Agent readiness is about to be free. That changes what is worth measuring.
We scanned 181 Australian brands for WebMCP support and found nine, all of them Shopify stores that did nothing to get it. Then Cloudflare announced it would give the same thing to any website with no code at all. Here is what actually stays scarce.
There is a new thing you are being told to buy. It is called agent readiness, and the pitch is that AI agents are about to start doing your customers’ shopping, booking and comparing for them, so your website had better be able to talk to a machine.
The underlying technology is real. The urgency is mostly manufactured, and the specific thing being sold is in the process of becoming free. We went and measured it, so this is what we actually found rather than what the category would like you to believe.
What WebMCP is, in one paragraph
WebMCP is a proposed browser standard that lets a web page hand an AI agent a set of typed, callable tools. Instead of an agent reading your page and guessing which button does what, your site says: here is search_catalog, here is get_shipping_estimate, here is what each one takes. The agent calls the function. It is the difference between a stranger fumbling with your website and a developer using your API.
It is worth being precise about scope, because a lot of coverage is not. WebMCP governs what an agent can do once it is already on your site. It has nothing to do with whether ChatGPT or Gemini names your brand in the first place. Those are two different problems, and only one of them has money attached to it today.
We measured 181 Australian brands. Nine had tools.
In August we ran Google’s Lighthouse agentic-browsing audit across 181 Australian brands that people actually transact with: banks, supermarkets, airlines, insurers, telcos, energy retailers, government services, universities. 175 could be scanned.
Nine of them exposed WebMCP tools. All nine were Shopify storefronts. Every one served the same ten tools from the same file on Shopify’s CDN. Not one merchant wrote a line of code to get there. Across banking, telco, travel, insurance, energy, government and education, we found nothing at all.
Key takeaway
Agent readiness in Australia is not being adopted. It is being distributed. It arrives when your platform vendor ships it, and it arrives complete.
One caveat belongs here rather than in a footnote. WebMCP is still an origin trial, so a site that has not enrolled with Google cannot register a tool even if it wants to. That number measures enrolment as much as intent. Nobody on that list has necessarily decided against agent tools. What it does show is that outside one vendor default, nobody has taken the step.
Then the interface layer started going free
During its Agents Week in August 2026, Cloudflare announced a developer preview that gives any website a WebMCP interface: a bridge injected at the edge that registers tools for unmodified pages, with no code and nothing changed at the origin. Its remote browser product had already added the other half, so an agent can discover and call whatever a site exposes.
Put that next to the Shopify finding and the pattern is hard to miss. The thing being sold as a strategic project is being handed out for free at the platform layer and at the CDN layer, by companies whose distribution you cannot match. If your answer to AI agents is a consulting engagement to hand-write tools, that answer has a shelf life.
Hand-authored tools are still better. A bridge that infers tools from your markup will never be as precise as one written by someone who knows which three things a customer actually wants to do. But having tools at all is about to stop being a differentiator, and pricing it like one is a mistake.
The scoring, meanwhile, does not work
The audit that agencies will start pasting into client reports is not measuring what people assume. If a site has no agent tools, Lighthouse does not mark it down. It decides those checks do not apply, removes them, and grades whatever is left. What is left is usually two general web-quality checks that existed years before agents did.
Twenty-two sites in our scan recorded a perfect 100%. Eighteen of them expose nothing at all to an agent. Telstra, AGL and Origin Energy each scored flawlessly while offering an arriving agent no way to do anything. Meanwhile six of the nine sites that actually have tools scored 75%, because registering tools switches more checks back on and gives you more ways to lose points.
We also re-ran a sample of sites a few hours later under identical conditions. One unchanged site moved 77 points. If you are reporting this number to a client month over month, you are reporting network conditions.
What the standard still has not solved
This is the part the vendor posts skip, and it is not our objection. It comes from the proposal’s own security and privacy questionnaire:
- Over-broad tools leak data. A tool can ask for far more personal information in its parameters than the task needs.
- State can cross origins. An agent browsing several sites in one session may carry state from one to another.
- There is no consent primitive. Hints for consequential actions are planned but not yet normative.
- There is no authentication story. A tool inherits the page’s session, and that session is the authorisation. No scoped tokens, no per-tool consent, no audit trail.
Which gives you a rule worth keeping: do not expose a tool that spends money, changes an account or messages a human until that is settled. Read-only tools, and tools that pre-fill a form for a person to confirm, are a completely different risk class.
So what should you actually do
- Check whether your platform already did it. If you are on Shopify you have ten tools and did nothing to get them. Look before you build.
- Fix the boring things first. A well-formed page structure, a stable layout and an llms.txt are cheap, useful today and read by systems that already exist. 77% of the Australian brands we scanned failed the page-structure check.
- If you do build, start read-only. Search, lookup, availability, a pre-filled form. Keep a human on anything consequential.
- Do not optimise for the score. It rewards absence, penalises effort and moves by tens of points on an unchanged site.
The honest timing note
Agent-driven transaction volume in Australia today is small. Nothing we measured says otherwise, and anyone presenting agent commerce as a live revenue channel right now is ahead of the evidence. WebMCP is a Community Group proposal, not a ratified standard. It runs in a Chrome and Edge origin trial and nowhere else. Firefox and Safari have not implemented it. Stable Chrome is projected for late 2026 and is not committed.
The argument for paying attention is that being early is currently cheap, not that the money has arrived.
Key takeaway
If exposure is becoming free, the scarce thing is no longer whether an agent can act on your site. It is whether anything is arriving at all, and whether the engines are naming you when a customer asks.
That second question has revenue attached to it today, and it is the one we would answer first. The free AI visibility check asks all six major AI engines a generic question in your category, never naming your business, and reports which of them recommended you. It takes about a minute and needs no account.
The full method, the site-level tables and every limitation we could find in our own work are in the agent readiness benchmark.
