llms.txt for small businesses: what it is and whether it does anything

Every so often a new file appears that is going to change how the web works, and within a fortnight somebody is selling it to small businesses as a package. The current one is llms.txt.

We publish one on this site. I also think most of what is written about it is nonsense, which makes for an awkward position to defend, so here is the honest version: what it is, who is genuinely reading it, and whether you should care.

Jeremy Howard of Answer.AI proposed it in September 2024. The reasoning is sound. Language models have limited room for context, and turning a modern web page into useful text means stripping out navigation, adverts, cookie banners and JavaScript, which is both slow and error prone.

So you publish a file at yoursite.co.uk/llms.txt holding a clean, plain summary of your site in markdown, with links to the pages that matter. A machine trying to understand your business reads that instead of wading through your HTML. The shape of the idea is borrowed from robots.txt, which has worked for about thirty years.

This is the part the people selling llms.txt packages tend to move past quickly.

As of early 2026, no major AI company has publicly committed to reading llms.txt in production. Not OpenAI, not Google, not Meta, not Mistral. Anthropic has been the most positive about it, and Perplexity has said it fetches the file to help decide which pages to prioritise, but that is a long way from the claim that AI search reads your llms.txt.

Google has been openly dismissive. Gary Illyes said in 2025 that Google does not support it and has no plans to, and John Mueller compared it publicly to the old keywords meta tag, which is roughly as cutting as search engineers get. OpenAI's own documentation points people at robots.txt for crawler control instead.

The server logs match the rhetoric. Sites that have published one report GPTBot fetching it occasionally, and most of the other crawlers essentially never touching it.

Adoption on the publishing side is real but lopsided. One study of 300,000 domains found roughly one in ten had a file, and those cluster heavily in software documentation, AI tooling and open source projects rather than among plumbers in Buckinghamshire.

Three files get mentioned in the same breath and they do completely different jobs, which causes real damage.

robots.txt tells crawlers what they may and may not fetch. It is decades old, honoured by every company that matters, and it is where you control whether AI companies can crawl your site at all. If you only touch one file, touch this one.

llms.txt runs in the opposite direction. It restricts nothing. It offers a tidy summary to anybody who wants one. ai.txt is a third proposal in the same territory with even less traction, and I would not spend an afternoon on it.

Somebody who believes their llms.txt is stopping AI companies using their content has protected precisely nothing, and that misunderstanding is common enough to be worth spelling out.

Four reasons, none of them dramatic.

It costs almost nothing. Ours generates itself from the same content that builds the website, so it took an afternoon once and then maintains itself. If it were a file somebody had to remember to update by hand, I would not bother.

It is useful to humans. Ours is a readable plain text version of the entire site, and I have used it myself to check whether a change actually went live.

The cost of being wrong is lopsided. If the idea gets adopted properly in two years, we already have one. If it never does, we lost an afternoon.

And writing one forces you to answer a question worth answering. If you had to describe your business in plain text, with no photographs and no design to hide behind, what would you actually say? More than one client has improved their homepage after seeing what their site sounds like with the pictures stripped out.

None of which means the idea is bad. Standards take years, and robots.txt itself spent a long stretch as a proposal nobody was obliged to honour. It means the honest status today is promising and largely unread, which is a very different sentence from the one in most of the marketing.

The format is deliberately plain. A heading with your business name, a short summary of what you do, some optional detail, then lists of links each with a single line explaining what is behind them.

For a small business that means: who you are and where you are, your main services, opening hours and contact details, links to the pages that matter, and anything you would want a machine to get right rather than infer. A restaurant would put the menu in it. A tradesman would put the areas covered and the fact that quotes are free.

The specification also allows an optional section for material that can be skipped when space is tight, which matters far more for a documentation site than for a plumber with nine pages.

There is a real shift underneath all the llms.txt noise, even if the file itself turns out to be a footnote.

For twenty years the job was to write for people and let search engines take a copy. Increasingly a machine reads your website, decides what your business is, and repeats a summary of it to somebody who never visits you at all. You do not get to review that summary before it goes out. It is assembled from whatever your site makes easy to understand.

That is a strong argument for plain, specific writing, for putting facts in text rather than burying them in images, and for saying what you actually do instead of describing yourself as a leading provider of bespoke solutions. All of which helped before any of this and helps rather more now.

It is not an argument for buying a new file every time somebody publishes a proposal.

If the goal is being recommended when somebody asks an assistant for a plumber in Aylesbury, llms.txt is not where the leverage sits. These are:

• Structured data on your pages, so a machine knows your opening hours are opening hours rather than a sentence that happens to mention Tuesday.
• Clean HTML with real text in it. If your prices live inside an image, no assistant can quote them to anybody.
• Consistency across the web. Your name, address and phone number matching everywhere they appear, down to the punctuation.
• Reviews and mentions on sites the models already trust, which for a local business mostly means Google.
• Content that answers the questions people actually ask, phrased the way they ask them.

Boring, all of it. Also the part that works.

For anyone technical, the site runs on Next.js with Sanity as the CMS. There is a route at /llms.txt that runs a GROQ query pulling the editorial content out of the CMS and renders it as plain text. When somebody edits a page in Sanity, the file changes with it. There is no separate document to maintain and no way for the two to drift apart, which was the whole point.

The trap worth avoiding is writing one by hand. A stale llms.txt is worse than none at all, because you have published a confident, machine readable description of a business that has since changed its prices and dropped half its services.

If your site sits on something that can generate the file automatically, publish one. It is cheap and it might matter later.

If it means paying somebody to write and then maintain a file by hand, do not. Spend that money on reviews, on decent photographs, or on making your existing pages load faster on a phone.

And if somebody sells you an llms.txt package as the route into AI search results, ask them which AI companies have committed to reading it in production. The honest answer, as things stand, is none of the large ones.

We build this into every site we make, mostly because doing it properly costs us nothing. If you want to see ours, it is right here. If you want an opinion on whether any of this is worth your attention, get in touch and I will tell you what I would do in your position.