Testing one article the way the widget would render it
The Ask tab runs a real assistant turn and draws the answer with the widget's own renderer, inside the same isolated shadow root residents actually see.
Every other tab on an article shows you data. The Ask tab shows you the actual experience — because it runs a real assistant turn against your live content and draws the response with the exact renderer your widget uses in production.
This is a real turn, not a simulation
Typing a question here and pressing run calls startAssistantTurn for real. It costs real model spend, and the screen treats that honestly: instead of a warning before you click, it shows a receipt after — the exact cost in dollars, taken from the turn's own result rather than estimated in advance.
There's no keyboard shortcut to submit. Enter is a plain newline in the question box on purpose, because a control that spends your organization's model budget shouldn't be reachable by a key someone presses while still thinking.
The answer renders inside the widget's own shadow root
The result isn't drawn by the console's own styling — it's handed to the real widget preview module, which renders it inside an isolated shadow DOM exactly as it would inside your website's widget. Nothing in the console can style what's inside it, and nothing inside it can leak out and style the console. What you see is what a resident would see, not an approximation of it.
The question is answered in the article's language, deliberately
The turn is sent with the article's own declared language and flagged as if the resident had explicitly chosen it — not detected from the words you typed. Reviewing the Arabic variant of an article means asking, in effect, "what would an Arabic-speaking resident get," regardless of what language you happen to type your test question in.
If the answer panel stays empty with a small error message naming a URL, that's not a broken article — it means the widget's own build hasn't been deployed to this environment yet. It's a missing file, not a content problem.
What the outcome tells you
Beside the cost receipt, the result names its own outcome — whether it was answered from a retrieved passage, an override, or came back with nothing. That's the same three-way split you'd see in Retrieval preview, applied to one specific article rather than an open question.