Does llms.txt Actually Work? What the Data Says, and What v2 Changed

You put an llms.txt at the root of your site. It was supposed to help AI engines understand you better. Do they actually read it? That question now has a data-backed answer, and the answer is uncomfortable. This article covers both the evidence and what the v2 release of August 2026 changed — and ends with the measured finding about what actually earns citations. Full disclosure: this site publishes an llms.txt too. This piece questions my own practice as much as anyone else’s.

What llms.txt is#

llms.txt is a file at the root of a site that presents its most important pages to language models in markdown. Jeremy Howard proposed it in September 2024 and the reasoning is sound: HTML is noisy, and a clean table of contents is more useful to a model.

The idea is sound; the problem is that no crawler asks for it.

What the server logs say#

This debate was settled by server logs rather than opinion. Ahrefs’ study published 15 June 2026 (May 2026 data, 137,210 domains) found:

• 28% of domains publish an llms.txt — adoption is genuinely broad.
• But 97% of those files received zero traffic.
• Among the 3% that were fetched, 96% of requests were bots, and only 19.5% came from named AI tools.
• The most striking finding: zero requests arrived for llms.txt files that do not exist. Nothing is proactively looking for the file.

That last point ends the argument. If nothing asks for a file, improving its contents cannot have an effect.

What Google says on the record#

Google’s position is documented. John Mueller, reported on 6 July 2026, was unambiguous: Google does not use llms.txt or llms-author.txt, and he knew of no other crawler or LLM confirming that they use these.

Google’s AI-features documentation supports that: it never mentions llms.txt. The only controls it names are robots.txt, nosnippet / data-nosnippet / max-snippet and noindex. The same document also answers the broader “GEO” expectation directly: there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.

And yet v2 shipped: what changed#

Here is the interesting part: with the evidence pointing that way, llms.txt v2 shipped in August 2026 — two years after v1, again authored by Jeremy Howard. And what v2 adds targets precisely the “nothing is looking” problem:

• Link relations for discoverability. You can now point at the file from the page: rel="alternate" type="text/markdown" and rel="describedby", deliverable as <link> elements or HTTP headers. You announce the file rather than hoping it is guessed.
• Flexible markdown URL forms: publish by appending the extension (/page.md) or by replacing it (/page.html → /page.md).
• Path-based coverage: a file covers pages beneath its path and the most specific file wins, so publishers covering only part of a site can participate.
• The llms_txt2ctx expansion tool was removed — and with it the mechanical semantics of the Optional section.

So v2 diagnosed the standard’s main weakness correctly. Whether the fix works will again be answered by server logs, not blog posts.

So what actually lifts citations?#

This is the more useful question, and it has been measured. The “What Gets Cited” study (arXiv:2605.25517, 25 May 2026) ran 252,000 trials across 6 LLMs, varying 18 content factors pairwise with brand anonymisation and counterbalanced ordering. The findings:

• Topical relevance and list position dominate. Everything else is small beside them.
• Explicit pricing data and recent timestamps consistently lift citation odds. Actionable and cheap.
• Formatting-only edits have minimal impact.

Two independent correlation findings sit alongside it. Ahrefs, 26 May 2025, across 75,000 brands: the strongest association with AI visibility was branded web mentions (r=0.664) and the weakest was backlinks (0.218). And the same team’s 11 August 2025 study (15,000 prompts) found only 12% of AI-cited URLs rank in Google’s top ten for the same prompt. Both are 2025 data; label them that way.

The click-through picture is a separate reality: Ahrefs’ 300,000-keyword study (updated 4 February 2026) found AI Overview presence correlating with a 58% lower click-through rate for the top-ranking page, up from 34.5% in their April 2025 measurement. Pew’s March 2025 data points the same direction: users clicked a traditional result on 8% of visits where an AI summary appeared, versus 15% without one.

A balanced conclusion#

1. Publish llms.txt, but set expectations correctly. The cost is near zero, there is no risk, and v2’s link relations may improve discoverability. Treat it as an insurance policy, not a traffic source.
2. Add v2’s <link> declaration. That is the one genuine improvement available: announce the file from the page.
3. Put your access policy in robots.txt. That is where it is actually enforced.
4. Spend your budget on content structure. The measured finding is clear: explicit pricing data, current dates and topical focus raise citation odds; formatting touch-ups do not.
5. Chase brand mentions. Branded mentions, not backlinks, show the strongest association with AI visibility.

And a warning: a significant share of the numbers circulating in this field trace to no verifiable source. When you hear a GEO claim, one question suffices — what date, how many samples, collected by whom? A claim without an answer is an opinion.

The most valuable thing in this field is not a new tactic but knowing which tactics have actually been measured. For the wider framing see generative engine optimization, for content and interface decisions in AI products see AI product UX, and for how models process text in the first place see how LLMs work.