Fixing Thin llms.txt Content: Notes a Language Model Can Use
A thin llms.txt file is one that parses correctly but tells a reader almost nothing: a site name, a row of bare links, perhaps a note that repeats each page title or promises "everything you need to know".
A language model or AI tool that opens it learns which URLs exist, which it could have found anyway, but not what is on them or which to read first. This guide shows how to rewrite the summary, the notes and the grouping so the file answers the questions a newcomer to your site would ask, using plain facts rather than marketing.
What a thin file looks like
The llms.txt Auditor scores the quality of the content as well as the structure, and a file can pass every format check while still scoring poorly. The weak spots tend to be the same from site to site. If you recognise two or more of these in your own file, this guide is the right one.
| Symptom | Example | Why it is weak |
|---|---|---|
| Bare links | - [Services](https://www.example.co.uk/services/) | The link text is the only clue. A reader has to fetch the page to learn anything. |
| Notes that echo the title | - [Pricing](…): our pricing page | Adds characters but no information. |
| Sales copy | : award-winning, unbeatable, trusted by thousands | Unverifiable claims a careful model will discount, and which say nothing about the page. |
| Empty or vague summary | > Welcome to our website! | The one place meant to explain the whole site says nothing about it. |
| One long section | ## Pages followed by sixty links | No signal about which pages matter or how they relate. |
| Sitemap dump | Every post, tag and product listed automatically | The file becomes longer than it is useful, and the key pages are lost in it. |
| Missing essentials | No link to prices, delivery, contact or policies | The pages people most often ask an assistant about are the ones left out. |
None of these is an error in the Markdown sense, which is why automatic generators produce them so often. A generator can list URLs and copy titles. It cannot tell a reader that your quotes are free within 20 miles of Leeds or that your software stopped supporting version 2 last year. That knowledge sits with you, and writing it down is the whole job.
Who actually reads the notes
It helps to picture the reader before writing for it. There are three realistic ones, and none of them is a search engine ranking algorithm.
- An AI tool pointed at your site. Someone pastes your address into an assistant or
research tool and asks a question. Some tools look for
/llms.txtfirst because a short Markdown file is quicker to work with than a set of rendered pages. The tool uses your notes to decide which one or two links to open. - A coding assistant or agent. Developers point these at documentation sites. A note that says which page covers authentication, and for which version, saves the agent from reading the wrong page.
- A person. Developers, journalists and curious customers do open llms.txt files. A clear file reads like a good site map with commentary.
What these readers have in common is limited attention. A model has a finite context window, and a tool that opens every link wastes time and money. Your notes are a routing aid: they let the reader pick the right page without opening the wrong ones. That is the test to apply to every line you write: does this help someone choose between pages?
Be realistic about what the file does. llms.txt is a proposal published by Jeremy Howard in 2024, not an agreed standard, and as far as we know none of the major AI providers has confirmed that its search or answer products read it. Better notes make the file more useful to the tools that do read it. They will not, on their own evidence, change how often an assistant cites you.
Writing the summary blockquote
The blockquote under the H1 is the only part of the file, apart from the name, that the proposal expects every reader to take in. It should let someone who has never heard of you answer four questions: what is this, who is it for, where does it operate, and what should I know before reading further.
Before
# Brightwater Dental
> Welcome to Brightwater Dental, where your smile is our passion!After
# Brightwater Dental
> Independent dental practice in Harrogate, North Yorkshire, offering
> private general dentistry, hygiene appointments, Invisalign and
> implants. We do not currently take NHS patients. New patients book
> online or by phone; emergency slots are released each morning.The rewrite contains nothing clever. It states the type of business, the place, the main services, one important limit and how to get started. The line about NHS patients is exactly the kind of detail an assistant would otherwise get wrong, because it is a common question and the answer varies from practice to practice.
A few rules keep the summary sharp:
- Two to four sentences. If you need more, the extra belongs in the details paragraph below it.
- Name the place. Town and county for a local business, country or region for anyone else. "UK-wide" is a fact; "nationwide" is ambiguous.
- State limits as well as offers. What you do not do is often more useful than another adjective about what you do.
- Avoid anything that dates quickly. A summer offer or this month's opening hours will be wrong soon. Link to the page that holds them instead.
Writing notes that help a reader choose
Each link can be followed by a colon and a note. The proposal leaves the wording to you, which is where most files go wrong. A reliable pattern is to say what the page covers, then add one or two specifics a reader might need, then, where it matters, the scope of the page.
| Weak note | Useful note |
|---|---|
| Our delivery information | UK mainland delivery times and charges, Highlands and Northern Ireland surcharges, how to track an order |
| Find out about our prices | Fixed prices for boiler servicing and gas safety certificates; repair work quoted after a visit |
| API docs | REST API reference for version 3: authentication with API keys, rate limits, pagination, error codes |
| Learn about us | Who runs the business, when it was founded, qualifications and memberships, registered company details |
| Amazing gallery of our work! | Photos of finished loft conversions in Sheffield and Rotherham, each with the room size and build time |
Notice what the useful notes avoid. They do not repeat the link text. They do not rate the page. They do not include a figure that will change next month; "fixed prices for boiler servicing" tells the reader where to look without printing a price that could go stale in the file while the page is updated.
Keep notes to a single line. Around 8 to 25 words is a comfortable range: long enough to carry a specific, short enough that a list of thirty links is still readable. Commas or semicolons are fine for listing the topics a page covers.
The fix, step by step
Step 1
Read the file as a stranger would
Open
your live /llms.txt in a browser, not the copy on your computer, and read it without looking
at the site. Write down every question you cannot answer from the file alone: where you operate, what
you charge for, how to buy or book, what you do not offer. Then run the llms.txt Auditor
and note which content findings it raises, so you know which gaps it has already spotted and which only
a human reader would notice.
Step 2
Rewrite the summary blockquote
Replace the text under the H1 with two to four plain sentences saying what the business or project is, who it serves, where it operates and the one or two limits people most often get wrong. Remove greetings, slogans and anything with a date attached. If there are further facts that change how the rest of the site should be read, such as prices excluding VAT or documentation covering one version only, put them in a short paragraph below the blockquote.
Step 3
Give every link a factual note
Go through the links one by one with each page open in a tab. After the link, add a colon and a single line naming the topics the page covers and one or two specifics a reader might need, such as the area served, the product version or the type of prices given. Delete words like best, leading and amazing. If you cannot write a useful note for a page, question whether it belongs in the file at all.
Step 4
Regroup the links by what readers need
Replace catch-all headings such as Pages or Links with H2 sections named after the tasks people come to the site for, for example Services, Buying and delivery, Help and support, or About the business. Put the most important section first and the most important link first within each section. Keep to roughly three to six sections; a section with a single link can usually be merged into its neighbour.
Step 5
Move secondary pages under Optional
Add
an ## Optional section at the end and move into it the pages a reader can safely skip when
short of space: galleries, case studies, older articles, press coverage and background reading. The proposal
describes this section as the place for material that can be left out, so tools that respect it will
read your core pages first. Delete, rather than move, pages that are thin, duplicated, out of date or
no longer relevant to customers.
Step 6
Check each note against its live page
For every note, confirm the page really says what the note claims: the same services, the
same areas, the same limits and the same version numbers. Correct the note, or the page, wherever they
disagree. Save the file as UTF-8, upload it to the site root, open /llms.txt in a private
browser window to confirm the new version is live rather than cached, and re-run the llms.txt
Auditor to check the content findings have cleared.
A full before and after
Here is a thin file of the kind generators and quick manual attempts produce, for a fictional pet grooming business.
# Pennine Paws
> The best dog groomers around!
## Pages
- [Home](https://www.example.co.uk/)
- [Services](https://www.example.co.uk/services/)
- [Prices](https://www.example.co.uk/prices/): our prices
- [Gallery](https://www.example.co.uk/gallery/)
- [Blog](https://www.example.co.uk/blog/)
- [Contact](https://www.example.co.uk/contact/)And the same file rewritten using the steps above. The page list barely changes; the information around it does.
# Pennine Paws
> Dog grooming salon in Hebden Bridge, West Yorkshire, for all breeds
> and sizes. Appointments only; we do not groom cats. A mobile van
> covers Calderdale on Tuesdays and Thursdays.
Prices depend on breed, coat condition and size, and are confirmed at
drop-off. Puppies must have had their first vaccinations before a visit.
## Grooming services
- [Full groom](https://www.example.co.uk/services/full-groom/): bath, dry, clip or hand strip, nails, ears; typical appointment lengths by size
- [Puppy introduction](https://www.example.co.uk/services/puppy/): short first visits for dogs under six months, what happens and what to bring
- [Mobile grooming](https://www.example.co.uk/services/mobile/): van service areas in Calderdale, days available, power and parking needs
## Booking and prices
- [Price guide](https://www.example.co.uk/prices/): starting prices by breed group, extra charges for matted coats and nervous dogs
- [Book or change an appointment](https://www.example.co.uk/contact/): online booking, phone hours, cancellation notice required
## About
- [About the salon](https://www.example.co.uk/about/): who runs it, grooming qualifications, insurance
## Optional
- [Gallery](https://www.example.co.uk/gallery/): before and after photos by breed
- [Coat care articles](https://www.example.co.uk/blog/): home brushing, seasonal shedding, choosing shampoosA reader now knows the business grooms dogs but not cats, where the van goes, that prices are confirmed on the day and where the cancellation policy lives. The home page has gone, because the summary already does its job, and the blog index has moved to Optional because it is useful background rather than something a new customer must read.
Harder cases
Large sites with hundreds of important pages
Do not list them all. Link to the section-level pages, such as category pages, documentation chapters or topic hubs, and write the note about the section: "Returns and refunds for all product types, including faulty goods and change of mind". If a section is large enough to deserve its own guide, some sites publish a separate Markdown file for it and link to that from the main file.
Very small sites
A five-page site does not need padding. A good summary, a details paragraph and five links with honest notes is a complete file. Do not invent sections to make it look more substantial.
Shops
Resist listing individual products unless you sell only a handful. Categories, size guides, delivery, returns, warranty and care instructions answer more questions than product pages, which change often and would make the file stale within weeks.
Regulated or sensitive sectors
Financial, legal, health and care businesses should keep notes strictly descriptive. "Explains how our fixed-fee conveyancing quotes are calculated" is safe. A note that sounds like advice, such as "the cheapest way to buy a house", invites an assistant to repeat it as if it were a recommendation. Nothing on this page is legal advice; if your sector has rules on promotional wording, they apply to this file as much as to your pages.
Several languages
Write notes in the same language as the page they describe, and make the language of each page clear. The language declarations guide covers how to lay out a multilingual file.
Keeping the content good
A file that reads well today will thin out again as pages are added, renamed and retired, unless someone owns it. A few habits prevent that:
- Write the note when you publish the page. If a page is important enough for the file, add its entry at the same time, while the content is fresh in your mind.
- Keep the page list in one place. A spreadsheet or a small data file with URL, section and note makes it easy to rebuild the Markdown and to spot entries with no note.
- Review the summary when the business changes. New areas, dropped services, a move of premises and a change of opening model all belong in the blockquote.
- Let automation handle links, not words. Scripts are good at checking that every URL still works and flagging new pages. They are poor at writing notes. Keep a human on the wording.
- Re-run the auditor after large content changes, such as a redesign or a migration, when both content and URLs tend to change at once.
Link failures and drift are covered separately in the guides on broken links in llms.txt and a stale llms.txt. If your key pages are long and mostly text, the next improvement after good notes is usually a companion llms-full.txt.
1. How long should each note in llms.txt be?
There is no rule in the proposal, but a single line of roughly 8 to 25 words works well for most sites. That is enough room to name the topics a page covers and add one specific, such as the area served or the product version, without turning the file into a wall of text. If a page needs a paragraph to describe, it is probably covering several subjects and the file may be clearer linking to more focused pages instead.
2. Should I use keywords in my llms.txt descriptions?
Write for understanding rather than keywords. llms.txt is not a ranking signal that any major search or AI provider has confirmed, so stuffing it with phrases achieves nothing and makes it harder to read. Naturally, a good note will contain the words people use for your services, because you are describing what the page covers. Use the plain terms customers use, such as boiler service rather than heating system maintenance solutions, and stop there.
3. Can I use AI to write my llms.txt notes?
You can use a tool to produce a first draft, but check every line against the page before publishing. Generated notes tend to be vague, repeat page titles or slip in claims the page does not make, which are exactly the weaknesses that make a file thin. The facts that make notes useful, such as areas covered, limits and exceptions, are often not stated clearly enough on the page for a tool to pick up, so they need your knowledge.
4. Should the home page be listed in llms.txt?
Usually not, if your summary blockquote already explains what the site is. The home page of most small business sites is a mixture of short teasers for other pages, so listing those deeper pages directly is more useful to a reader. The exception is a site whose home page carries real content that exists nowhere else, such as a single-page business site or a project whose home page is its main documentation.
5. What should go in the paragraph after the blockquote?
Facts that change how the rest of the site should be read. Examples include whether prices include VAT, which country or currency the site deals in, which software version the documentation covers, whether you sell to trade or public customers, and whether services need an appointment. Keep it to a few sentences. The proposal allows paragraphs and lists in this part but no headings, so anything that needs a heading belongs in a linked page instead.
6. Is it bad to have a very short llms.txt file?
No. Short is fine; empty is the problem. A small site with a clear summary and six well described links has a complete, useful file. The auditor is looking for content that helps a reader, not for a particular number of links. What makes a file thin is the absence of information: bare links, notes that say nothing, or a summary that could describe any business. Add detail to what you list before adding more links.
7. How often should I rewrite the notes?
Review them whenever the page they describe changes in substance, and read the whole file through at least every few months. Services, areas, versions and policies change, and a note that was accurate last year can now contradict the page. Tying the review to your publishing routine works better than a calendar reminder, because the person changing the page is the one best placed to update the note that describes it.
The llms.txt Auditor fetches your /llms.txt and /llms-full.txt files, validates them against the proposed format, checks that each linked URL responds, confirms robots.txt allows the relevant AI crawlers and scores the quality of the content.
Open llms.txt Fixes →