We Checked 63 AI Visibility Agencies. Half Publish llms.txt.

What we measured, and what this does not prove

On 18 August 2026 we took the top twenty organic results for four commercial queries: generative engine optimization agency and aeo agency in the United States, agencia geo and agencia aeo in Spain. We removed social platforms, user forums, directories and encyclopedias, which leaves the website of a company that either sells the service or publishes about it. That is 63 domains, 34 in Spain and 29 in the United States.

Then we asked each domain three questions that any crawler asks: what does /robots.txt say, is there an /llms.txt, and what structured data sits on the homepage. No JavaScript, no rendering. We read what a crawler receives in the response.

Two controls are declared rather than hidden: our own site and Hyperdot Australia, an agency with the same positioning. Neither counts in any percentage below.

What this does not prove. It does not measure whether llms.txt works, because nobody outside the assistant vendors can measure that. It does not measure content quality. And a sample built from four SERPs is a sample of who ranks, not of the industry. We publish the method so the numbers can be argued with.

Half publish llms.txt, and a third of those files are broken

llms.txt is an optional plain text file at your domain root that lists and summarizes your important pages for language models. The proposal is simple: a heading, then sections of Markdown links.

32 of 63 sites publish one. Spain 16 of 34, the United States 16 of 29. That is a striking adoption rate for a file with no proven effect.

The quality is where it falls apart:

Four files have no links in them. A file whose entire purpose is to route a model to your pages, containing no pages. One site, greenbananaseo.com, ranks third in the United States for generative engine optimization agency and serves an llms.txt of zero bytes.

Ten of the thirty-two carry the signature of Rank Math or Yoast, meaning a plugin generated the file and nobody opened it afterwards. The largest we found is 59 KB, which is not an index of a site, it is the site pasted into a text file.

Publishing it does not separate who ranks

Here is the finding that changed how we talk about this file with clients.

Of the nine companies we measured in the top ten for agencia geo in Spain, seven do not publish an llms.txt at all, including the sites at positions one and three. Exactly one, at position four, publishes a valid file.

In the United States adoption among top results is higher, six of the eight we measured. So in one market the leaders skip it and in the other most have it. Neither pattern separates the ranked from the unranked, which is what you would expect from a file Google says its Search does not use.

That is not an argument to delete yours. It is an argument to stop selling it as a ranking requirement, and to stop buying it as one.

If you want the version of this with your own domain in it, our free AI crawler access checker reads your robots.txt with the real precedence rules, and our AI visibility checker scores the signals that actually decide whether an assistant names you.

The fear being sold does not happen

The most repeated warning in this industry is that you are blocking AI without knowing it. In this sample it is close to fiction.

Two sites out of 63 block an AI crawler, and both did it deliberately, with their own rules for GPTBot and ClaudeBot. Nobody blocked one by accident through a stray wildcard.

What we did find is a mix up worth more than the fear. Look at which crawler gets named:

Twice as many sites wrote a rule for the crawler that cannot cite them as for the one that can. If you only ever learn one distinction from this study, learn that one. Blocking GPTBot is a licensing decision about training. Blocking OAI-SearchBot is a decision to leave the answer.

Seven sites cannot serve their own rules

Eleven percent of the sample, 7 of 63, do not return readable rules to an ordinary HTTP client. Five answer with an HTML anti bot challenge page instead of text, one returns 404 because the file does not exist, and one returns a hard 403.

Two of those seven are notable. xufa.es sits at position three for agencia geo in Spain. minuttia.com publishes one of the "best GEO agencies" listicles that the rest of the industry links to.

The honest caveat, because we corrected ourselves on it mid study: an anti bot layer often keeps a whitelist of known AI crawlers, so this does not prove GPTBot gets the same challenge page. It proves the file is not universally readable and that nobody has checked. Checking takes one request.

We also corrected a defect in our own measurement. Our first pass flagged 23 valid llms.txt files because the heading test was not multiline, so any file with a line above the # failed, which is exactly what the Rank Math template produces. The corrected number is 22 valid out of 32, and it moved in the direction that was less flattering to our narrative. Both fixes are written down in the source document.

Structured data is healthier than expected, with two clean breaks

55 of 63 homepages carry at least one JSON-LD block, and 53 declare an Organization or an equivalent. That is a better result than we predicted before running it.

• 8 sites carry no structured data at all.
• 2 publish JSON-LD that does not parse. A broken block does not degrade gracefully. It simply does not exist for whatever reads it, and one of the two is a site publishing its own ranking of GEO agencies.
• 15 use FAQPage, the type most recommended in AEO writing, which means the majority of the people recommending it have not implemented it.

One more pattern, because it matches something we measured separately. Nine of the 34 Spanish sites put a city in the homepage title. In the United States, two of 29. Spanish agencies are competing on agencia geo plus a city while American ones fight the national head term, and the Spanish ones are right: the geographic qualifier is the cheapest way to reach a SERP you can actually win.

Five checks worth running on your own site this week

1. Request your own robots.txt from outside the browser. If it answers with HTML, a bot layer is sitting in front of your rules.
2. Decide GPTBot and OAI-SearchBot separately. They are different decisions: training versus citation.
3. If you publish llms.txt, open it. Confirm it has a heading and real links, and that a plugin did not write it for you.
4. Validate your JSON-LD. Not "does it exist", does it parse.
5. Write down today's numbers. Whatever you cannot measure later, you can at least have a baseline of.

What we would like taken from this

The tactical layer of AI visibility is sold with more confidence than the evidence supports, and the companies selling it do not follow their own advice on their own websites. That is not a scandal, it is a young discipline. But it is a reason to ask any provider, including us, which of their recommendations they have measured and which they inherited.

If you want the underlying decisions rather than the file checklist, we wrote about what Google actually documented for AI search, what llms.txt is for, and how to choose an agency for this work. And if you would rather we run this measurement on your domain instead of on ours, tell us what you sell and to whom.

We repeat the study in November 2026 with the same sample. If your site is in it and the numbers look wrong, write to us and we will recheck it in public.