Your website has two audiences now. Only one of them is human.

Last week I audited a site that looked great in a browser. Clean design, full service pages, real photography. Then I looked at what an AI agent gets when it reads the same address: a nearly empty shell. The content arrives through JavaScript after the page loads, and most agents read the page before that happens. For the fastest-growing way buyers do research, this business barely exists.

That site is normal. Most small-business websites are built for exactly one reader, and the second reader showed up without an announcement.

Every buyer now has a research assistant

People ask ChatGPT and Claude things they used to ask Google: who does this kind of work, is this company legitimate, what do they charge. The assistant goes and reads websites to answer. Whether it can actually read yours decides whether you are in the answer.

You can see this shift in a strange place: robots.txt, the file that tells crawlers what they may visit. Mine names seventeen different AI crawlers, each with its own permissions. Two years ago that file had one audience worth addressing. Now the crawler list reads like a customer list.

What an AI agent actually sees on your site

Load your homepage, then imagine it stripped of everything that requires JavaScript to appear. For a lot of modern sites, what remains is a logo, a menu, and silence.

Humans get the rendered page. Agents get the source. When your copy, your services, and your proof all render client-side, the agent's version of your business is a blank page with good SEO intentions.

The audited site above scored well on design and dismally on this. Same address, two completely different businesses depending on who asked.

The fix is a handful of plain text files

I run this setup on my own properties (byteflowai.com, aiforlife.app, aiguymatt.com), so everything below is deployed and checkable, not theory.

llms.txt. A plain-language summary of who you are, what you offer, and where the important pages live. Agents that look for it get the whole business in one read. Mine lists the offer funnel, common questions, and every discovery endpoint.

A real sitemap. Machine-readable list of your pages. Most platforms generate one; most owners have never checked whether it includes the pages that matter.

Structured data (JSON-LD). A block in your page's code that states, in a format agents parse natively: this is an Organization, here is the founder, these are the services and prices, these are the FAQ answers. It removes the guesswork from being understood.

Content negotiation. The sharpest tool of the set: when an agent asks my site for markdown instead of HTML, the site serves it. You can try this on my homepage right now:

A human gets the designed page. An agent that asks gets clean, readable text of the same business. Two audiences, one address, each served in its own language.

Crawler signals. The robots.txt policies that say: search engines yes, AI reading yes, AI training no. You decide the terms instead of defaulting into whatever happens.

None of this requires a rebuild. These are additions that sit alongside the site you already have, on any host and any platform.

What this looked like on my own property

Earlier this month I took aiforlife.app, my free Claude-learning site, through the full treatment in one working session: a 22-post article section, structured data on every post, twenty-three new sitemap entries, an llms.txt describing the whole site, and markdown variants served to any agent that asks. Every artifact went live with automated tests confirming it, and the whole surface is public if you want to see the pattern deployed rather than described.

The point of the receipt: this is a bounded piece of work with a visible finish line, done in days. It sits at the opposite end of the effort scale from a redesign.

One rule: describe only what exists

A warning from my own playbook, because the temptation is real. Agent-discovery files are statements of fact, and agents treat them as such. My hard rule when deploying these artifacts: publish metadata only for infrastructure that actually exists. A site that advertises an API it never built, or claims capabilities it does not have, is lying to the most literal-minded readers on the internet. They will act on the claim, fail, and file your business under broken.

Describe the business you have, accurately, in formats agents parse. That alone puts you ahead of nearly every competitor's site.

How to check your own site in five minutes

Three checks, no developer required:

1. The source test. Open your homepage, view the page source (right-click, View Page Source), and search for a sentence from your main copy. Missing from the source means agents are missing it too.
2. The robots test. Visit yoursite.com/robots.txt and yoursite.com/llms.txt. A 404 on both means you have given AI readers zero guidance.
3. The assistant test. Ask an AI assistant with browsing enabled: "What does the business at [your URL] do, and what does it charge?" The answer it gives is the answer your prospects are getting.

Score yourself simply: three passes means your site already speaks to both audiences. One or two misses means the fixes above are your shortlist, in the order listed. Three misses puts you where the site I audited last week sits, invisible to the reader that researches on your buyers' behalf, and that is worth knowing this week rather than next year.