Back to Research Lab
Issue 01 · Agent Infrastructure

The Action Layer: What 122,000 Website Scans Showed Us About AEO vs. Agent Tools

Answer engine optimization gets your business quoted by an AI. It does not get a job booked. Here is what we found when we scanned the websites of a whole country to measure the difference.

By Gabriel Foley · 1 November 2026 · Sydney, Australia · 6 min read

A painted green front door at golden hour, with a phone lying lit on the step.
The phone on the doorstep: an assistant that has found the business and still cannot get through the door.

Ask an assistant who fixes a leaking tap near you and it will give you a confident, well-written answer. Ask it to book one for Thursday morning and, most of the time, it will hand you back a phone number.

That gap is the whole story of this issue. For two years the web has been reorganising itself around the first sentence: making sure an AI can read your site and quote it. The industry has a name for that work, answer engine optimization, or AEO, and it is real and worth doing. But an answer ends in a sentence. A booking ends in a changed calendar, and between the two sits a list of things a reading AI cannot do.

The answer is not the action

When an assistant reads your page, it is doing what a very fast customer does with a brochure. It can learn that you exist, what you charge and where you work. It cannot check that Thursday is free, get past the form that wants a CAPTCHA solved, or tell you who it is acting for. It can describe the door. It cannot open it.

We call the missing piece the Action Layer: the part of the web where an assistant calls a tool instead of reading a page. A tool has a name, a list of the fields it needs and a defined result. It does not care how nicely the website is laid out, and it does not break when the designer moves a button.

What 122,357 websites looked like

We build infrastructure for this, so we went and measured it. Our directory now holds 193,533 Australian businesses, and we have scanned the websites of 122,357 of them the way an agent would arrive: no human, no logged-in session, just a machine trying to find something it can do.

The first number is encouraging. 77,340 of those sites, 63.2 per cent, offer some action a human can take: a booking button, a quote form, a phone link. The business is, in the ordinary sense, open for work online.

The second number is the one that matters for an assistant. Only 13,654 of the sites, 11.2 per cent, offer an action an agent can call directly. That leaves 88.8 per cent of the sites we scanned with a front door built for people and no handle for software.

The numbers, 1 November 2026

Eight in nine sites give an AI nothing it can call.

National directory

Businesses indexed
193,533
Websites scanned
122,357

With an action a human can take63.2% (77,340)

With an action an agent can call11.2% (13,654)

88.8%Action Layer failure rate: scanned sites with nothing an agent can call.

Deep Sydney benchmark, 2,946 scored sites

Average readiness
41.5 / 100
Agent-ready sites
0.5% (16)
  • 99.9% publish no MCP endpoint
  • 31.5% are blocked by a CAPTCHA
  • 5.4% use a scheduling widget
  • Square41
  • Calendly29
  • Fresha25
  • Timely25
  • Cliniko20
  • Acuity19
Figures from the Frontlatch directory, queried 1 November 2026. The Sydney benchmark is the 2,946 scored sites in our Sydney index.

Why the page fails the tool call

To see why, we took a closer look at Sydney, the city where we have scored the most sites in depth: 2,946 of them. The average site scored 41.5 out of 100 on our readiness scale, and only 16, 0.5 per cent, were what we would call agent-ready.

The reasons are mundane, which is what makes them fixable. Nearly all of them, 99.9 per cent, publish no MCP endpoint, so there is nothing for an assistant to connect to. 31.5 per cent put a CAPTCHA in front of a form, which is the one thing an agent must never try to get around. And only 5.4 per cent use a scheduling widget at all, spread across a handful of vendors, each with its own markup and its own frame an agent cannot see into.

The website was built to be looked at. Almost nothing on it says what can be done there, to software.

From the answer layer to the action layer

None of this is a prediction about a distant future. Google Connected Apps, Claude and ChatGPT already let an assistant reach into other software and do something, not just say something. The direction of travel is the same everywhere: out of the answer layer, where AEO lives, and into the action layer, where a tool call either succeeds or fails.

In that world a business is not found by a ranking. It is chosen by whether the assistant can finish the job. A site that reads beautifully and cannot be called loses to a plainer one that can.

This is why we expose our gateway as 12 named tools over MCP, the open standard these assistants use. An assistant can find a business, list what it can do, describe one action in full and request it. The shape is the same for every business, so an agent does not have to learn each website separately.

What stays human

Making sites callable only works if it is safe, so the safety rules are part of the layer, not an extra. No action runs on a business that has not approved that kind of action. A CAPTCHA is never bypassed, and a password or a card number never passes through an agent or through us. When a step needs one of those, the assistant stops and hands the step back to the person.

The owner stays in charge of the result. A booking arrives as a request with the details attached, nothing confirms until the owner approves it, and when it does both sides hold a signed receipt.

AEO gets you cited. The Action Layer gets you booked, and keeps the owner in the loop while it does. The only way to know which side of that line your own website is on is to test it.