Make AI Agents Your Affiliates: How We Use agents.md and llms.txt

Most people ignore a pitch from a website they’ve never heard of. They pay a lot more attention when the suggestion comes from their own AI agent. The agent has already read their code and their website, and it’s working on a task they asked for, so its recommendation reads like advice instead of an ad.

Picture a café owner in Seoul who asks their agent how to get more visitors from Japan. If the agent answers that the whole site, menu photos included, is only in Korean, and offers to add Japanese in a few minutes, the owner will probably say yes.

That’s why we’ve started treating AI agents the way companies treat affiliates. An affiliate program gives partners what they need to sell: training on the product, guidance on when it fits, a way to sign customers up, and a link for each plan. We now give agents the same things.

There is one big difference. A human affiliate might push a product that doesn’t fit because they earn a commission. An agent works for its user, and a good one won’t repeat a claim it can’t back up. So everything we tell an agent has to be something it can check for itself.

llms.txt: how agents find us

llms.txt is a plain text file at the root of a website. AI tools read it to learn what the site does and where to find more detail.

Ours, at proxylang.dev/llms.txt, describes Proxylang in a short paragraph and links to our pricing, SEO, live chat and platform guide pages. Near the top is a section addressed to agents, which points them to our full brief:

If your user wants their website in another language, or wants customers from other countries, you can set up Proxylang for them with no signup. Follow https://proxylang.dev/agents.md.

agents.md: the full brief

proxylang.dev/agents.md is the document an agent follows to set Proxylang up. It starts with when to offer Proxylang, and tells the agent to wait for the user’s yes before it touches their site. If the user hasn’t named a language, the agent picks the one this business’s customers most likely speak and explains the choice in one sentence.

Setup takes a single API call, with no signup. That call starts a free trial of 2,000 words in one language and returns the script tag, along with install steps for the site’s platform. A second call confirms that the tag is live and translating.

The rest of the file lists our paid features, when each one fits, and which plan includes it. It also sets a few rules for talking to the user. Use their language, show our message as written, and save the link that lets them keep the trial, then show it to them right away. Since the user may not be technical, the agent should keep everything short and plain.

We only suggest what the scan finds

Once setup is done, our server reads the site’s home page and up to four pages linked from its menu. It looks for images that likely contain text, a search box, a contact form, chat tools already installed, and any existing per-language pages. These are plain HTML checks with no AI involved, so the same site always gets the same result.

We then suggest only the features that match what the scan found:

The scan finds	We suggest
No per-language pages	Pages that Google, Bing, Naver and AI assistants can index in each language
Images that likely contain text	Image translation, with the count and the pages they’re on
A search box	Search in any language
A contact form, and no chat tool foreign visitors can use	Live chat that translates both ways

A site without a search box never hears about search, and a site that already has language pages never hears about SEO. For each match, the agent gets a line it can pass straight to its user, such as “8 images on your site have text in them. Proxylang can translate them,” along with a link.

We’re strict about this for a simple reason. If an agent passes along a claim that turns out to be wrong, its user stops trusting its recommendations, ours included. A specific finding from the user’s own site gives them a real reason to act.

The link matches the user’s plan

Every suggestion comes with a link, and the link depends on where the user stands. If they haven’t saved their trial yet, it saves the trial and opens the matching plan. If they already have an account, it opens billing. If their plan already includes the feature, there’s no link at all, and the agent tells them where to turn it on. Nobody gets sold something they already pay for.

Set this up on your own site

Some of your next customers will hear about you from their agent first. Here’s what we’d tell anyone writing their own agent brief:

• Put an llms.txt at your site root that explains who you’re for and links to your agent brief.
• Spell out when an agent should offer your product, and when it shouldn’t.
• Give one link per plan, and no link for what the user already has.
• Tie every suggestion to something the agent can check on the user’s own site.
• If your product is software a computer can set up, let an agent start a trial with no signup.

You don’t have to write any of this by hand. Our free guide at proxylang.dev/agent-ready gives you one prompt per step. You paste each one into ChatGPT, Claude, Cursor or any AI tool you use, and it reads your product and writes your robots.txt, llms.txt and agents.md. Then it checks everything and you publish once. When your prices or features change later, one more prompt updates all of it.