The evidence chain
Evidence is what we saw, linked to the fix it supports. Every fix in your fix list points back to the page response our checker received: the address, the status code, any redirects and the part of the page or file the check looked at. Each result records the date and the version of our checks, so you can compare two results measured the same way.
Today the evidence comes from the automatic site check. Our checker identifies itself as AiscoveryAudit, reads public pages at most one page per second and follows robots.txt. It doesn't run JavaScript, log in or submit forms, so its evidence is the page response, not a picture of the page.
Coming later: evidence from key task test runs. When AI agent task tests are released for your plan, each test run keeps a step log, step screenshots, step timings and the list of pages the agent requested. The run also records the exact agent and model, so a result is never attributed to an agent that didn't run.
Step screenshots
Coming later. We don't record video, and there is no session replay. Instead, each test run keeps:
- a step log: each page the agent opened, each link it followed and each site search or filter it used;
- a step screenshot for each step, taken in our controlled browser with sensitive fields masked before the image is stored;
- the literal checks each step was held to and whether they passed;
- the point where the agent stopped, such as before cart or before a form that sends data.
A screenshot alone can't prove that a goal was reached. Literal checks on the page text decide the result. An AI check can read the same evidence and confirm the result or flag it for review, but it can never turn a failed check into a pass.
Completion time
Coming later. Each step of a test run records when it started and how long it took, so a team can see where time accumulated:
| Measure | What it explains |
|---|---|
| Time per step | How long each page, link, search or filter took. |
| Page response time | How long your website took to answer each request. |
| Total run time | The end-to-end cost of the task for that agent. |
Compare timing only across the same task, the same agent and the same website environment. A fast run that didn't reach the goal is still a run that didn't pass.
Layout shifts
Coming later. During a test run our browser measures Cumulative Layout Shift on each page the agent visits, using the browser's standard session-window algorithm. It is a bounded lab measurement in one browser, not field data from your visitors, and a measurement the browser can't support is shown as unsupported, never as zero.
We don't yet measure how far the agent's intended target moved around each click. Field CLS from your own visitors, lab CLS from our browser and target movement stay separate values.
Redaction and retention
Evidence is only useful when it is safe to review.
- Site check evidence holds page responses from your public pages only.
- Step screenshots mask password, payment and personal-data fields before they are stored. Coming later, with AI agent task tests.
- AI agents stop before login, cart, checkout, payment and any form that sends data, so test runs don't create customer, payment or account data.
- Evidence is kept for a limited period, and access to it follows your workspace's roles.
If masking would remove what a fix depends on, we keep the evidence restricted and say so in the result.