We shipped llms.txt. The evidence says it changes nothing.

Last week we published an llms.txt file on this site: a plain-text summary of what we do, what each service costs, and where to find things, written for AI assistants rather than people. Plenty of agencies are selling this as the thing that gets you cited by ChatGPT.

We think that claim is not supported, and since we sell an audit that produces one of these files, we should say so plainly.

What the evidence actually shows

Three findings, none of them encouraging.

An analysis of roughly 300,000 domains found that llms.txt adoption is low and has no measurable link to how often a site is cited by AI assistants. A separate crawl by SE Ranking put adoption at 10.13 % of domains, so nearly nine in ten sites do not have one. And as of the first quarter of 2026, no major AI company — OpenAI, Google, Anthropic, Meta or Mistral — has publicly committed to reading or acting on the file in production.

Google's John Mueller compared it to the old keywords meta tag: a thing sites declare about themselves that the systems they are declaring it to have not agreed to read. That comparison is uncomfortable and roughly fair.

Meanwhile SEO tools flag a missing llms.txt as a site issue, which manufactures urgency without evidence. That is how a convention with no demonstrated effect becomes a line item on a proposal.

So why did we publish one

Three honest reasons, none of which is "it will get us cited".

It costs about twenty minutes and one file. It is a genuinely useful, accurate summary of the business that a person or a machine can read in one pass, which is more than can be said for most navigation. And if any assistant does start reading it, we are already there.

That is the whole case. It is cheap, it is truthful, and it might pay off later. What it is not is a growth lever, and anyone selling it as one is ahead of the evidence.

The traffic is small and the visitors are not

The more useful question is what AI assistants send you at all. The answer is: not much yet, growing fast.

A third of one percent is easy to dismiss, and dismissing it on volume alone would be a mistake, because the composition is unusual. Where the traffic does come from is concentrated almost entirely in one place.

ChatGPT's share is falling even as its absolute traffic grows, because the rest of the market is growing faster. For a small Australian business deciding where to spend attention, though, the practical read is that this is close to a single-channel question today.

The visitors also behave differently. Reported conversion advantages over organic search range from about 4 times to more than 20 times depending on who is measuring and what they count, which is a spread wide enough that you should treat any single figure with suspicion. The direction is consistent across studies even though the magnitude is not: people arriving from an assistant have usually done their comparison already.

The number that should worry professional services

Citation rates are not evenly distributed, and the category most Australian consultancies sit in is the worst one.

Overall citation presence in US ChatGPT prompts rose from about 1.6 % in June 2025 to about 6.8 % by May 2026, so the direction is up sharply. But if you sell advice rather than flights, assistants are still answering most questions in your category without naming anyone. Being the exception is worth more here than it is in travel, and it is harder.

What appears to matter instead

Nothing here is a guarantee, and anyone who offers you one is guessing. But the things that plausibly affect whether an assistant can cite you are unglamorous and mostly under your control.

Being readable at all: assistants cannot cite a page their crawler cannot fetch or render. Blocking their crawlers is a legitimate choice, but it is a choice with a cost, and most businesses have made it by accident rather than on purpose.

Being answer-shaped: a page that states the answer in its first paragraph and supports it underneath is easier to lift a correct sentence from than one that builds to a conclusion over eight hundred words. This is good writing anyway.

Being consistent about facts: prices, service names, locations and hours that match across your site, your structured data and third-party listings. Assistants synthesise from several sources, and contradictions get you dropped or misquoted.

Being cited elsewhere: what other credible sites say about you appears to matter more than what you say about yourself, which is the same lesson search taught everyone twenty years ago.

And one more, which is where the honest measurement lives: ask them. Put your twenty real customer questions to ChatGPT, Claude, Perplexity, Copilot and Google, write down what they say about you, then check again in thirty days. It takes an afternoon and it is the only number in this article that is actually about your business.

What we did on this site

For the record, so you can check it: we publish llms.txt and a longer llms-full.txt; we allow the major assistant crawlers explicitly in robots.txt rather than by default; and every service page carries structured data whose prices match the prices printed on the page. We have a baseline set of questions and answers recorded from before those changes.

When we have a second reading, we will publish it here, including the parts that did not move. If llms.txt turns out to do nothing measurable, that is what this post will say.

Sources

• Search Engine Journal, llms.txt shows no clear effect on AI citations, based on 300k domains
• SE Ranking, llms.txt: why brands rely on it and why it doesn't work
• SE Ranking, analysis of top AI search engines and AI referral traffic
• SE Ranking, referral traffic from ChatGPT hit an all-time high in May 2026
• Similarweb, generative AI search statistics 2026