Everything else measures how your site is built. This measures whether an agent can use it.
A check can tell you a Product block is missing. It cannot tell you that an agent sent to buy from you gave up on the third page, or that it reported a price you have never charged. Only sending one can.
The errand
A task is one instruction a customer might give an assistant, pointed at one site: find this product's price, reach the checkout, say when you open today, cite this article. Which ones apply is read from what your site already is, so a shop is never asked for a byline.
48 errands across 7 kinds of site
A site with no article template owes nobody an article. That is not a zero. It is not applicable, and the difference is deliberate: scoring a site for an absence it was right to have is the mistake the plugins were careful not to make.
What it is allowed to do
Three tools and nothing else. It can fetch a page, it can read a well-known file, and it can stop and state its answer. Every request is a GET. It cannot submit a form, sign in, add anything to a cart or spend money, because those verbs were never built.
- 01
fetch_pagefetch a page
- 02
read_well_knownread a well-known file
- 03
donestop, and say what it found
A run is capped at a dozen pages and under a minute. Running out is itself a finding: an agent that could not get there in twelve pages is an agent your customer would have abandoned.
It reads what your server serves, with no JavaScript: the first 6,000 characters of text and up to 80 links. That is roughly what an agent sees of you.
How a run ends
Six outcomes, and the last two are the reason this is not a boolean.
how far it got
- success
- It did the errand.
- partial
- It got most of the way, and something was missing.
- fail
- It could not.
- blocked
- Your site refused it.
- hallucinated
- It answered confidently, and the answer is on none of your pages.
- not_applicable
- You did not owe this one.
The answer nobody else finds
When an agent reports a price that appears nowhere on your site, we say so. Nothing in your markup is wrong, no checklist can see it, and the customer who acted on that answer acted on nothing. This is the finding that only exists because something tried.
A rate, not a tick
The same errand is run several times and reported as a frequency with the count beside it. Something with a temperature was asked to do a job, so one run is an anecdote. Three out of five is a different sentence from five out of five, and both are different from a tick.
From a failure to something to press
Every failure carries a reason code, and the code is the way back into the rest of the product: to the check that measures it, to the guide section that fixes it, and where the fix is a file, to the file itself, generated from what your site already publishes.
it stopped here
a reason code
- checks
- the check that measures it
- guides
- the guide section that fixes it
- artifact
- the file, written from your pages
- null
- sometimes: nothing to press
Sometimes the honest answer is that there is nothing to press. A dead link deep in a shop is a page somebody has to restore, and saying so is better than offering a button that generates a file which does not address it.
Then prove the fix
Run the errand again and the two recordings sit side by side: the agent failing at step three, and the agent getting past it. Nothing about that is a claim.
Beside the score, never inside it
The task success rate sits next to the AX score and is never folded into it. The 26 checks stay comparable across every site and every date only because nothing is ever added to them.