llms.txt: what the data actually says (and why I still have one)

llms.txt: what the data actually says

llms.txt does almost nothing. Ahrefs analysed 137,000 domains: 97% of llms.txt files received zero requests in May 2026, and Google states in writing that it does not use them. I still maintain one on this site, for one precise and measurable reason I explain below. This is not a reversal, it is an order of priorities.

What the Ahrefs study found

Ahrefs published an analysis in June 2026 based on server logs from 137,000 domains. The figures:

Measure	Value
Domains publishing an llms.txt	28%
Valid files receiving zero requests in May 2026	97%
Valid files receiving at least one request	~1,100 out of ~38,000
Share of AI retrieval bots in requests	1.1%
GPTBot share	4.51%
ClaudeBot share	0.80%
Top actual requester: SEO audit tools	21.7%

Source: Ahrefs, We Analyzed 137K Sites, June 2026.

The number worth keeping is not the 97%. It is the 21.7%. The main consumer of llms.txt files is not AI engines, it is SEO audit tools checking whether the file exists. Put plainly: the file is mostly read by the tools that tell you to have one.

What Google says

Google has not been ambiguous, and never was.

Gary Illyes, at Search Central Live in July 2025, stated that Google does not support llms.txt and does not plan to.

John Mueller compared the file to the meta keywords tag: a file declaring what a site claims to be about, when Google would rather read the site. Asked on Bluesky whether Google properties carrying llms.txt files amounted to an endorsement, his answer was one word: no.

Finally, Google's official AI optimisation documentation, updated 15 June 2026, is explicit:

You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them.

So why do I have one

Because I measured one real effect, and exactly one.

On 21 July 2026, this site's Lighthouse audit in the Agentic Browsing category scored 2/3. The failing audit was named llms-txt and the reason was precise: File does not appear to contain any links. My file existed but listed URLs as plain numbered text instead of Markdown links, which is what the llmstxt.org specification requires.

I rewrote the file to spec: a single H1, a summary blockquote, H2 sections with entries formatted as - [Title](url): description. The audit moved to 3/3. That is measured before and after, on the same URL, with the same Lighthouse version.

So here is the only effect I can attest to: a Lighthouse category going from 2/3 to 3/3. Not one extra AI citation. Not one lead. A box ticked in a Google measurement tool, while Google states it does not use the file. The irony is complete and I own it.

What I did not measure

I have not run an A/B test with and without llms.txt on comparable sites. Nobody serious has published one that I know of, and I am not going to invent figures to make this article sell better.

What it would actually take to settle this:

Two comparable sites, same sector, same authority, same publishing cadence
llms.txt on one, not the other
Citation tracking across ChatGPT, Perplexity, Claude and Google AI Overview over at least 90 days
Server logs to verify who actually reads the file

Until that protocol has run, everything circulating on this topic is opinion, including the forward-looking part of this article.

Where the budget should go instead

llms.txt costs thirty minutes. It is not a cost problem, it is an ordering problem. If your GEO budget is limited, here is what produces verifiable effects before it.

Priority	Lever	Why it comes first
1	AI bot access in robots.txt	If GPTBot and ClaudeBot are blocked, everything else is pointless. It is binary and checkable in 10 seconds.
2	Valid Schema.org	AI engines resolve entities. Correct markup tells them who you are, where free text leaves it to interpretation.
3	Direct-answer content	A passage that already answers the question can be lifted as-is. An article with a preamble has to be summarised first.
4	Speed and server rendering	Content a crawler cannot see does not exist, however good it is.
5	External citations and mentions	Engines cite the corpus they trust, not the domain that declares itself competent.
6	llms.txt	Thirty minutes, one Lighthouse box, no demonstrated effect beyond that.

What I actually recommend

Do it, but last, and expect nothing from it. Thirty minutes for a spec-compliant file is a negligible cost and a free option if the specification does get adopted. What would be wrong is billing it to a client as an AI visibility service, or presenting it as a citation lever. The data says otherwise.

If a provider sells you llms.txt as a pillar of their GEO method, ask them about the Ahrefs figures. Their reaction will tell you more than their deck.

FAQ

Is llms.txt useful in 2026?

No visibility effect has been publicly demonstrated. Ahrefs measures that 97% of files received zero requests in May 2026 and that AI retrieval bots account for 1.1% of the requests received by those that did. Google states it does not use them. The only effect I have personally measured is a Lighthouse Agentic Browsing audit moving from 2/3 to 3/3.

Does Google use llms.txt?

No. Google's official AI optimisation documentation, updated 15 June 2026, states that no machine readable file is needed to appear in Google Search, including its generative AI capabilities, because Google Search does not use them. Gary Illyes had already said so at Search Central Live in July 2025.

Does ChatGPT read my llms.txt?

Rarely. In the Ahrefs data, GPTBot accounts for 4.51% of requests to llms.txt files, and AI retrieval bots as a whole for 1.1%. The top actual requester is the SEO audit tool category, at 21.7%.

Should I create one anyway?

Yes, but last and with no dedicated budget. Thirty minutes for a file compliant with the llmstxt.org specification is a free option in case adoption grows. What is not defensible is selling it as an AI citation lever, or prioritising it above bot access, Schema.org and content format.

How do I write a compliant llms.txt?

The llmstxt.org specification asks for a single H1 heading, a summary blockquote directly below it, then H2 sections containing entries formatted as - [Title](url): description. The most common mistake, and the one I had myself, is listing URLs as plain text: the Lighthouse audit then returns File does not appear to contain any links and fails.

Why does this article contain no A/B test?

Because I did not run one. A credible A/B test needs two comparable sites tracked over at least 90 days with server log access. Publishing an invented figure on a topic where everyone is looking for proof would be exactly the problem this article calls out.

Sources

Ahrefs, We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read, June 2026
Search Engine Journal, 97% Of llms.txt Files Got No Requests, Ahrefs Data Shows
Search Engine Journal, Google Confirms LLMs.txt Has No Current Implementation
Search Engine Roundtable, Google Search Team Does Not Endorse LLMs.txt Files
Google AI optimisation documentation, updated 15 June 2026
First-party measurement: Lighthouse audit of thibautcampana.com, Agentic Browsing category, before and after rewriting the file, 21 July 2026