What have the authority sites measured recently?

Let me start with something important: in the 2026 public research I have seen, when mainstream organizations talk about AI search, they focus on content and measurement, not on a particular configuration file. Ahrefs’ “How ChatGPT Does Query Fan-Out (15 Million Queries Studied)” of 2026-10-12 analyzes how the model breaks a question apart and which pages it retrieves. That bears directly on how much llms.txt is worth: if the model fetches real web pages by sub-query, then a list on your home page can only do so much.

Semrush’s “How to build an AI visibility dashboard” of 2026-10-12 is about turning AI visibility into metrics you can watch day to day. Work like that starts from accepting that visibility comes from your content being adopted, not from a declaration file.

Ahrefs’ 331,000-page study of 2026-07-06 points the same way: what gets processed is bad content. Put the three together and you land in one place — the deciding factor in AI search is content quality and accessibility, not whether some new file exists.

What exactly is llms.txt?

It is a plain-text Markdown file at your site’s root directory, with the path /llms.txt. Inside, you list your most important pages, each with a title and a one-line description, so that large models can more easily find “what on this site is worth reading.”

The logic behind the idea is reasonable: a model’s context is limited, and if a ready-made list exists, it does not have to crawl from scratch. But reasonable is not the same as adopted. Search engines accepted sitemaps back then because the mainstream engines supported them together; llms.txt has not reached that point, and each company’s crawling and retrieval strategy is not public.

So its actual status today is a “low-cost optional action”: doing it may not pay off, and not doing it shows no visible loss. Anything in that state does not deserve a heavy bet. For background, compare what a sitemap is to understand the difference between a standard and a non-standard.

How much evidence is there right now?

I have not found a credible source showing that AI citations or referral traffic rise steadily after adding llms.txt. Nor has any mainstream model vendor publicly said it reads this file first. That does not mean it is certainly useless; it means the evidence is insufficient at present.

Conversely, the evidence about content and crawlability is ample. Ahrefs’ query fan-out study of 2026-10-12 and its 331,000-page study of 2026-07-06 both point at the content itself. Semrush’s 2026-05-06 piece on tracking ChatGPT traffic also starts from the data side.

There is another point that is often overlooked: if the model cannot crawl your pages at all, the best-written list is useless. You can confirm the basics first — the page is accessible, no crawler is blocked by mistake, and the main content does not appear only after scripts run. These checks matter far more than adding a file. For how to do them, see Googlebot content accessibility and blocking AI crawlers on WordPress.

I ran one small check myself: I fetched the HTML of our main pages and looked at whether the body text could be read without executing JavaScript. It turned out that on several pages the product specifications were rendered by scripts, so a direct fetch returned only an empty shell. Problems like this matter far more than whether llms.txt exists, because if the model cannot get the content, everything discussed afterward is meaningless. The fix is usually not complicated: have the theme or plugin output the key content as static HTML. Work through it together with JavaScript and SEO.

How does it relate to sitemap and robots?

The order is that clear. A mistake in robots.txt can block crawlers outright, and that is a real loss; a sitemap helps search engines discover new pages, and that is the foundation; llms.txt is a nice extra. Some people call llms.txt “the robots.txt of the AI era,” but the analogy is inaccurate — robots is control while llms.txt is suggestion, and the two carry completely different force. For the relevant rules, see how to use robots, noindex, canonical, and sitemap.

If you do write one, how do you avoid wasting time?

If you decide to add one, keep it under half an hour. The structure is simple: one line describing the site at the top, then pages listed by category, each in the format “title: link: one-line description.” Pick 20 to 40 pages, choosing the ones that genuinely matter: main product pages, core Q&A pages, the company introduction, certification pages. Do not put the whole site in.

Once it is written, leave it alone; do not maintain it weekly. Unlike a sitemap, it does not need to update automatically with your content; if the upkeep costs more than the benefit, then simply do not add it. Spend the saved time revising one page that earns impressions but no clicks — the return is more certain. For how to do that, see how to write citable answer paragraphs.

One kind of site deserves a little more attention here: sites with many product lines, a complex page structure, and a tendency for models to grab the wrong emphasis. In that case a list really can help point the way. But it is still an aid, not a deciding factor. You can sort out the site structure together with structured data.

One last reminder: do not turn llms.txt into a marketing page. Some sites make it into an introduction to “how excellent our company is,” and that misses the point. Its job is navigation, not persuasion. Links should point to real, crawlable pages, and descriptions should say what question that page answers — not stack up adjectives. If a link points to a 404 or requires a login, it wastes one chance to be read.

There is an even plainer test: if deleting this file changes nothing on your site, then it is probably just psychological comfort. Technical actions that genuinely matter break something the moment you delete them — sitemaps, robots, sitemap submission. Measure against that standard and you will see more clearly where your time belongs. You can put the saved time into updating three pages that earn impressions but no clicks; the return is more certain.

A low-cost action checklist

Three steps, no more than an hour in total.

1. Check robots and accessibility first: confirm that no major crawler is blocked by mistake and that your main pages show their content without scripts.
2. Confirm the sitemap is healthy: it is reachable, includes the main pages, and has no large number of 404s.
3. Write llms.txt only if you have spare capacity: list 20 to 40 core pages, finish within half an hour, and stop maintaining it after that.

The value of these three steps declines as you go, so do them in order. If you only pick one, pick the first. To see the overall priority order for AI visibility, read how to set up AI visibility monitoring.

The questions owners ask most often

Is llms.txt an official standard?

No. It is a convention proposed by the community; it has not become a formal W3C or search-engine specification, and no mainstream model has publicly committed to reading it. Writing one costs something, but do not expect a certain return.

Will not writing llms.txt affect AI citations?

There is currently no reliable evidence that skipping it gets you ignored. Models mainly crawl and retrieve pages themselves; whether you get cited depends on content quality and accessibility.

Is it still worth adding?

If it only takes half an hour, adding one does no harm, and at least it lists your important pages clearly. But it should not take budget and time you meant to spend improving content.

How is it different from a sitemap?

A sitemap is a long-standing standard search engines use, oriented toward crawl discovery; llms.txt is a page list for large models, oriented toward content navigation. The former is infrastructure; the latter is more like an experiment.