Declares what an agent can do here

What it is worth

Two numbers rather than one, because a check's weight is inside its pillar and the pillar is a share of the whole. The largest check in the smallest pillar moves your score less than a middling one in the largest, and knowing which you are looking at is the difference between fixing the right thing and fixing the loudest thing.

Why an agent cares

Inference is where an agent gets it wrong or gives up. Declaring the action, its URL template and the parameters it takes is the difference between an agent that can complete a task on your site and one that reads about it and leaves.

Worth saying plainly: this is not a search-engine ranking factor and passing it will not move you up a results page. It measures something else — whether software with no browser, no JavaScript and a few hundred tokens of budget can tell what your page is and do something with it. Those are the conditions an AI assistant reads you under, and they are much harsher than a browser's.

How to fix it

Declare the actions this page supports with schema.org potentialAction.

Any page. The site-wide SearchAction belongs on the home page; an OrderAction or ReserveAction belongs on the page that offers it.

Copy{
  "@context": "https://schema.org",
  "@type": "WebSite",
  "url": "https://example.com/",
  "potentialAction": {
    "@type": "SearchAction",
    "target": {
      "@type": "EntryPoint",
      "urlTemplate": "https://example.com/search?q={query}"
    },
    "query-input": "required name=query"
  }
}

That remediation is the one the scanner itself offers when this check fails, read from the same code that does the scoring rather than written alongside it. If it were typed out separately it could drift from the check that judges it, which is the failure that makes published advice worthless.

How often the web fails this

Measured on 2026-09-06 against AX specification v1.5.1, across a sample of 292 public websites.

The last row is kept out of the percentages deliberately. A check that did not apply to 0 of 292 sites is not a check most of the web passes, and folding those in is the commonest way this sort of table is used to say something untrue.

At 84% this is one of the things nearly everybody gets wrong, which cuts both ways: you are not unusual for failing it, and fixing it separates you from most of your competitors rather than catching you up with them.

Testing it on your own site

Copynpx axray-cli your-site.com

All 70 checks against one URL, this one among them, with what was found rather than only whether it passed. Add --min 70 in a pipeline and the command exits non-zero below that score, so a change that breaks what agents can read fails the build instead of shipping.

Questions about this check

What does the "Declares what an agent can do here" check mean?

Inference is where an agent gets it wrong or gives up. Declaring the action, its URL template and the parameters it takes is the difference between an agent that can complete a task on your site and one that reads about it and leaves. It sits in the Actionability pillar of the Agent Experience rubric and is worth 5 points within that pillar, which works out at about 1.2 points of the overall score out of 100.

How do I fix potential-action?

Declare the actions this page supports with schema.org potentialAction. Any page. The site-wide SearchAction belongs on the home page; an OrderAction or ReserveAction belongs on the page that offers it. It is a configuration or content change rather than an engineering one, so it can usually ship the same day.

Is it unusual to fail potential-action?

No. Of 292 public sites where this check applied, 244 failed it — 84%. The measurement was taken on 2026-09-06 against AX specification v1.5.1, and no site is named.

How do I test potential-action on my own site?

Run npx axray-cli your-site.com from a terminal. It runs all 70 checks against one URL and prints this one among them, with what it found rather than only whether it passed. It is free, needs no account and installs nothing.

The rest of the Actionability pillar

Can the agent take the next step - navigate, search, submit, buy - without guessing?

• Navigation uses real links — 12 points
• Link text describes the destination — 8 points
• Links onward into the site — 6 points
• Form fields are labelled — 8 points
• Form fields have stable names — 6 points
• Fields declare autocomplete tokens — 4 points
• Controls are real buttons — 6 points
• Reachable by a human, discoverably — 5 points
• Site search is machine-usable — 4 points
• Programmatic surface is discoverable — 6 points