Your next website visitor might not be a person

We have spent years asking whether AI will replace websites.

I think that is the wrong question. Or at least the lazy one.

The nearer problem is more practical. Can AI agents use the websites we already have?

Because that is where things are moving. Not in a neat, finished, everyone-agrees-on-the-standard way. This is the web. It will be messy, partial, uneven, and full of vendors pretending they saw it all coming. The direction is still hard to ignore.

A recent piece by Slobodan Manic pulled the pattern together well. Cloudflare, Shopify, Stripe, Supabase, Netlify, and Google have all been building for agent-facing web activity in different ways. Not one coordinated campaign. Different companies, different incentives, same bet. Agents are becoming a new visitor class.

That phrase matters.

A visitor class is not a content trend. It is a design constraint.

Search engines became a visitor class, so websites had to care about crawlability, page structure, internal links, sitemaps, metadata, canonical URLs, and all the boring bits that make SEO people either useful or unbearable.

Mobile users became a visitor class, so sites had to work on small screens and flaky connections.

Now agents are turning into another one.

These agents might browse on behalf of a human. They might compare suppliers. They might check stock. They might read reviews. They might pull facts into an AI answer. They might monitor a market for changes. They might fill a form. In some categories they may eventually complete checkout or trigger a workflow.

This does not mean your buyer disappears.

It means your buyer may send something ahead of them.

And if that something cannot understand your website, you may not make the shortlist.

That is the uncomfortable bit. A human can forgive a messy site for a while. They can squint at a bad page, click around, infer what the business means, ignore the vague hero copy, and fill in the gaps from context.

An agent is less generous. It parses what it can parse. It follows the structure it can see. It uses the facts it can extract. If your offer is trapped inside a carousel, a poetic headline, a PDF, a sales deck, or four paragraphs of “tailored solutions” mist, good luck.

This is where the marketing conversation needs to grow up a bit.

AEO cannot just mean “do we show up in ChatGPT?” That is too narrow. It is a useful question, but it is not enough.

The better question is this. Can machines understand, trust, compare, and act on this business?

That takes us beyond prompt tracking and into website operations.

Cloudflare has already launched an Agent Readiness score. Its own scan of major domains found the basics are still early. Most sites have a robots.txt, but only a small minority declare AI usage preferences, only a small minority serve clean Markdown when agents ask for it, and newer standards such as MCP server cards and API catalogues barely show up at all.

That does not mean every small business needs to sprint into experimental protocols this week. Please do not let anyone sell you a five-grand “WebMCP transformation sprint” because they read a blog post and got excitable.

Start with the floor.

Can an agent fetch the important pages? Are the product or service pages clear without JavaScript gymnastics? Does the HTML say what the page is actually about? Is the sitemap current? Are the business, products, locations, pricing rules, service areas, FAQs, case studies, reviews, and proof points structured enough to be understood? Are forms labelled properly, or do they rely on visual tricks and fragile client-side behaviour? Can an agent tell what changed recently? Is the content specific, or is it the usual brochure fog?

That last one is bigger than people think.

Most websites are written for a human who already half-knows what the company does. They are not written for a machine trying to decide whether this supplier fits a buyer’s question.

“We deliver innovative solutions for ambitious teams” is useless.

What do you sell? Who is it for? Where do you operate? What does it cost, or what affects the price? What proof do you have? What makes you different from the other five options? What should happen next?

This is not glamorous work. It is plumbing.

But plumbing is usually where the money leaks.

There is another trap here. Magic-file theatre.

The llms.txt idea is useful in the right context. A simple machine-readable guide to important pages can help agents orient themselves, especially on developer documentation and content-heavy sites. Fine. Cheap hygiene. Worth considering.

But Ahrefs analysed 137,000 domains and found that 97% of llms.txt files received zero traffic in May 2026. Zero. No bots. No humans. Nothing. Google has also been lukewarm about it for search visibility, with John Mueller framing it more as a temporary crutch for some AI tooling than a generative-search ranking lever.

So if someone tells you llms.txt is the thing that will win AI visibility, be careful.

It might be part of the hygiene layer.

It is not the strategy.

The strategy is making the business legible in the places agents actually look and the systems actually use. That includes normal web pages. Product feeds. Documentation. Reviews. YouTube. Third-party mentions. Comparison pages. Fresh updates. Schema where it helps. Proper HTML. Public proof. Clear entities. Search and AI visibility data. And yes, perhaps agent-specific files and endpoints when the use case justifies them.

The point is not to worship any one standard. The point is to reduce ambiguity.

Bing’s latest AI visibility work points the same way. Its new Webmaster Tools preview adds Intents, Topics, Citation Share, and Compare for AI-generated answers. Microsoft is careful to say Citation Share is observational, not a ranking system. That caveat matters. The serious platforms know AI visibility is dynamic, contextual, and hard to flatten into one metric.

Good.

That is where useful marketing work begins.

Not “we added a file, therefore AI loves us”. More like this. When AI systems answer buying questions, do they include us? When they include us, do they describe us correctly? Which sources do they cite? Which competitors show up? What facts are missing? What proof is too buried for them to use? Where does the site block an agent that is trying to complete a reasonable task?

That is a much better brief than “write more blogs”.

This also changes what an agency should be selling.

The old agency model was mostly output. Pages, campaigns, assets, reports, a strategy deck if everyone had suffered enough.

The post-agency model has to be more operational. It needs to build the loops that keep the business legible and useful across human search, AI answers, agentic browsing, and commercial follow-up.

That means an Agent Readiness Audit should sit next to SEO and AEO work. Not as a gimmick. As a practical check.

I would want to know a few things. Can agents discover the important URLs and entities? Can they extract the offer, proof, pricing context, locations, availability, and differentiators? Can they understand what changed recently? Can they cite first-party and third-party evidence? Can they complete simple tasks through accessible forms or structured flows? Are risky actions blocked behind human approval or secure login? Are bot policies clear rather than accidentally hostile? Are analytics separating useful machine visits from human behaviour? Does the site tell a specific story, or does it dissolve into generic marketing foam?

That is not just a technical audit. It is a commercial audit.

Because if agents become part of how buyers discover and compare businesses, then agent-readiness becomes part of distribution.

This is especially obvious in ecommerce, SaaS, local services, B2B providers, marketplaces, publishing, events, travel, finance, jobs, property, and anything where comparison matters. An agent does not need to love your brand. It needs enough clean evidence to decide whether you are worth putting in front of the human.

That should make a lot of businesses slightly nervous.

Many websites are already weak for humans. They hide the offer. They avoid pricing. They bury proof. They make forms annoying. They rely on brand language that means nothing outside the boardroom. They publish content because someone said “SEO” in 2018 and nobody has questioned the ritual since.

Agents will not make that better.

They will expose it.

The good news is that most of the first fixes are not exotic. Make the offer clearer. Name the entities properly. Write pages that answer real buying questions. Put proof where it can be found. Keep important pages fresh. Use normal web standards. Make forms accessible. Stop hiding critical information in images, PDFs, videos, sliders, and vague copy. Track citations and representation over time, but do not pretend a single dashboard has the whole truth.

Then, once the basics work, look at agent-specific surfaces. Feeds, APIs, commerce protocols, action endpoints, MCP-style tools, and whatever standards actually get adoption rather than applause.

This is the sensible order.

Plumbing first. Protocols second. Hype never, ideally.

For business owners, the line is simple.

If an agent cannot understand your business, it probably will not recommend your business. If it cannot find your proof, it cannot use your proof. If it cannot tell what action to take next, it will either stop or pick someone easier to deal with.

That is not a future-of-the-internet keynote. It is a website problem.

Which means it is fixable.

But only if you stop treating AI visibility as a content trick and start treating it as part of the operating system around the business.

That is where the useful work is now.

Not more slop. Better surfaces. Cleaner evidence. Safer actions. Less fog.

The web is getting another kind of visitor.

Best not make it guess.

---

Jason Sibley is the founder of Cleo, a post-agency marketing and AI company. JasonVsTheNoise is where he writes about what is actually happening with AI, marketing, and how businesses should be thinking about both.