What Is llms.txt? 3 Steps to Help AI Read Your Website in 2026

Over the past six months we have received quite a few questions like this from business owners: I heard I need an llms.txt file for ChatGPT to read my website, is that right, how long does it take, how much does it cost. The honest answer is that it takes about one afternoon, costs nothing if you write it yourself, and whether your website gets mentioned by AI depends on entirely different factors. This article explains what llms.txt is according to its original proposer, gives the latest numbers on whether the file is actually read, and then walks through 3 steps to create it. The last part explains what most Vietnamese articles are getting wrong: to block AI from taking your content, you have to write it somewhere else.

What llms.txt is and what is inside the file

llms.txt was published by Jeremy Howard, founder of Answer.AI, on September 3, 2024 as an open proposal, with the specification at llmstxt.org. The original idea is very simple and comes from a real pain for developers. A modern web page has hundreds of kilobytes of HTML, JavaScript, menus, banners, and popups, while the text actually worth reading is only a few percent. Large language models are limited by their context window, so reading a full HTML page burns tokens and returns mostly junk. So Jeremy Howard suggested that site owners clean up a lean version themselves, written in Markdown, so machines can grab exactly what they need.

Think of it as the table of contents of a catalogue. The catalogue is still there in full. The table of contents just tells you: pricing is on page 12, warranty on page 30, contact on the last page. A machine that reads the table of contents knows where to go directly, without flipping from the start.

The structure of the file, exactly as in the original specification

The specification sets the format quite strictly, and that is exactly why machines can read it without guessing. A valid llms.txt file has four parts in the following order.

• One line with the only H1, giving the business or project name. Only one, never a second.
• A professional a quote block (blockquote, starting with the >) summarizing what the business does, in about 2 to 5 sentences.
• The business's H2 sections, each grouping a set of links of the same kind, written as - [Page name](https://...): one-line description.
• A section named Optional at the end, containing less important pages, so that any machine with a tight context can skip this part first.

There is one detail that many Vietnamese tutorials get badly wrong, and we need to spell it out here because it is the root of every misunderstanding that follows. llms.txt has no User-agentsyntax, and no Allowsyntax, and no Disallow. Those three command lines belong to file robots.txt, which is completely different. Anyone who writes Disallow: /admin/ into llms.txt still ends up with a valid text file, but no bot understands it, and nothing gets blocked. You think you have locked the door, but in reality it is wide open.

llms-full.txt is a combined version for people who want to paste a whole website into a chat

Alongside llms.txt there is a secondary convention called llms-full.txt. This file merges all of a website's text into one Markdown file, much heavier, sometimes several hundred kilobytes to a few megabytes. Its real use is quite everyday: an engineer opens Claude or ChatGPT to ask about your product, pastes one llms-full.txt link into the chat box, and the AI has your whole documentation set to answer from, with no need to dig through page by page. Anthropic, Cloudflare, Vercel, and the documentation platform Mintlify all publish both files for their technical docs.

This is also the place to see clearly who really uses llms.txt. It was born in the technical documentation community, serving developers and coding tools. For sales websites, service websites, and store-chain websites, the benefit is much thinner, and you should know that before paying anyone to do it.

Open the mona.media llms.txt file and take a look

Talking in the abstract is hard to picture, so open https://mona.media/llms.txt and you will see it right away. We rechecked this file on August 5, 2026: 2,018 bytes, 28 lines, with one H1 line for the company name, a summary paragraph, then four sections: Main services, AI applications, Contact, and Resources. All of it fits on one screen. Nothing complicated.

And while reviewing it to write this article, we caught two wrong numbers in our own file, written back in June 2026: one place stated a headcount that MONA has removed from all outward-facing documents, and another wrongly said "14,000+ customers" when the correct figure is 14,000+ projects. We fixed both on August 5, 2026, and if you open the file now you will see the corrected version. We tell this story not to beat ourselves up but so you can see a very real risk: an llms.txt file is something written once and forgotten, yet it is the business summary that machines read first. One wrong number in it is wrong in the place most likely to be copied.

The hard truth: Google says llms.txt does not help rankings

We wrote this part against most articles circulating online, and put it before the how-to, so you have the facts before deciding whether to spend money.

On June 15, 2026, Google updated its documentation on optimizing for AI features, stating plainly that machine-readable files like llms.txt have no effect, good or bad, on Search rankings or on AI Overviews. Earlier, Google's John Mueller said something rather blunt when asked: to him, llms.txt is the equivalent of the old meta keywords tag, a place where site owners declare what they sell. Google dropped the meta keywords tag in 2009 for a reason anyone can guess: people lied.

97% of llms.txt files received not a single read in May 2026

Ahrefs ran a survey of 137,210 domains with real traffic in May 2026. The results are a little shocking.

• 28% of those domains have published an llms.txt file, about 38,360 websites.
• 97% of those files received exactly 0 visits during the whole of May 2026. Only about 1,100 domains saw anyone stop by.
• Among the small group that got visits, 96% visits were bots, and real people made up only 4%.
• Breaking that bot group down further: well-known AI tools accounted for 19.5% of visits, of which AI agents were 10.5%, training bots 5.3%, and answer-retrieval bots only 1.1%.
• Most reads came from SEO audit tools at 21.7%, general web crawlers at 13.1%, and technology detection tools at 11.6%.

Translating that table into plain language: most visits to your llms.txt file come from bots that score SEO, not from the AI composing answers for your customers. Another survey, by SE Ranking across 300,000 domains, found an adoption rate of 10.13%, and only 0.3% among the largest websites. After nearly two years of talk, none of the big four has confirmed that it uses the standard.

So do we recommend doing it

Yes, but at its true scale, and we explain why so that nobody can oversell you.

The first reason is that the cost is close to zero. One 2-kilobyte file, one afternoon, no programmer, nothing touched in the website code. The second is that it is genuinely useful when real people use it, for example a customer or partner pastes your llms.txt link into ChatGPT to ask quickly about your company, or your sales team loads the file into an internal assistant. The third is that writing the file forces you to sit down and answer two questions you have long been dodging: how would you describe your business in three sentences, and out of two hundred pages on your site, which twenty are most worth reading. Just answering those two questions is worth the effort.

And if anyone pitches you an "llms.txt declaration to reach the top of AI" package for tens of millions of VND, put the 1.1% figure above right in front of their quote. Whether your website gets cited by AI is decided by three other things: content that answers questions directly, having structured data declared correctly, and a website with strong enough trust signals. We split those three into three separate articles, arranged as a cluster for you to read in order: how to write answer-first content for AI to cite, how to build E-E-A-T in the AI era, and the set of metrics for measuring SEO as zero-click grows.

3 steps to declare llms.txt for a business website

We wrote the three steps below in the exact order you do them, so you can check as you go. Assume the website runs on WordPress, the hosting gives you file access, and you do not need to know how to code.

Step 1: Choose 20 to 40 pages worth a machine's time

This step takes the most time and also decides whether the file has any value. Open Google Search Console, go to Performance, sort by impressions over the last 12 months, and take about the top 40 pages. Remove pages that say nothing about your products: cart pages, login pages, thank-you pages, tag pages, and pagination pages like /page/2/.

Group the rest into 4 to 6 groups. For a multi-branch spa chain whose website we built, the four groups came out like this: services and treatments, price list, skincare knowledge, branch information and booking. For a real estate company, the groups would be projects on sale, legal procedures, area analysis, contact. The right way to group is by the questions customers often ask, not by the website's menu structure. The menu is how the company arranges itself. Questions are how customers think.

Step 2: Write the file in the standard format

Open Notepad on Windows or TextEdit on Mac (remember to switch TextEdit to plain text mode before typing), then copy the template below and replace it with your own content.

# ABC Company > ABC Company was founded in 2015 and distributes industrial kitchen equipment to > restaurants and hotels in southern Vietnam. It has 3 warehouses in Ho Chi Minh City, Can Tho and Da Nang. > On-site warranty within 24 hours. Hotline 1900 xxxx. ## Products - [Industrial stoves](https://abc.vn/bep-cong-nghiep/): High-capacity gas and induction stoves for restaurants, with specs and prices. - [Refrigerators and freezers](https://abc.vn/tu-mat/): Food storage units by capacity, with a power-consumption comparison table. ## Guides and knowledge - [Choosing stove power by number of tables](https://abc.vn/blog/chon-cong-suat-bep/): A power lookup table for 20, 50 and 100 tables. - [Kitchen equipment maintenance schedule](https://abc.vn/blog/bao-tri/): What to do weekly, monthly and quarterly. ## Policies - [Warranty and returns](https://abc.vn/bao-hanh/): Conditions, duration and how to submit a request. - [Delivery and installation](https://abc.vn/giao-hang/): Fees, timing and service areas. ## Optional - [About us](https://abc.vn/gioi-thieu/): Company history and team. - [Careers](https://abc.vn/tuyen-dung/): Open positions.

Three mistakes we see most often when reviewing customers' files. One is writing advertising-style descriptions, such as "reputable, high quality, good price"; a machine can pull no information from that, and you are better off leaving it blank. Two is stuffing in a hundred links to look impressive, which swells the file to tens of kilobytes and kills the filtering purpose the file was created for. Three is leaving in links that point to 404 pages or renamed pages, which happens a lot on websites that have been running for years.

Step 3: Put it in the root folder and check with three tests

Name the file exactly llms.txt, all lowercase, then upload it to the website's root folder, the same place as the robots.txt and index.phpfile. For WordPress on cPanel hosting, that is the public_htmlfolder. FileZilla, Cyberduck, or the hosting's file manager all work. Do not put the file in a subfolder, because machines only look in the root.

Then check three things, which takes about two minutes. First, open your browser at tenmien.com/llms.txtand you should see the text appear as plain text. If the browser downloads the file instead, the server is returning the wrong content type, so ask your hosting provider to set it to text/plain. If you use hosting Mona.Host just message their technical team, and they can fix it in a few minutes. Second, click each link in the file, and there must be no dead links. Third, set a reminder to review it once a quarter, on the same day you review the price list.

The same approach also applies to: fashion retail chains, construction material manufacturers and distributors, real estate developers, clinic systems, online training centers, and travel agencies.

To block AI from taking your content, write it in robots.txt, not in llms.txt

Here is the part we think is worth the most in the whole article, because it corrects the very misunderstanding that is giving many businesses a false sense of safety. Many people think that creating an llms.txt file protects their exclusive content. Not at all. llms.txt is an invitation and has no blocking power whatsoever.

The place to command a block is still robots.txt, and the major AI companies have all published their own bot names for site owners to target. We rechecked the list below on August 5, 2026.

• OpenAI: GPTBot (collects data for training), ChatGPT-User (when a user clicks a link in the chat), OAI-SearchBot (indexes pages for search in ChatGPT).
• Anthropic: ClaudeBot, Claude-User, Claude-SearchBot.
• Google: Google-Extended, used only for training Gemini and Vertex AI.
• Other operators: PerplexityBot, CCBot from Common Crawl, Bytespider from ByteDance, Amazonbot, Applebot-Extended, meta-externalagent from Meta.

To forbid a bot from reading an internal documents folder, add exactly three lines to robots.txt , for example User-agent: GPTBot then a new line Disallow: /tai-lieu-doi-tac/. Repeat for each bot you want to block. This is the real syntax, and it only works in the robots.txt file.

Google-Extended blocks training, but does not remove pages from AI Overviews

This spot traps many people, so we will go through it carefully. Blocking Google-Extended only stops Google from using your content to train Gemini models. It does no remove your pages from AI Overviews, because AI Overviews draws on the regular Search index, which is crawled by Googlebot . To keep a page from being cited in AI Overviews, you have to use the nosnippet or max-snippettag, but doing so also loses the description snippet in regular search results. And blocking Googlebot outright makes the page vanish from Google. As of August 5, 2026, there is no switch that separates the two.

Does blocking everything cost you customers

We do not advise blocking everything, and this is a position we are willing to commit to. Blocking all AI bots means that when a customer asks ChatGPT "which company in Ho Chi Minh City does X", your name will certainly not be in the answer, while a competitor who did not block will be. What do you get in return? Your public marketing content stays out of training sets, which in practice has long since been copied everywhere anyway.

What we usually advise is selective blocking, just three groups. Group one is folders of documents meant for partners, internal price lists, and business forms. Group two is areas that hold customers' personal data, even if you think they are already hidden. Group three is assets that earn money by themselves, such as a training center's course materials or a spa chain's before-and-after photo sets. Everything else, meaning blog posts, service pages, and about pages, stays wide open, and the more machines read it the better. That is where you want to be named.

Which pages each industry should put in llms.txt

The same template applies, but which pages go in differs by industry. Below is how we usually choose when working for customers.

• E-commerce. Include the main category pages, return policy pages, and guides to choosing sizes or specifications. Do not include individual product pages, or you run out of room immediately. Supporting platform: e-commerce websites and MONA eCommerce.
• Real estate. Include the page for each project on sale, legal procedure pages, and the price and progress pages. Leave out internal market analysis entirely if it is what your sales team uses to close deals. See also real estate management software.
• Education and training. Include program overview pages, tuition, instructor profiles, and start dates. Keep course materials and exams out, and block them in robots.txt as well. Supporting platform: LMS software with built-in AI from Mona.Software.
• Restaurants and hotels. Include menus, prices, opening hours, each branch's address, and the booking policy. This group benefits most clearly because customers often ask machines exactly these things. Supporting platform: restaurant management software.
• Spas, salons, and clinics. Include the service list, prices, the process of a treatment session, and the refund policy. Block the before-and-after photo sets and customer records. Supporting platform: spa management software.

One thing in common for every industry: any page you put in llms.txt must answer the question directly in its opening paragraph. Put a meandering page into a table-of-contents file and the machine still gets nothing from it. How to write it properly is covered in the article on writing answer-first content, and for how bots read the web, see also what crawl is and What is Googlebot.

What you can do this week

To sum up in three jobs, all done in one afternoon.

Job one, open tenmien.com/llms.txt and tenmien.com/robots.txt and read it. Just read, change nothing yet. If your llms.txt file contains the word Disallow then you have just found a hole you thought you had plugged. Job two, take the 40 pages with the most impressions from Google Search Console, group them by the questions customers ask, and write the file in the template above. Job three, decide firmly which three groups of content to block in robots.txt, and only those three.

The MONA Cloud and MONA Pay ecosystem gathers all its documentation into a single llms.txt file

Type one command to create a VPS into the Claude Code window, and in under 11 seconds a brand new server in Vietnam is ready to run at a cost of only 550 VND per hour. The AI assistant knows which API to call and what it costs, without you having to open the documentation and look it up. The secret is that every product in the MONA Cloud ecosystem, from the MONA Mail email system and MONA Domain registration to the bank API and MONA Pay automatic payment confirmation service, comes with an llms.txt file prepared by us. One file is enough.

Picture it this way: instead of making the bot wade through dozens of links, the MONA Cloud llms.txt lists each page clearly, with a plain-text version of each. Inside is everything from the hourly price list and wallet and top-up documentation to guides for migrating from foreign platforms. Likewise, if you open the MONA Pay file, the system has already prepared the list of supported banks, the API for AI agents, and how to generate accurate VietQR codes. Unlike international platforms that host servers in Singapore with latency of 30 to 50ms measured from VNPT on September 19, 2026, our infrastructure sits in Vietnam and delivers response times of 5 to 15ms. Apps run smoothly, payment is in VND, and most importantly your accounting department always has VAT invoices to book transparently every month.

Behind that smoothness are hard lessons paid for out of our own pocket. In mid-September 2026, while testing the AI calling the wallet top-up feature on MONA Cloud, the bot successfully printed a VietQR code right in the chat window for the user to simply scan. It seemed everything was smooth, but on September 15, two internal products had mismatched webhook keys, so customers transferred money and the wallet was not credited. We had to stay up through the night and patch it immediately. By September 17, every transaction returned the correct 200 status and money was credited accurately to the last dong. We took this bug before our customers did. Since then, MONA's internal rule requires every payment-notification key to be cross-checked with a real transaction before the system opens to businesses. AI agents now call directly through the MCP platform, backed by the MONA engineering team that, since 2016, has been through more than 14,000 projects with 85% of customers returning. Your application will always run stably.

If you have personal projects that you only want to test for a few days, with no real users yet, do not rush to renting a server and then pay for maintenance. Instead of spending money early, use the free simulation environment we built into MONA Cloud to test first. Only when your business needs an app with an admin system, a database, and long-running tasks that need source code security and safe hosting in Vietnam should you bring the app to run on MONA's official infrastructure.

A manifesto for growing together

MONA is a software, digital transformation and AI transformation company, but above all, MONA is a B2B service company that benefits directly from this: THE MORE SUCCESSFUL VIETNAMESE BUSINESSES BECOME, THE MORE MONA BENEFITS, while a declining, slow market with weak technology is exactly what kills MONA's future business opportunities. So, more than anyone, MONA wants you to succeed, and MONA actually steps in to help you succeed.

And AI transformation is the key to that success for all of us.

MONA

Helping machines automate the entire technical infrastructure of a business

Instead of having staff spend hours digging through technical documentation, you can fully let AI read the llms.txt file and act on its own. MONA Cloud bills by the hour from 550 VND, with a free sandbox for AI to try first. Go to https://monacloud.vn/ and try it first, which takes about 5 minutes. One click and you are done.

Let the AI try a free sandbox deploy Call 1900 636 648

FAQ

How is llms.txt different from robots.txt?

robots.txt is a command file. It uses User-agent, Allow, and Disallow syntax to allow or forbid each crawler from each folder. llms.txt is an invitation file, written in Markdown, that only lists the pages worth reading with a short description so machines can pick up the right parts. llms.txt has no Disallow syntax and cannot block any bot.

Will a website with llms.txt reach the top of Google faster?

No. On June 15, 2026, Google updated its guidance documentation and stated clearly that llms.txt has no effect on Search rankings or on AI Overviews. Google's John Mueller compared this file to the meta keywords tag, meaning a place where site owners describe themselves but which search engines do not use for ranking.

How many steps does it take to create an llms.txt file?

Three steps. Step 1, choose 20 to 40 pages worth a machine's time, taken from the Performance report in Google Search Console, then group them into 4 to 6 groups by the questions customers often ask. Step 2, write the Markdown file in the standard format: one H1 line with the company name, a short summary paragraph, H2 sections containing link lists with a one-line description each, and an Optional section at the end. Step 3, upload the file to the website root, in the same place as robots.txt, open it in a browser, click through every link, and schedule a review each quarter.

How do I stop AI from taking my website content?

Write blocking rules in robots.txt, targeting each company's bot by name: GPTBot, ChatGPT-User, and OAI-SearchBot from OpenAI; ClaudeBot, Claude-User, and Claude-SearchBot from Anthropic; Google-Extended from Google; PerplexityBot, CCBot, Bytespider, Amazonbot, Applebot-Extended, and meta-externalagent from the others. Note that blocking Google-Extended only stops Google from training Gemini. It does not remove your pages from AI Overviews.

What is llms-full.txt, and do I need it?

llms-full.txt merges all the text of a website into a single Markdown file, so a user can paste one link and load the whole content library into ChatGPT or Claude. It is useful for technical documentation sites like Anthropic, Cloudflare, and Vercel. Most sales and service websites do not need it, and a lean llms.txt is enough.