Somewhere in your inbox right now, there’s probably a pitch promising that llms.txt is the missing piece of your AI search visibility. A vendor will install it. An agency will “optimize” it. A LinkedIn post will tell you the sites that publish one are quietly winning ChatGPT citations while everyone else sleeps.
I want to give you a straighter answer than that, because I publish an llms.txt file on my own site — and I still think it’s one of the most oversold artifacts in AI-search marketing. Both of those things are true at the same time. Understanding why is worth ten minutes of your attention before you spend a dollar on it.
What is llms.txt?
llms.txt is a proposed community convention: a markdown file placed at the root of your website that lists your key pages and content in a format designed for AI systems to read. Think of it as a curated map of your site for language models — robots.txt-adjacent in spirit, but offering a table of contents rather than crawl rules.
That word “proposed” is load-bearing. llms.txt came out of the developer community, not out of Google, OpenAI, or Anthropic. Some sites have adopted it. Some tools generate it automatically. But no major search engine has announced that it uses llms.txt as an input to ranking, and no major AI assistant has committed to it as the way they understand websites. It exists in the same category as many web conventions before it: a reasonable idea, published openly, waiting to see whether the platforms that matter ever adopt it.
The file itself is almost disappointingly simple. A heading with your site name. A short description. A list of links to your most important pages, each with a one-line summary, written in plain markdown. That’s the whole artifact. Anyone telling you it requires specialized expertise to produce is telling you something about their pricing model, not about the file.
Does llms.txt affect your Google rankings?
No. Nothing in Google’s published guidance for Search — including its guidance for AI features in Search — gives llms.txt any role. Google’s documentation on appearing in AI experiences points at the same fundamentals it has always pointed at: content that can be crawled, content that can be indexed, and content that is genuinely useful and original. It announces no special file, no secret markup, no side door that unlocks AI visibility.
I want to say this kindly, because a lot of smart people have been told otherwise: the emperor has no clothes here. When Google describes what earns visibility in its AI features, it describes ordinary search fundamentals. If a proprietary file at your site root were part of the equation, Google would say so — the company publishes extensive documentation about exactly what it reads and how. The absence is the answer.
This doesn’t make llms.txt worthless. It makes it what it is: a courtesy convention with no announced ranking function. The problem isn’t the file. The problem is the sales pitch wrapped around it, which converts “a community proposal some crawlers might read” into “the lever that determines whether AI recommends you.” Those are very different claims, and only one of them is supported.
Why do vendors keep selling it?
Because a checkbox artifact is easy to sell. It’s visible, it’s deliverable, and it fits in a screenshot. A vendor can install llms.txt on Tuesday, send you the URL on Wednesday, and mark the line item complete. You can see it. Your board can see it. Everyone gets the satisfying feeling of having Done Something About AI.
Compare that with the work that actually moves AI search visibility. Restructuring your content so it answers real buyer questions directly, in the first two sentences, instead of warming up for four paragraphs. Making sure your company’s basic facts — what you do, who you serve, where you operate — are stated consistently across your site, your directories, and the third-party sources assistants draw on. Earning mentions from publications and communities that AI systems treat as credible. None of that ships in a day. None of it screenshots well. All of it compounds.
This is the pattern I’d want you to recognize, because it’s industry agnostic and it predates AI by decades. When a new channel emerges, the first wave of products sold into it are artifacts, not outcomes. Artifacts are legible to buyers who don’t yet have a mental model for the channel. If you can’t yet evaluate whether your content answers buyer questions well, you can absolutely evaluate whether a file exists at yoursite.com/llms.txt. So that’s what gets sold.
The tell is proportion. A vendor who mentions llms.txt as one small item in a larger program built on content and structure is being honest with you. A vendor whose pitch leads with llms.txt is selling folklore, and you should read the rest of the proposal with that in mind.
Why do we publish one anyway?
Because at the right price, a courtesy to machines is worth extending. Here’s the honest cost-benefit, first-principles style.
The cost side: our llms.txt is generated automatically from real content at build time. Every time the site rebuilds, the file regenerates from the pages that actually exist. Zero marginal maintenance. Nobody updates it by hand, nobody bills hours against it, nobody thinks about it. If the convention dies quietly, we’ve lost nothing.
The benefit side has three parts. First, some LLM crawlers do fetch the file today, and when they do, they get a clean, curated map instead of having to infer site structure from navigation menus and footer links. That’s a small legibility win. Second, if the convention gets adopted more broadly — if ChatGPT, Gemini, or Claude ever announce they use it to understand sites — we’re already there, at no additional cost. That’s optionality, purchased for free. Third, the exercise of generating it forces a small useful discipline: deciding which pages actually represent the business. If you can’t produce a curated list of your ten most important pages, that’s a signal about your content, not about the file format.
Notice what’s missing from that list: any claim that the file affects rankings, citations, or visibility. I publish llms.txt the way I’d leave a well-labeled directory at the front desk of an office building. Most visitors won’t read it. The ones who do will have a slightly easier time. It costs me nothing to keep current. That’s the whole business case — and priced accordingly, it clears the bar.
The point is proportion. llms.txt is a cheap courtesy, not a strategy. The strategy is the content itself.
What should you spend the effort on instead?
Spend it on the fundamentals that AI assistants demonstrably lean on when they decide what to surface and cite — the real dividing line between AI-search strategy and conventional SEO. There are four, and none of them are glamorous.
Content that answers real buyer questions directly. AI assistants assemble answers, which means they favor sources that contain answers — stated plainly, near the top, in language a machine can lift cleanly. Take the questions your prospects actually ask in sales calls and make each one a page or section where the first two sentences resolve the question. Most B2B sites do the opposite: they bury the answer under positioning language. If a founder asked where to put the next quarter of content effort, this is it, because it serves human buyers and AI systems with the same work.
Consistent entity facts. Assistants build a model of who you are from everything they can read: your site, your LinkedIn page, industry directories, press mentions. When those sources disagree — different descriptions of what you do, different service categories, stale headcount or locations — the machine’s picture of you blurs, and blurry entities get recommended less confidently. Audit the basic facts about your business everywhere they appear and make them agree. Tedious, unglamorous, and more valuable than any file at your site root.
Structure machines can parse. Real headings that describe what each section contains. Lists where content is genuinely list-shaped. Schema markup where it fits. Pages organized so a system skimming them can tell what’s being claimed and by whom. This is the same architecture that has served search for years; AI retrieval raises the stakes on it rather than replacing it.
Sources worth citing. Assistants corroborate. A claim that exists only on your own site is a claim; the same claim echoed by third parties — trade publications, industry communities, review platforms, partners — starts to look like a fact. Earning that corroboration is slow relationship work, which is exactly why it’s defensible and exactly why nobody sells it as a checkbox.
Here’s the test to apply to any AI-visibility proposal that crosses your desk: does the effort improve what your content says and how legibly it says it, or does it only add an artifact alongside the content? The first category compounds. The second category is llms.txt — fine at zero cost, foolish at any real price.
If you’re getting pitched llms.txt as a strategy, what you’re really being told is that your vendor hasn’t done the harder thinking about how AI systems will find, understand, and recommend your business. That thinking starts from first principles: what do these systems actually read, what makes an entity legible to them, and which of your existing assets already do that work. Thirty days spent understanding how your buyers ask questions and how machines retrieve answers is worth more than thirty minutes installing a file — because the first one builds a revenue engine, and the second one builds a screenshot. A senior partner who can tell you which is which, in your business, with your budget, is where that conversation begins.