 3% of domains received all the traffic, and about 96% of requests came from bots, many of them SEO and "AI visibility" checkers rather than agents.
When It Helps
The picture changes when an agent is pointed at the file.
- Agents follow links to it. In AgentReady's traces of real agent runs, 86% of llms.txt fetches came from a page that linked to it. Agents asked for Markdown on about 65% of their web fetches.
- It helps navigation, together with Markdown. Mintlify's own benchmark (2,400 tests across 20 docs sites, vendor data) counted 404 errors per task. HTML pages averaged 2.23, Markdown pages 1.42, and Markdown pages with llms.txt 0.11: agents guessed far fewer wrong URLs when they had the index.
So the value isn't in having the file. It's in agents finding it, and that mostly happens when your pages link to it.
What We Found on 102 Developer Companies
We scanned the 102 reachable companies on our leaderboard with ARS 3.3 before releasing it:
| Sites | |
|---|---|
| Serve a real llms.txt (not an HTML error page) | 54 of 102 |
| …whose sampled links resolve (at least 80%) | 46 of 54 |
| …linked from the homepage | 7 of 54 |
Return Markdown when asked (Accept: text/markdown or a .md page) | 30 of 102 |
Advertise a Markdown alternate (<link rel="alternate" type="text/markdown">) | 7 of 102 |
Developer companies adopt llms.txt far more than the web at large: 53%, against SE Ranking's 10%. But nearly all of them leave it unlinked, which is exactly what the evidence says matters.
Sites with an llms.txt averaged 58 on ARS 3.3, and sites without averaged 39. Don't read that as cause and effect. llms.txt now earns almost nothing in ARS. Companies that publish one tend to invest in the rest of their agent surfaces too, and that's what the gap reflects.
The scan also caught a bug of ours. When checking that llms.txt links resolve, our scanner had counted any page over its 200 KB read limit as broken, even when it answered with HTTP 200. On the first run only 13 of 54 sites passed. With the bug fixed, 46 do. It had affected ARS 3.2 scores too, and it's fixed in 3.3.
What We Changed in ARS 3.3
| Check | ARS 3.2 | ARS 3.3 |
|---|---|---|
llms-txt-exists | 1 base point: absence cost points | 1 bonus point: absence costs nothing |
llms-txt-formatting | 2 base points | 1 bonus point |
llms-txt-links-resolve | 2 base points | 2 base points, with oversized pages no longer counted as broken |
llms-txt-graph-reachability | Bonus if "llms.txt" appeared anywhere in the HTML | Bonus only for an actual link to it |
| A 200 that isn't an llms.txt | Counted as a file unless it contained <html | Rejected unless it's non-HTML text with a Markdown heading or link |
markdown-alternate-link | — | New bonus: the page advertises a Markdown version that resolves |
A broken llms.txt still costs points. Agents that follow it reach dead ends, which is worse than having no index at all.
What to Do
- Serve Markdown versions of your pages, and answer
Accept: text/markdown. That's what agents ask for most. - Advertise them with
<link rel="alternate" type="text/markdown">or aLinkheader. - If you have an llms.txt, link to it from your docs pages, and keep its links working.
- Don't expect it to change your AI citations. Three studies say it won't.
- Make sure agents can get in at all. In AgentReady's experiment, blocking agents or rendering only with JavaScript did the damage; llms.txt didn't fix it.
Sources
- SE Ranking: LLMs.txt: Why Brands Rely On It and Why It Doesn't Work (15 Oct 2026)
- CueScout: Does llms.txt work? (15 Oct 2026)
- Ahrefs: We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read (15 Jun 2026)
- AgentReady.org: spec and data (CC BY 4.0 data)
- Mintlify: The State of Knowledge 2026, highlights
Third-party figures as published, read on 23 October 2026. Our figures come from scanning 102 leaderboard companies with ARS 3.3 on 24 October 2026. Disclosure: Glintbase sells agent-readiness scoring.
Run a Free AI Readiness Audit
Identify broken parameter signatures, missing Model Context Protocol schemas, and sitemap issues in your developer documentation.
