llms.txt Does Not Get You Cited. Here Is What It Is Actually For.

There is a small text file being sold as a shortcut through a problem that has no shortcuts.

Put llms.txt at the root of your site, the pitch goes, and ChatGPT, Claude, Gemini, and Perplexity will finally understand what you do. You will get cited. Your rankings will move. The robots will stop overlooking you.

That would be a lovely product feature. It is not how this works.

The best public crawl data we have says the major answer engines barely request llms.txt at all. One recent study of more than 100,000 sites found an association between the file and AI visibility, but association is exactly what you would expect when technically mature sites do many sensible things at once. Other server-log studies found that crawlers request robots.txt hundreds of times more often, and that most llms.txt requests come from SEO tools checking whether the file exists.

So should you create one? Probably. It takes twenty minutes. Just stop treating it as a growth strategy.

What llms.txt Actually Is

The proposal is simple: a Markdown file at /llms.txt that gives a language model a compact map of your site. A heading, a sentence explaining the product, links to important documentation, maybe a list of endpoints or guides.

That is a reasonable convention. Documentation is full of navigation, sidebars, marketing copy, version selectors, and JavaScript that sometimes makes the useful part harder to extract. A clean Markdown index can be genuinely helpful to a coding agent that is trying to answer, "How do I authenticate with this API?"

This is why Stripe and other developer-first companies have something in the shape of an LLM-friendly docs entry point. Their use case is not search ranking. It is making the correct current information easier to retrieve when an agent is already reading their documentation.

That distinction matters:

• llms.txt can help an agent navigate your docs after it arrives.
• It does not make the agent arrive.
• It does not prove that your claims are true.
• It does not turn a site nobody mentions into a recommended product.

The file is documentation hygiene. Good hygiene matters. It does not replace a product, a reputation, or pages that answer a real question better than the alternatives.

Does llms.txt Improve AI Visibility?

Not in a way anyone has shown convincingly.

No major answer engine has said that llms.txt is a ranking factor or a citation input. Google has been especially clear that its ordinary search guidance still applies to AI features. Make pages accessible. Make them useful. Use descriptive titles. Give users a reason to choose them. The answer is boring because the answer has always been boring.

The temptation to believe otherwise is understandable. Zero-click search is now the majority of Google searches, so a file that promises a direct line to the answer box feels like a thing you should not miss.

But a file cannot solve the actual citation question: why should a model name you rather than the company it has already seen in documentation, reviews, Reddit threads, comparison pages, editorial coverage, and its own training data?

It cannot.

The strongest signal in the studies is not file format. It is clarity plus corroboration. Brands that appear in several independent, trustworthy places get named more often. Brands whose own site states exactly who they serve, what they do, how they compare, and what it costs are easier to cite accurately once the model encounters them. That is work. A text file is not.

The Correlation Trap Is Doing Most of the Selling

Here is how a bad conclusion becomes a marketing claim.

Someone scans 100,000 websites. Sites with llms.txt score better in an AI visibility tool. They publish the graph. A software company turns the graph into "add an llms.txt file to increase citations." Another company builds a scanner that gives you a red warning until you add one.

What the graph actually captures is that a team that publishes an LLM file is more likely to also have:

• documentation that is current and crawlable
• schema markup and clean canonical URLs
• comparison pages with a clear category definition
• an active product with reviews, backlinks, and community mentions
• someone who notices when a page is broken

Those things have a chance of helping. The file is a useful proxy for "this team cares about its website," not evidence of causation.

You can see the same problem with the whole category of AI visibility dashboards. A graph showing that your competitor was cited 18 times and you were cited 6 times is a diagnosis, not a plan. That gap is exactly why I built beseen.so: I wanted the missing page and rewrite, not a more decorative thermometer.

When You Should Publish llms.txt

I would add it in three cases.

You have developer documentation. If your users paste prompts into coding agents, a compact entry point reduces the odds that the agent follows an old blog post, a deprecated endpoint, or a random answer from Stack Overflow. Point to versioned guides and make the current path obvious.

Your site is large or hard to navigate. A docs site with hundreds of pages, a dashboard that renders client-side, or a product with multiple APIs can benefit from an explicit map. It is the same argument for a good README: people should not have to play scavenger hunt to understand the system.

You can keep it current. An inaccurate llms.txt is worse than no file. It gives a confident agent a clean, concise way to be wrong. Put ownership on it, review it when you change a public API, and remove old links.

For a small marketing site with five pages and no docs, it is not a priority. Your homepage should already be easy to understand. If it is not, fixing the homepage will do more for people and crawlers than creating a separate map of the confusing homepage.

What To Do Before You Touch It

If you want to be named by assistants, use this order instead.

1. Make the category and audience painfully clear

The first screen of your homepage should answer three questions without a scroll: what is this, who is it for, and what does it replace or improve? Vague claims such as "the operating system for modern teams" make a design team feel clever and a retrieval system useless.

Use the words customers use. If you are a rank tracker that checks AI prompts, say that. If you are an invoice tool for freelancers in Romania, say that. Clarity is not boring. Clarity is a feature.

2. Publish pages with a reason to exist

Do not make a hundred generic FAQ pages because an agent can write them cheaply. Google has been aggressive about low-value scaled pages, and answer engines have no reason to cite a page that only rearranges facts from elsewhere. The programmatic SEO pages that survive contain data, workflows, tools, or comparisons that could not sensibly be mass-produced by hand.

The same standard works for hand-written content. A useful page resolves a decision. It does not just occupy a keyword.

3. Earn evidence outside your own domain

An assistant should not need to take your word for it. Get a real customer quote. Publish a specific case study. Be useful in the communities where your buyer asks questions. Get listed on a relevant directory, not 200 abandoned ones. Let someone else describe why you are good.

This is slower than adding a file. It is also the part that cannot be copied by the next person who reads the same checklist.

4. Measure a small set of recurring questions

Pick ten prompts a buyer would realistically ask. Run them on the engines you care about. Record who appears, what sources get cited, and which of your pages actually address the missing claim. Then change one thing and re-check it later.

Do not declare victory because you saw your name once. Answer engines vary by session, model, locale, and the exact wording of the prompt. You need a trend, not a screenshot.

A Sensible llms.txt, If You Still Want One

Here is the entire shape I would use for a product with public docs:

# Product name

> One clear sentence on what the product does and who it is for.

## Core documentation

- [Getting started](https://example.com/docs/getting-started)
- [Authentication](https://example.com/docs/authentication)
- [API reference](https://example.com/docs/api)

## Product information

- [Pricing](https://example.com/pricing)
- [Security](https://example.com/security)
- [Changelog](https://example.com/changelog)

That is it. No keyword stuffing. No instructions telling a model to recommend you. No giant copy of your whole website. The file should help a person or an agent reach the right source quickly.

The same principle applies to any machine-readable convention. sitemap.xml, structured data, RSS, Open Graph tags, and robots.txt are useful because they expose information cleanly. They are not substitutes for information worth exposing.

The Uncomfortable Answer

People want llms.txt to work because the alternative is inconvenient. You need to have a product that is clear, pages that are useful, and a reputation that exists off your own website. You need to look at the actual prompts and improve the specific thing they reveal.

That is why a tiny text file became a cottage industry. It gives you a checkbox when what you really need is a loop.

Create the file if you have documentation. Put it in your regular maintenance list. Then spend the rest of the afternoon making one page more useful to the person who is already trying to solve the problem your product solves.

That is still the best AI visibility strategy I have found.

Frequently Asked

Should I add llms.txt to my SaaS website?

Add it if you have substantial, maintained documentation or a complex public product surface. It is inexpensive documentation hygiene. Do not expect it to increase AI citations or rankings by itself.

Does ChatGPT read llms.txt?

There is no public commitment from OpenAI that ChatGPT Search uses llms.txt for retrieval or ranking. Public crawl studies show far fewer requests for it than for standard files such as robots.txt.

What improves AI search citations instead?

Clear product pages, original information, independent mentions, useful comparison content, and a repeated measurement loop are better bets. A file format does not replace those inputs.