Does llms.txt work?
Here is what the data actually shows

A great deal has been written about llms.txt, and nearly all of it by people with something to sell. We sell something too — our platform can publish and maintain the file automatically. That is precisely why this page puts the numbers first and the conclusion second. If you only want to know what the file is, the definition lives on what is llms.txt.

What do the engine builders say?

Google has been asked directly. In June 2026, John Mueller of Google said the question of llms.txt's effect is purely speculative for now — the file has existed for years without AI systems using it. His practical advice: create the file the day an AI platform that actually sends you customers asks for it.

It is no better elsewhere. Neither OpenAI, Anthropic nor Perplexity documents llms.txt as a citation or visibility signal in their official material. As of today, no major provider has committed to the file in production.

What does the traffic data show?

Opinions are one thing; fetches are another. In May 2026, Ahrefs studied 137,210 domains and looked at who actually retrieves these files. The result is bleak for anyone who has made one:

• 28 % of the domains studied publish an llms.txt.
• 97 % of those files received zero fetches in May 2026. Not few — zero.
• Among the remaining 3 %, most requests came from SEO tools and general crawlers. AI search bots accounted for 1.1 % of total requests.
• And the detail that settles it: no AI bot ever requested an llms.txt that did not exist. They do not go looking for the file. They stumble across it.

That last point deserves a moment. A signal nobody actively seeks out cannot, by definition, be a prerequisite for being found.

What do our own measurements show?

We measure too. In a scan of 465 search-visible Danish company websites (21 July 2026), 89 % had no llms.txt — the full analysis with method and caveats. Across the Nordic panel of 1,252 sites the figure is 88.2 % (±1.8 percentage points).

That number is often used — by us as well — as an argument to hurry. Note what it actually says: it measures adoption, not effect. That nine in ten are missing something does not prove the tenth benefits from having it.

We have, however, measured something close to the question. We asked four AI engines 576 questions about those 465 companies and compared the answers with each site's technical setup:

• The pages AI cited as a source had a technical score of 77.9 against 72.7 for the rest. The difference is strongly significant (p=0.000).
• The companies AI recommended scored 74.9 — against 74.9 for those not recommended. No difference whatsoever (p=0.988).

Technical setup therefore decides whether your page is used as a source, and nothing about whether your company is recommended. Two outcomes, two entirely different answers — see the full analysis.

What we do not know

This is where our data stops, and it should be said plainly: we have not isolated the effect of llms.txt on its own. The technical score above is composed of 15 dimensions, of which llms.txt is one. That the score predicts citation does not prove the llms.txt part is doing the work — it could just as easily be crawler access, server-side rendering or structured data, which the file merely travels alongside.

Settling that requires a controlled experiment: same sites, same period, the file as the only change. We have not run it. Until someone does, honest claims about the independent effect of llms.txt are not possible — ours or anyone else's.

So should you create one?

Yes — but for the right reason, and with the right amount of time.

The argument is not that the file works today. It is that it costs ten minutes once, cannot hurt, and leaves you ready the day an engine starts using it. That is cheap insurance, not a growth engine. The generator here is free and requires no login.

Skip it if you are considering spending hours maintaining a parallel set of markdown pages for robots. That investment has no documented return — and those hours do more elsewhere.

And if someone sells you llms.txt as the way into ChatGPT: ask for their data. That is the whole test.

What actually moves the needle

Our own measurement points to where the effort belongs if you want to be cited:

1. Crawler access. If your robots.txt blocks GPTBot, nothing else matters. In our scan, 5 % of sites did exactly that — often without knowing.
2. Content readable without JavaScript. If the AI crawler fetches an empty page, your content does not exist to it.
3. Structured data. Machine-readable marking of who you are and what you offer.
4. Citable content. Concrete numbers, clear answers and a source that can be named — that is what gets pulled into an AI answer.

Those four are also what GEO is about. To see where your own site stands on all 15 dimensions, the free GEO test takes a minute.

Frequently asked questions

Does llms.txt work?

Not as a standalone signal today. Google called the effect purely speculative in June 2026, and 97 % of all llms.txt files were never fetched at all in May 2026 (Ahrefs, 137,210 domains). Create one anyway: it costs ten minutes once, cannot hurt, and leaves you ready if an engine starts using it.

Does ChatGPT use llms.txt?

It is not documented. Neither OpenAI, Anthropic nor Perplexity describes the file as a citation or visibility signal. In Ahrefs' traffic data, AI search bots accounted for 1.1 % of fetches — and no AI bot ever asked for an llms.txt that did not exist.

Can having an llms.txt hurt?

No. The worst realistic cost is your own time, if you start maintaining a whole parallel set of pages for robots. The file itself sits harmlessly at the root of your domain.

Should I remove the one I already made?

No. It costs nothing to leave in place, and it is there if the picture changes. Just do not spend more time on it until something fetches it.

What should I do instead?

What our measurement actually points to: give AI crawlers access in robots.txt, make sure the content can be read without JavaScript, mark it up with structured data, and write something worth citing. The free GEO test shows where you stand on all 15 dimensions.

Sources: Ahrefs, llms.txt crawl analysis (137,210 domains, May 2026) · Search Engine Journal on John Mueller's answer, June 2026 · Geoa, own scan 21 July 2026 (n=465) and AI answer analysis 26 July 2026 (576 answers). The method behind our own figures is on the methodology page.