When a new category appears, the first thing that gets sold is whatever can be delivered fast. In the business of showing up in AI answers, two deliverables are almost always in the first proposal a company receives. Both are cheap to produce and both look good in a report.
It is worth looking at what the available evidence says, because it is less enthusiastic than the proposals.
One: the llms.txt file
The idea is nice. A text file at the root of the site, similar to robots.txt, where a company explains to language models what it is and which pages matter. It takes an afternoon.
The problem is that, for it to work, somebody has to go and fetch it.
Ahrefs checked 137,210 domains with traffic in May 2026. 28% had the file published, around 38,000 sites. Of those, 97% received not a single request to that file all month. Not from a robot, not from a person. All the measured traffic went to the remaining 3%.
In other words: the file is there, and nobody opens it.
This doesn't mean it does harm. It costs next to nothing and it might be used someday. What it does mean is that it is not a lever, and if it sits at the top of a proposal as though it were, that deserves a question.
Two: structured data
This one is more interesting, because the evidence has two layers and the first one misleads.
Structured data is an invisible block of code that tells a machine, in an orderly way, what this page is: a company, a product, an article, with its name, address, author.
The first layer is a correlation, and it is strong: across six million URLs, pages cited by an AI were almost three times more likely to carry structured data. That is where most sales decks stop.
The second layer is the one that counts. The same team tracked 1,885 pages that added structured data between August 2025 and March 2026, and compared them against about 4,000 similar control pages that didn't, measuring 30 days before and 30 days after.
The result: citations in Google's AI Overviews fell 4.6%, while AI Mode and ChatGPT rose 2.4% and 2.2%, two figures the authors themselves describe as statistically indistinguishable from zero.
There is an honest caveat worth stating, and it usually isn't: the sample was pages that already got cited heavily. So what was measured is "adding structured data to a page that is already cited doesn't increase citations". That is not the same as "structured data is useless".
We still add it, for a different reason: it helps Google, it shapes how a link looks when somebody shares it, and it keeps a company's details saying the same thing everywhere. Those are reasons enough. What doesn't hold up against what has been measured so far is "add structured data and you'll show up in ChatGPT".
Why this matters beyond the technical detail
Because both deliverables share a trait: they ship fast, they show up in a report, and they are hard to disprove from the outside. They are ideal proposal filler.
Meanwhile, what is strongly measured asks for more work: pages that can be read without running JavaScript, content that answers something concrete, and a company being named in places that are not its own website.
A practical way to use this: when a proposal in this category reaches you, ask for the evidence behind each line and the size of the effect. Not to make anyone uncomfortable, but because this is a young category where almost anything can still be claimed without being checked.
That is how we build the work, and it is laid out in Get recommended by AI.
