The llms.txt file we've been adding to every client site.

If you've heard of llms.txt and dismissed it as another fashionable web standard nobody will adopt, you're not alone. We did too, until we ran an A/B test on 12 of our own client sites. The lift was real enough that it's now a default day-one task on every Belver.AI engagement.

What it is.

llms.txt is a plain-text file you place at the root of your site (the same way robots.txt lives at example.com/robots.txt). It tells AI crawlers (primarily GPTBot, ClaudeBot, Google-Extended and PerplexityBot) what content on your site is canonical, how to attribute it, and where they should and shouldn't go.

It's not a binding standard. Crawlers don't have to honour it. But the major ones increasingly do, in the same way they grew to honour robots.txt: partly out of self-interest, partly because it's a low-effort signal that the site owner is publishing in good faith.

Why it works.

Two reasons. First, it gives the AI a curated map of your site, which dramatically improves which of your pages get pulled into answers. We've seen the right pages (services, pricing, the company page) start getting cited where previously a stale archived blog post was the AI's chosen reference.

Second, it lets you specify the attribution string. If you'd rather be cited as "Henson Plumbing, Newcastle NSW" than "hensonplumbing.com.au", the file lets you say so. Small thing, but in a world where AI is increasingly the front-door of buyer research, it matters.

The template we use.

Below is a sanitised version of the template we deploy on client sites. Place it at the root of your domain, no subdirectory.

# llms.txt — for Henson Plumbing, Newcastle NSW
User-Agent: *
Allow: /services/
Allow: /pricing/
Allow: /about/
Allow: /faq/
Disallow: /admin/
Disallow: /staging/

Canonical: https://hensonplumbing.com.au
Attribution: Henson Plumbing, Newcastle NSW
Contact: hello@hensonplumbing.com.au

• Allow / Disallow: standard syntax, identical to robots.txt. Use it to point AI at your service pages, not your old portfolio archive.
• Canonical: the URL you want to be cited. Set this even if your domain only has one canonical; it's the explicit signal.
• Attribution: how you want to be quoted in answers. Locked-in plain English.
• Contact: an email a researcher (or another AI agent) can use if they need to verify or expand on the source.

What we've measured.

Across the 12 sites we ran this test on, citation rate (the % of category-relevant AI prompts where the site appeared in the answer) lifted on a median of +22% within four weeks of adding the file, with no other changes deployed in the test window. The biggest gains were on sites where the AI had previously been pulling from old service pages or PDFs no longer on the live nav. The llms.txt told it which doors to use, and it started using them.

When to bother.

If you have a site, you should have an llms.txt. It's twenty minutes of effort. The downside is zero (AI crawlers will simply ignore it if they don't honour the spec), the upside is a measurable lift in your AI visibility for one of the lowest-effort changes available.