blog: llms.txt is overhyped (#76)

Author: adamaltman

blog/images/llms-txt.jpg

@@ -0,0 +1,98 @@
+---
+template: '../@theme/templates/BlogPost'
+title: llms.txt is overhyped
+description: We built it, tested it, and checked the logs. llms.txt isn’t the “robots.txt for AI” — it’s mostly ignored. Here’s what actually matters.
+seo:
+  title: llms.txt is overhyped
+  description: Redocly experimented with llms.txt and found it mostly smoke, not fire. See the results, the logs, and what really matters for docs + AI.
+  image: ./images/llms-txt.jpg
+
+author: adam-altman
+date: 2025-08-20
+categories:
+  - developer-experience
+  - learning
+  - company-update
+  - dev-portal
+image: llms-txt.jpg
+---
+# LLMS.txt is overhyped
+
+Every now and then, the industry invents a new “standard” that's supposed to solve everything.
+Right now the hype train is parked at llms.txt.
+People call it the _robots.txt_ for AI.
+Cute analogy.
+The problem is: it doesn't actually work that way.
+
+## We built it anyway
+
+At Redocly, we like to experiment.
+So we added automatic llms.txt support to our platform.
+Turn it on, it generates the file.
+Easy.
+We even ran a full [Phronesis project](./phronesis.md) on it — testing across models, prompts, and scenarios.
+
+The results? Pretty underwhelming.
+- Unless you explicitly paste the llms.txt file into the LLM, it doesn't do anything.
+- When you do paste it, you'd get better results just pasting the actual Markdown docs.
+- No model we tested spontaneously “read” or respected llms.txt on its own.
+
+That's not a governance breakthrough.
+That's a parlor trick.
+
+## The logs don't lie
+
+We also pulled logs.
+How often are `llms.txt` and `llms-full.txt` even being accessed?
+Answer: basically never.
+When they are, it looks like someone experimenting in a single LLM session, not systematic use by the models.
+
+Michael O'Neill at the University of Iowa checked too — same conclusion: [don't lose sleep over llms.txt](https://www.linkedin.com/pulse/dont-worry-llmstxt-yet-maybe-ever-michael-o-neill-huifc/?trackingId=epKdG7eoRpmJ5sNO7dFNwQ%3D%3D).
+
+## The silver lining
+
+The best thing about building llms.txt wasn't llms.txt. It was what came after:
+- one-click copy of any page in Markdown,
+- links you can drop straight into ChatGPT or Claude,
+- smooth handoff from docs → AI assistant.
+
+That's useful today.
+That's how people actually want to interact with docs in an AI-first world.
+
+## A tale of two experiments
+
+Not all experiments flop.
+Last week, we ran another Phronesis project, this time on two of our new MCP features (not yet public).
+The difference was night and day.
+
+With Docs MCPs, we saw real value.
+They made docs instantly more useful inside AI workflows.
+The debriefs weren’t full of head-scratching like with llms.txt — they were full of smiles.
+
+That’s the difference between smoke and fire.
+LLMS.txt is smoke.
+Docs MCPs are fire.
+
+
+## What really matters
+
+Focus on making good content.
+
+If we want content governance in AI, it won't come from a text file no one reads.
+It'll come from:
+- licensing,
+- attribution,
+- legal clarity,
+- real standards AI companies can't ignore.
+
+Until then, `llms.txt` is just… there.
+More checkbox than standard.
+
+## My take
+
+We tried it. We measured it.
+We learned from it.
+And now we can say it out loud:
+**llms.txt is overhyped.**
+
+The sooner we move past the illusion, the sooner we can focus on solutions that actually matter.