Case Study: How We Made agenticplug.ai Website Agent-Ready

Key takeaways

• AI assistants are now the middleman. Customers ask ChatGPT, Claude, and Perplexity to find and compare businesses instead of scrolling search results.
• Most websites cannot be read or used by them. Not because the content is weak. It's because nothing on the site is built for a machine visitor.
• agenticplug.ai started at 36 out of 100. Level 1 of 5, the lowest tier. And that's a firm whose own business is AI visibility. Most business sites score worse.
• We took it to 100 out of 100, Level 5, in a single focused build. Days, not months.
• The expensive mistake is faking it. Several checks pass on claims alone. Publish a capability you can't back, and AI assistants will try it, fail, and stop recommending you.

Disclosure: agenticplug.ai is backed by 360 Automation AI.

The problem: your customers are asking AI, and AI can't see you

For twenty years, getting found online meant one thing: rank on Google, get the click. That's still true. It's no longer the whole picture.

A growing share of customers now start somewhere else. They ask an AI assistant to recommend a contractor, compare two vendors, or explain what a company does. The assistant reads a handful of websites and returns an answer naming two or three businesses. If you aren't one of them, you never knew you were in the running. There's no bounce rate for this. No lost-lead report. It's invisible.

Here is what most business owners miss, and it's the crux of everything below: an AI assistant reads your website nothing like a person does. A person sees your design, your photography, your headline. A machine visitor is looking for structured, predictable signals: a clean version of your content, a plain statement of what it's permitted to do with that content, and some way to actually ask your business a question.

Most websites provide none of the three. They are built entirely for human eyes. To an AI assistant, they're a locked door with a nice paint job.

The baseline: 36 out of 100

Cloudflare, one of the companies that runs a large share of the internet's infrastructure, published a tool that measures exactly this. It visits a site, evaluates how usable it is to AI agents, and returns a score from 0 to 100 plus a level from 1 to 5.

We ran it against agenticplug.ai, an AEO/GEO and agentic-commerce consultancy whose entire business is making companies visible and transactable to AI.

It scored 36 out of 100. Level 1, "Basic Web Presence."

How we diagnose a site

We evaluate every site against three questions. Strip away the technical vocabulary and every single check reduces to one of them.

1. Can an AI assistant find you?

Not "find you on Google," but locate the machine-readable parts of your site. Most sites carry no signposts telling an arriving agent where anything lives.

2. Can it read you cleanly?

Websites are wrapped in layers of design code. An assistant trying to extract your services and pricing has to fight through all of it. There is a standard way to offer a clean, plain version instead. Most sites don't.

3. Can it actually do anything with you?

This is the decisive one. Even when an assistant reads your site perfectly, can it ask your business a real question and get a real answer? Look up your services? Start an inquiry? For nearly every website on the internet, the answer is no.

agenticplug.ai sat exactly where most of our clients sit. The first two were in reasonable shape: the basics were present, and the site already served a clean, readable version of its pages. On the third, it scored 1 out of 7. An AI assistant could read the site and do nothing whatsoever with it.

That is the gap on the overwhelming majority of sites we audit.

What was already right

What agenticplug.ai already had right according to the audit:

• robots.txt and a valid sitemap. The crawler basics, correctly configured. These are table stakes, and plenty of sites still get them wrong.
• Rules for named AI crawlers. GPTBot, ClaudeBot, PerplexityBot and others were addressed explicitly rather than left to a default.
• Markdown versions of every page. A clean, plain-text alternative that an AI assistant can read without wading through design code. This is the hardest item on the entire list to retrofit onto an existing site, and it was already done.
• WebMCP. A browser-based standard that lets a web page offer tools to an AI assistant while someone is actively viewing that page. This earned the site its only point in the category it was otherwise failing.

What we built

Here is the work in plain terms, grouped by what each piece accomplishes.

• Link response headers. A small piece of information now travels with every page response, invisible to humans, telling an arriving agent where to find the site's machine-readable resources.
• DNS for AI Discovery (DNS-AID) records, secured with DNSSEC. We added entries at the DNS level, the system that turns domain names into destinations, so an AI assistant can discover the site's services before it ever loads a page. DNSSEC is the security layer that proves those entries genuinely belong to the domain.
• Content Signals, published in robots.txt. A clear, standard declaration of what AI systems may do with the content: use it in search, use it to answer questions, use it for training. We set all three to yes, because visibility is that company's business. For a law firm or a publisher we would advise differently. The point is that the decision is now explicit rather than left to interpretation.
• An API catalog and an OpenAPI specification. A machine-readable directory of what the business exposes, so an agent does not have to guess or scrape it out of page text. The OpenAPI file describes each endpoint in a format that every developer tool and AI agent already understands.
• An Agent Skills index. Portable capability documents an AI assistant can load and act on. We wrote real ones rather than shipping placeholders, and each is fingerprinted so an agent can verify it received the genuine file.
• A Model Context Protocol (MCP) server, advertised by an MCP server card. This is the piece almost no business has. MCP is the emerging standard for connecting AI assistants to outside tools. The server lets an assistant query the company's knowledge base directly and get accurate answers from the source instead of inferring them from marketing copy. The server card is how agents find it in the first place.
• An OAuth 2.1 authorization server, documented in auth.md. Agents can now register themselves and authenticate securely, using the same category of standard sign-in technology that protects banking and enterprise software. We built it for real rather than merely claiming it, which is the whole point of the next section.

What came next

One piece was added after the score work was finished. The site already exposed WebMCP tools to assistants browsing the page, so we added one more: an anonymous consultation request. An AI assistant helping someone evaluate the company can now submit an inquiry on that person's behalf, with no account, no login, and no form to fill in.

The rule we don't break: never publish what you can't back

This is the part most agencies will skip, and it is the most important thing in this article.

The scoring tool checks whether a site publishes the right information. It does not verify that what you published actually works. Several checks can therefore be passed with claims alone:

• Publish a document explaining how to securely connect to your system, pointing at a system that doesn't exist. The check passes.
• Publish a listing for a service that isn't running. Point scored.
• List capabilities whose links lead nowhere. Still passes.

Any competent agency could game a client to a high score in an afternoon. We won't, for one blunt reason: a promise your website can't keep is worse than making no promise at all.

The score is a proxy. The actual objective is a real AI assistant arriving on behalf of a real customer, using what you advertised, and succeeding. When it reads your claims and hits a dead end, you have not gained visibility. You have published a broken promise, and that assistant's next move is to recommend a competitor.

So every document we published for agenticplug.ai is backed by something that genuinely works. The corollary matters just as much, and we apply it to every engagement: if you can't back it, don't publish it. A straightforward marketing site with nothing to protect does not need the secure front door. Skipping it and landing at 12 or 13 out of 14 is an honest, defensible result, and a better outcome than a fabricated perfect score.

Why this is harder than it looks

Four things went wrong that no documentation warns you about:

• A security key was generated one size too small. Every page loaded, every check passed, and the one function that actually mattered failed silently. Finding it took genuine diagnostic work.
• A copy-paste inserted invisible characters. Typing a setting on a Mac quietly converted plain quote marks into "smart quotes," which the DNS system rejects outright. The record looked correct on screen and was broken in reality.
• One change appeared broken for hours. A security setting reported as failing well after it had been configured correctly. Nothing was wrong. That layer of the internet simply takes time to propagate. Less experienced hands would have "fixed" a non-problem and made it worse.
• Leftover code from an old website tool surfaced. Scripts from a builder long since abandoned were still loading on every page, for every visitor. Removing them made the site faster and more secure.

None of these appear in a manual. They are the difference between a two-day project and a two-week one, and they are most of what you are actually paying for when you hire a team that has done this before.

The result

• Score: 36 to 100 out of 100.
• Level: 1 (Basic Web Presence) to 5 (Agent-Native), the top tier.
• Checks passing: 5 of 14 to 14 of 14.
• The weakest category (can an agent do anything with the site): 1 of 7 to 7 of 7.
• Timeline: a single focused build cycle.
• Pages redesigned: none.

What it changes in practice:

• AI assistants now get answers from the source instead of guessing. When someone asks an assistant about the company, it can query the business's own knowledge base directly rather than inferring from marketing copy. That materially reduces the risk of an assistant confidently telling a prospect something wrong about services, scope, or pricing.
• An assistant that arrives can finish the job. Before, it could read the site and stop. Now it can look up what the business offers and act on it. That is the difference between being mentioned and being chosen.
• The business is discoverable before a page even loads. The DNS-level entries let an assistant find the company's capabilities without crawling the site first.
• AI usage rights are explicit and on the record. What AI systems may and may not do with the content is now stated in a machine-readable way, rather than left to whatever each crawler decides to assume.
• It is a credential a prospect can verify. The score is public and anyone can re-run it in a minute. Top tier on a standard published by Cloudflare is a claim that survives being checked.
• Nothing broke. No pages redesigned, no URLs changed, no disruption to existing search rankings or traffic. Removing the abandoned scripts made every page load slightly faster and more secure.

‍
We verified every piece by hand rather than trusting the score, confirming that a real assistant could use what we had published. That distinction, between passing a test and actually working, is the difference between an agency that optimizes for a number and one that optimizes for your customers.

Frequently asked questions

What does it mean for a website to be "agent-ready"?

An AI assistant can find your site, read your content cleanly, understand what you permit it to do with that content, and interact with your business, not merely look at your pages.

How do I check my own site?

Use Cloudflare's free scanner at isitagentready.com. Enter your URL and you'll get a score out of 100, a level from 1 to 5, and a list of what passed and failed. It takes about a minute.

Is this the same as SEO?

No, though they're related. SEO determines whether you rank in search results. Agent-readiness determines whether an AI assistant can read and use your site at all. You can rank well on Google and remain invisible to ChatGPT.

Does this matter if I don't sell products online?

Yes. A small number of checks apply only to businesses selling directly through AI agents, and those are excluded from your score when they don't apply. Everything else, being found, read, and queried, applies to any business with a website.

What this means for your business

The pattern is consistent across nearly every site we audit:

• Most businesses are adequate on content and close to zero on the layer that lets an AI assistant actually use them.
• That layer is the cheaper half to fix, and the half that determines whether an assistant can do more than skim you.
• This is not a future problem. AI assistants are already the first stop for a growing share of customers deciding who to call.
• A business an assistant cannot read or query is quietly losing those customers to one it can, with no signal that it ever happened.

Anyone can publish claims. Making them true is what determines whether an AI assistant recommends you, and that is the standard we hold every engagement to.

360 Automation AI builds agent-ready websites for businesses that want to be found and used by AI, not just seen by people. Run the free scan and see where you land. If you want the gap closed properly, with nothing advertised that isn't real, get in touch.

360 Automation AI is an AI automation consultancy serving businesses in the Kansas City metro. Learn more about AI search optimization or contact us.

‍