Google’s new agent-readiness score gives its highest marks for exposing nothing

A perfect score is awarded for exposing nothing

If a site has no agent tools, Google does not mark it down for that. It decides those checks do not apply, removes them from the report card, and works out the score from whatever is left.

That happened to 166 of the 175 sites we scanned, out of 181 attempted. All three agent-tool checks were removed. For most, the llms.txt check was removed as well. What was left were two checks that have nothing to do with agents and existed years before them: whether the page structure is well formed, and whether the layout jumps around while it loads. Those two checks produced the entire grade.

So the headline number is not a pass mark. It is a score out of two questions, presented as though it answered six.

The result is that 22 sites scored 100% on their first pass, and 18 of them have no agent-callable tools and no llms.txt file. Finding 03 shows why “on their first pass” is doing real work in that sentence: some of these scores did not survive a second run hours later. Telstra, AGL and Origin Energy each score a flawless 100 while offering an arriving agent no way to do anything at all.

Every site we found with agent tools is a Shopify store

9 of 175 sites expose WebMCP tools. All 9 run Shopify. Not one of the other 166 sites, across banking, telco, travel, insurance, energy, government or education, exposes a single callable tool.

One caveat belongs here rather than in the footnotes, because it changes how the zero should be read. WebMCP is still an origin trial: a site that has not enrolled with Google’s trial cannot register a tool even if it wants to. So this number measures enrolment as much as intent, and no brand on that list has necessarily decided against agent tools. What it does show is that nobody outside one vendor default has taken the step yet.

None of the nine built anything. Every tool on every one of those sites is served from the same file:

This is the single most consequential finding in the dataset. Agent readiness in Australia is not being adopted, it is being distributed. It arrives when your platform vendor ships it, and it arrives complete, including tools that change state: update_cart, proceed_to_checkout and manage_orders are all registered by default.

The strategic implication is uncomfortable for anyone selling agent-readiness consulting. Exposure is becoming free. What remains scarce is knowing which tools matter, whether agents actually call them, and whether those calls turn into revenue.

The same site, the same day, a 77-point swing

We re-ran 35 sites under identical conditions a few hours after the first pass. 7 changed score. The average change was 26 points.

Amazon Australia was rated perfectly agent-ready in one run and near-worst in the cohort a few hours later. Nothing on the site changed. The two checks carrying the entire score are performance-sensitive, so the grade tracks network conditions rather than agent readiness.

A measurement that moves 77 points on an unchanged site cannot support a decision. If you are reporting this number to a client month over month, you are reporting noise.

Where Australian business actually stands

Setting the score aside and counting only what is observably true: tools present, and llms.txt present.

Travel is the sharpest gap in the dataset. Airlines and booking sites attract exactly the intent an agent is built to act on, a search with dates and constraints, and score lowest of any transacting sector at 35%. Not one exposes a callable tool.

Only 31% of the sites in this table publish an llms.txt, and 77% of the cohort fail Google’s page-structure check, which is a plain-language signal that agents struggle to read the page at all.

What to do with a score that cannot be trusted

• The score is not a target. It rewards absence, penalises effort, and moves by tens of points on an unchanged site. Optimising for it is optimising for nothing.
• The underlying checks are still worth passing. A well-formed page structure, a stable layout and an llms.txt are real, cheap and useful, and 77% of Australian brands fail the first one. Do that work because it helps, not because it moves a number.
• Exposure is arriving without you. Every one of the 9 tool-bearing sites in this cohort is a Shopify storefront, and not one of those merchants wrote a line of code to get there. If your platform ships it, you will get it. If it does not, the question is whether you need it yet.

The honest timing noteAgent-driven transaction volume today is small. Nothing in this report says otherwise, and anyone presenting agent commerce as a live revenue channel in Australia right now is ahead of the evidence. The argument for paying attention is that being early is currently cheap, not that the money has arrived.

The question worth asking is not “what is my agent-readiness score.” It is which agents are arriving, what they are trying to do, and whether anything on the page lets them do it. No static audit can answer that, because the answer only exists in your own traffic.