If you want to write an `llms.txt`, choose a short Markdown file in the root of the domain, as AI models and tools parse simple titles, descriptions, H2 sections, and link lists most easily. In practice, check: H1, description in blockquote, 4-6 sections, absolute links, short descriptions, `/llms.txt`, sitemap, and no blockage in robots.txt.
Template to copy: "If my site has a [store/services/blog], I create /llms.txt with [name, description, main pages, offer, usage rules] because [AI needs a compressed site map]. In practice, I check [file address, length, links, descriptions, robots.txt, sitemap]."
A client asks Perplexity: "Which dog food store offers pickup in Krakow?" The agent can access your site, but if they have to navigate through the menu, cookie banner, slider, and 40 links in the footer, they start guessing. llms.txt is like a note at the entrance: here is the store, here are the most important departments, here are the delivery rules, and here is the contact.
Why This Matters in 2026
llms.txt is not an older sibling to robots.txt. It is a different file with a different purpose. robots.txt tells bots where they can go. llms.txt tells models and agents what is most important on the site and where to find the readable version.
The specification described at llmstxt.org was published by Jeremy Howard on September 3, 2024. Its premise is simple: models have a limited context window, and a complete HTML page has too much navigation, advertising, scripts, and repetitive elements.
In 2026, this issue is practical for small businesses. AI Search and shopping agents no longer need to display 10 blue links. They increasingly attempt to respond immediately: "Choose this company because it has X, Y, and Z." If your most important information is buried in HTML, the model may not select it.
At Audit AI, llms.txt is part of the Discovery category. We check if the file exists and if it has at least 100 characters of useful content. This is a low technical threshold, but a good starting point: without the file, the agent does not receive any condensed instructions.
How This Differs from Classical SEO
SEO says: "build a sitemap, meta title, and content for queries." llms.txt says: "give AI a short, current map of the most important content, preferably in Markdown."
SEO Example: The category "women's trekking shoes" has a 2500-character description, H1, meta description, and internal linking. llms.txt Example: one line leads to this category and states that it is the main collection of trekking shoes with filters for size, membrane, and season.
This does not replace schema.org, Product schema, or good descriptions. It is a table of contents for AI. If the table of contents is clear, the agent gets to the right page faster and has fewer reasons to guess.
Step by Step: Structure That's Easy to Parse
- Start with One H1 with the Name of the SiteBad"Welcome to our online store."Better`# ZooKarma.pl - a store for dog and cat food and accessories`.
The first line should immediately state what type of site it is. Do not write a marketing slogan. Write the name and function of the company.
For a cosmetics store: `# NaturalnaPolka.pl - a store for natural cosmetics`. For a physiotherapy clinic: `# FizjoRuch Poznań - physiotherapy and sports rehabilitation`.
# YourCompany.pl - short description of the site type
This H1 is the only section required by the specification, but an H1 alone is insufficient for a good file.
- Add a Blockquote with a Description in 1-2 SentencesBad"We are a dynamic company offering high-quality services."Better`> The store sells veterinary food, walking accessories, and beds for dogs. It offers shipping in Poland and in-store pickup in Krakow.`
The blockquote under H1 is the place for a brief summary. The model should know within 10 seconds whether the site is a store, clinic, blog, SaaS, or local service.
For a dentist, write specifically: city, scope of services, form of contact. For a store: type of products, market, delivery, pickup, most important constraint.
- Use H2 Sections as DrawersBadone large block of text with 30 links.Better4-6 H2 sections: `About`, `Main pages`, `Products`, `Policies`, `AI usage notes`, `Optional`.
The `llms.txt` specification assumes H2 sections with lists of links. Tools can then easily recognize which links are documentation, which are products, and which are additional.
In a furniture store, sections might look like this: `Main pages`, `Product categories`, `Buying guides`, `Policies`, `Optional`. In a beauty salon: `Services`, `Booking`, `Prices`, `Policies`, `Optional`.
## Main pages - [Home](https://example.pl/): main page and most important categories - [Contact](https://example.pl/kontakt): company details, phone number, and form
Each link should have a short description after the colon. Without a description, the model knows the address, but does not understand the reason for going there.
- Link Only to Pages That Truly HelpBaddropping the entire menu, all blog tags, and 200 products.Better10-25 links to pages that answer real customer questions.
`llms.txt` is not a sitemap. A sitemap can have 5000 URLs. `llms.txt` should be a selection. If the list is too long, the model has to decide for you again.
For a shoe store, select: the homepage, 5 categories, a size chart, delivery, returns, contact, 2 guides. For an accountant: services, pricing, supported forms of business, contact, privacy policy, FAQ.
- Add `AI usage notes`, but Don’t Pretend It’s a Legal AgreementBad"AI can use everything without restrictions."Better"Public product descriptions can be summarized. Prices and availability should be verified on product pages. Content from the customer panel is not public."
This section should provide guidelines, not replace a licensing agreement. Don’t include things here that you technically cannot enforce. If you block model training in robots.txt, say so consistently in `llms.txt` as well.
Example for a shop:
## AI usage notes - Public product descriptions may be summarized by AI assistants. - Prices, stock, and delivery dates should be verified on product pages. - Customer account pages, checkout, and payment pages are not public content.
Example for a clinic:
## AI usage notes - Public service pages may be summarized. - The website does not provide medical diagnosis. - Appointment availability must be confirmed through the booking form.
- Use the `Optional` Section for ExtrasBadeverything has the same priority.Bettermain pages are higher, and the `Optional` section contains the blog, old guides, social media, and supplementary materials.
In the specification, `Optional` has special significance: links in this section may be skipped when the tool builds a shorter context. This is useful because models have limits. You don’t want the section detailing the company history to push delivery rules or pricing out.
For a pet food store, add a blog about dog care to `Optional`. For a beauty salon: a gallery of implementations and helpful articles. Keep important transactional links higher up.
- Keep the File Short and Up-to-DateBad`llms.txt` is 6000 words because "the more, the better".Better700-1500 words, 10-25 links, updated once a month or after major changes in the offer.
Audit AI's own template suggests keeping the file under 1500 words. This is not a hard limit of the specification, but a good limit for small businesses. A short file is easier to maintain and easier to paste into an AI tool as context.
If you have a large documentation site, you may consider `llms-full.txt`. For a small store, a good `/llms.txt` and a normal sitemap.xml usually suffice.
Ready-Made Template for E-commerce Store
Copy this template and replace the domain, description, and links.
# ZooKarma.pl - a store for dog and cat food and accessories > The store sells veterinary food, walking accessories, and beds. It offers shipping in Poland and in-store pickup in Krakow. ## About - Domain: https://zookarma.pl - Type: e-commerce - Main language: pl - Market: Poland ## Main pages - [Home](https://zookarma.pl/): main page of the store - [Contact](https://zookarma.pl/kontakt): company details, phone number, and form - [About](https://zookarma.pl/o-nas): information about the company and physical store ## Product categories - [Dog Food](https://zookarma.pl/karmy-dla-psow): main category of dry and wet food - [Veterinary Food](https://zookarma.pl/karmy-weterynaryjne): products requiring veterinary recommendations - [Walking Accessories](https://zookarma.pl/akcesoria-spacerowe): leashes, harnesses, and collars ## Policies - [Delivery](https://zookarma.pl/dostawa): costs, shipping time, and in-store pickup - [Returns](https://zookarma.pl/zwroty): return and complaint rules - [Privacy policy](https://zookarma.pl/polityka-prywatnosci): information on data processing ## AI usage notes - Public product and category pages may be summarized by AI assistants. - Prices, stock, and delivery dates should be verified on product pages. - Checkout, payment, and customer account pages are not public content. ## Optional - [Blog](https://zookarma.pl/blog): guides on feeding and caring for pets
Service Company Variant
# FizjoRuch Poznań - physiotherapy and sports rehabilitation > The clinic offers physiotherapy for adults, rehabilitation after injuries, and spinal pain therapy in Poznań. Appointments can be booked online or by phone. ## Main pages - [Home](https://fizjoruch.pl/): description of the clinic and main services - [Services](https://fizjoruch.pl/uslugi): list of physiotherapy services - [Prices](https://fizjoruch.pl/cennik): prices for visits and packages - [Booking](https://fizjoruch.pl/rezerwacja): appointment booking form - [Contact](https://fizjoruch.pl/kontakt): address, phone number, and hours of operation ## Services - [Back Pain](https://fizjoruch.pl/uslugi/bol-kregoslupa): therapy for back and neck pain - [Sports Rehabilitation](https://fizjoruch.pl/uslugi/rehabilitacja-sportowa): return to training after an injury ## AI usage notes - Public service pages may be summarized. - The website does not provide medical diagnosis. - Appointment availability must be confirmed through the booking page.
Table: What to Add and What Not to Add
| Element | Add to llms.txt | Do Not Add | | ---------- | --------------------------------- | ------------------------------- | | Homepage | yes | duplicates with UTM parameters | | Categories | 5-10 most important ones | all filters and sorting options | | Products | only bestsellers or landing pages | hundreds of product cards | | Policies | delivery, returns, privacy | outdated regulations | | Blog | best guides | tags, archives, pagination |
Implementation Checklist
- The file is accessible at `https://yourdomain.pl/llms.txt`.
- The first line is one H1 with the name and type of the site.
- Under H1 is a blockquote with a description in 1-2 sentences.
- The file has 4-6 H2 sections.
- Links are absolute, with the full `https://`.
- Each important link has a description after the colon.
- The list contains a maximum of 25 most important links.
- The `Optional` section contains only second-priority content.
- The file does not contain slogans without specifics.
- The file does not contain prices if prices change frequently.
- The file refers to the product page, where the price and availability are current.
- The privacy policy, delivery, and returns are linked.
- `robots.txt` does not block `/llms.txt`.
- sitemap.xml still exists and is not replaced by `llms.txt`.
- The file has more than 100 characters of useful content.
- The file has less than approximately 1500 words.
- After implementation, the HTTP response code 200 is checked.
- After any major changes in the offer, the file goes to the update list.
7-Day Mini Plan
list 10 questions a customer might ask AI about your company.
select 10-25 URLs that best answer these questions.
write H1, description in blockquote, and `About` section.
add `Main pages`, `Products` or `Services`, `Policies`, and `AI usage notes` sections.
upload the file as `/llms.txt` in the root domain or ask hosting/developer for upload.
check the address in a browser, HTTP code 200, links, and no blockage in robots.txt.
run an audit on [auditai.cc](https://auditai.cc) and improve the Discovery score if the file is too short or inaccessible.
Common Mistakes
User-agent: * Allow: / Disallow: /cart
This is the `robots.txt` syntax, not `llms.txt`. In `llms.txt`, use Markdown: H1, blockquote, H2, and list of links.
- [Contact](/contact): company details
Better:
- [Contact](https://yourdomain.pl/contact): company details, phone number, and form
How to Measure Effects
First, measure technicalities. Does /llms.txt return HTTP 200? Does Audit AI pass checkpoints llms.txt exists and llms.txt has content? Is the file visible without logging in?
Then measure quality. Paste the link to the file into ChatGPT, Claude, or Gemini and ask 5 questions: what do we sell, who is the offer for, where is the contact, what is the return policy, which pages are the most important. If the model responds inaccurately, improve the descriptions of the links.
Finally, observe business signals: more branded queries, AI client inquiries with specific service names, visits to policy and guide pages, fewer erroneous responses about the offer.
Who llms.txt Is Not Enough For
If you have a store with thousands of products and frequent price changes, llms.txt will not replace a product feed, Product schema, and current inventory status. If you have a SaaS application with an API, you also need OpenAPI documentation, MCP, or WebMCP.
If the site blocks crawlers in WAF, has SSL errors, or generates content solely through JavaScript, llms.txt alone won’t fix visibility. It is a map. It won’t fix the road if it’s closed.
FAQ
Does llms.txt guarantee that ChatGPT will quote me?
Does llms.txt replace robots.txt?
Do I need llms-full.txt?
Should the file be in Polish or English?
Summary
A good llms.txt is not a lengthy manifesto. It’s a short map: company name, description, most important links, usage rules, and an optional section. Start with 10-25 links, stick to simple Markdown, and after implementation, check the file on auditai.cc before adding more protocols.
