Here is the strange state of llms.txt in Australia right now. When we crawled 589 Australian business domains in August 2026, 181 of them, which is 31 per cent, had an llms.txt file. That is far higher adoption than anyone expected for a standard that is barely two years old. But a third of those files were generated automatically by an SEO plugin, 32 per cent contained not one fact worth quoting about the business, and only around 4 per cent of sites had a genuinely hand written file carrying real information. The full numbers are in our llms.txt study.
In other words, most Australian businesses now technically have the file, and almost none of them have the thing the file was actually invented to be. This guide covers what llms.txt is actually for, what belongs in it, what to leave out, and a template you can adapt in half an hour.
What llms.txt is, and what it is not
The llms.txt convention, proposed in 2024 and documented at llmstxt.org, is a plain text markdown file at the root of your domain that gives language models a curated, readable summary of your site. The reasoning is simple. Your website is built for humans, with navigation, scripts, pop ups and design furniture wrapped around the substance. An AI system reading your site has to dig the substance out. The file hands it over directly: who you are, what you do, and where the important pages are.
Three honest caveats before the template, because this topic attracts hype in both directions.
First, no major AI provider has formally committed to reading llms.txt the way search engines committed to robots.txt. Treat it as a low cost, no downside way to state your facts in the most readable possible format, not as a guaranteed channel. Half an hour of work for a plausible benefit and zero risk is a good trade. Betting your strategy on it is not.
Second, llms.txt is not robots.txt. It does not control access, block crawlers or protect content. Access control still lives in robots.txt and your firewall, and if your firewall is silently blocking AI crawlers, which we found on 13 per cent of Australian sites, no llms.txt file will matter because nothing can reach it. Check access first. Our AI SEO audit guide puts that check at step one for a reason.
Third, the file supplements your pages, it does not replace them. Assistants quote facts from your actual pages far more often than from any manifest. If your six key facts exist only in llms.txt and nowhere on your homepage, you have solved the wrong problem. Publish the facts on the site, then mirror them in the file.
Why plugin generated files miss the point
The plugin generated files we found, and we could identify Yoast, All in One SEO and Rank Math output by their signature lines, are essentially link lists. They enumerate your pages the way a sitemap does, without saying anything about the business. A model reading one learns your URL structure and nothing else.
The median llms.txt file in our crawl was around 7,000 bytes and averaged 119 links. Think about that from the reading side. A file with 119 undifferentiated links gives a model no way to know which three pages matter or what the business would want said about itself. Curation is the entire value of the format. A good file is short, opinionated and factual. If your SEO plugin already generates one, you can usually override or replace its output, and you should.
What belongs in the file: the six facts, then the map
When we audited 500 Australian homepages, we scored them against the six facts AI assistants most consistently quote when recommending a business: service area, pricing or pricing approach, licence or ABN, year founded, response time, and review presence. The average homepage published fewer than two of the six. Your llms.txt should carry all six, stated plainly, because this file is the one place where there is no design pressure, no brand voice debate and no reason to be vague.
After the facts, give the map: your eight to fifteen most important pages with a one line description of what each answers. Services, pricing, about, proof pages, and your best genuinely useful guides. Not every blog post. Not tag archives. The pages you would show a new customer if you could only show them ten.
The template
Here is the structure we use, adapted for a fictional Brisbane electrician. Swap the details and keep the shape. The format is markdown: a single H1, a summary block, then H2 sections with link lists.
# Smith Electrical Brisbane
> Licensed electrical contractor serving Brisbane's northern suburbs
> since 2011. Family owned, 6 electricians. QLD licence 78901.
> Specialists in switchboard upgrades, EV charger installation and
> solar maintenance. Typical response inside 4 hours for emergency
> calls in our service area. Callout fee $95 including first 15
> minutes. 4.8 stars across 214 Google reviews.
## Key facts
- Established: 2011
- Service area: Brisbane north side, Chermside to Redcliffe
- Licence: QLD electrical contractor licence 78901
- Team: 6 licensed electricians, all directly employed
- Pricing: callout $95, switchboard upgrades from $1,400,
EV charger installs from $1,100 installed
- Response: same day for emergencies, quotes within 48 hours
- Reviews: 4.8 from 214 Google reviews
- Insurance: $20m public liability
## Services
- [Switchboard upgrades](https://example.com.au/switchboard-upgrades):
what an upgrade involves, pricing and compliance
- [EV charger installation](https://example.com.au/ev-chargers):
single and three phase, load management, brands we install
- [Emergency electrician](https://example.com.au/emergency):
after hours faults, response times by suburb
## About
- [About us](https://example.com.au/about): our team, licence
details and history
- [Reviews](https://example.com.au/reviews): what customers say,
linked to sources
## Guides
- [EV charger cost guide](https://example.com.au/ev-charger-cost):
real installed prices for Brisbane homes
Notice what the summary block does. In under eighty words, a model reading it can answer where you work, how long you have existed, what you charge, how fast you respond, whether you are licensed and how you are rated. That is more quotable substance than most entire Australian websites offer, and every claim in it is checkable against a register, a review platform or the site itself, which is precisely what makes it safe for an assistant to repeat with your name attached.
The rules that keep it honest and useful
Every claim must be verifiable somewhere else. The licence number should match the public register exactly, under your exact trading name. The review figures should match your live profile. Assistants cross check, and we have watched ChatGPT run register lookups by name before recommending businesses. A file that overclaims is worse than no file.
Keep it under 200 lines. Curation is the entire value of the format. If everything on your site is important, then nothing is.
Use absolute URLs, because the file gets read out of context.
Update it on a trigger, not a schedule. New pricing, new service, new licence, review milestone: update the file the same week. Stale facts in a file that exists to state facts is the one way this format can hurt you.
Mirror, do not replace. Every fact in the file should also live on a human visible page, because pages are what get cited. The file is the index, the site is the evidence.
llms-full.txt, FAQs and the variations worth knowing
The convention has grown a few extensions since the original proposal, and two are worth a decision.
The first is llms-full.txt, a companion file that contains the full text of your important pages in markdown rather than links to them. The idea is to hand a model everything in one fetch. For documentation heavy sites and software products it makes sense. For a service business we mostly advise against bothering: your pages change, the file drifts stale, and the duplication creates exactly the inconsistency risk this whole exercise is meant to remove. A tight llms.txt pointing at well structured pages achieves the goal with one source of truth.
The second is including a short FAQ block directly in the file, five to eight of the questions buyers actually ask, each answered in a sentence or two. This we do recommend, because it costs twenty minutes and mirrors how assistants consume content: question in, answer out. Keep each answer self contained and factual. Do you service my area, what does a typical job cost, how fast do you respond, are you licensed and insured. If a question deserves a longer answer, answer it briefly in the file and link the page that covers it fully.
Formatting details that trip people up: the file must be plain text, UTF 8, served at the root with a normal 200 response, not behind a redirect chain or a bot challenge. Check it renders in a browser at the URL directly. We found files in our crawl that existed but were served with errors to non browser user agents, which combines the two most common failures in this field into one perfectly useless artefact: a file written for machines that machines cannot read. After any firewall or CDN change, re-check it along with your general crawler access, using the sequence in the audit.
And one maintenance habit that costs nothing: put the file’s review date inside the file itself, as a last updated line at the bottom. It reminds you, it signals currency to anything reading it, and it turns staleness from invisible to obvious.
Does it actually work?
The honest answer: the file is one layer of a system, and nobody can isolate its effect because AI answers vary between identical runs. What we can say from our own measurement is that businesses publishing the six facts consistently across homepage, schema and llms.txt get named and quoted in AI answers at rates that businesses publishing nothing simply do not, and the stated reasons assistants give trace back to the published facts almost every time. The file is the cheapest part of that system to build. Do it after crawler access and on page facts, before the slower external work, and spend no more than an hour on it. Then measure like everything else: fixed prompt panel, monthly, rates not screenshots, as laid out in how to get recommended by ChatGPT.
Frequently asked questions
Where does the llms.txt file go?
At the root of your domain, so yourdomain.com.au/llms.txt, the same place robots.txt lives. It is a plain text file in markdown format. No plugin is required to create one; any text editor and FTP or file manager access does it.
Should I use a plugin or generator to make it?
Use them for the skeleton if you like, but rewrite the content by hand. A third of Australian llms.txt files are plugin generated link dumps carrying zero business facts, which defeats the purpose of the format. The value is a curated summary a human wrote and stands behind.
Will llms.txt get me recommended by ChatGPT?
Not by itself. No provider guarantees they read it, and assistants quote your actual pages more than any manifest. It is a cheap, zero risk layer on top of the things that demonstrably matter: crawler access, facts published on your pages, schema, and presence on the sources assistants cite.
What is the difference between llms.txt and robots.txt?
robots.txt controls which crawlers may access which paths. llms.txt describes your site to language models that are already allowed in. One is a gate, the other is a welcome sheet. They live side by side and do not overlap.
How long should an llms.txt file be?
Under 200 lines. A summary block, a key facts list carrying the six quotable facts, and eight to fifteen curated links with one line descriptions. The median Australian file is 7,000 bytes of undifferentiated links, which is exactly what not to do.
Does a bad llms.txt file hurt me?
Two ways it can. Stale or inflated claims that contradict your site or public registers undermine the verification assistants run. And a plugin dump of a hundred links does nothing while letting you believe the job is done. An accurate short file or no file both beat a bad one.
