There Is No Secret File That Makes AI Recommend You

Here is a claim doing the rounds in marketing circles: add a file called llms.txt to your website and you will start showing up when people ask ChatGPT, Claude or Gemini for a recommendation. It is an appealing idea, because it promises a clean technical lever for something that otherwise feels out of your hands. The trouble is that the evidence for it is thin to the point of absent, and it is worth saying so before more businesses spend effort chasing it.

What the data shows

llms.txt is a proposed standard: a plain file that tells AI crawlers which pages matter on your site. The intention is reasonable. The results are not there. An analysis by Ahrefs found that 97% of valid llms.txt files received no fetches at all, and among the small share that did see any traffic, AI tools accounted for less than a fifth of requests. A separate study of nearly 14,000 domains found that having the file did not predict whether a site got cited in AI answers. Another reported no statistical correlation between implementing it and how often a site was cited.

The platforms are unusually direct about this too. Google's published guidance states that llms.txt will not positively or negatively affect your visibility or rankings in Search. OpenAI and Anthropic do publish llms.txt files, but for their own developer documentation, and their guidance to site owners points to the long-standing robots.txt instead. None of the major AI companies tells you to rely on the file to get recommended.

Why the myth keeps spreading

Two things keep it alive. Adoption is genuinely rising, reportedly up around 8.8 times in a year to more than 36,000 sites, which looks like momentum until you notice that most of the growth comes from website platforms adding the file by default rather than owners choosing it for results. And the broader market around AI visibility has an incentive to sell simple fixes. Several of the loudest sources on this are vendors with a product to push, which is reason enough to read their claims carefully.

This is the same shape as other tidy-sounding shortcuts we have examined, from the belief that answers in ChatGPT are quietly for sale to the trust people place in chatbots as reliable fact-checkers. The common thread is a wish for a dial you can turn, when the real mechanism is messier.

What actually seems to move the needle

In fairness, absence of a measurable effect today is not proof the file can never matter, and keeping a low-maintenance llms.txt around does no harm. But if the goal is to be mentioned by an AI assistant, the signals that correlate are the unglamorous ones: being written about, cited and talked about elsewhere. One analysis of a million AI citations found that the large majority traced back to earned media rather than anything on the brand's own site. In other words, there is no secret file. Chatbots tend to recommend what the wider web already talks about, which is harder to fake and a good deal more durable.