The agent draws, it does not retype
When you ask the agent something about your numbers, it reads them and then draws them: a headline figure, a table of the rows behind it, and a trend line where there is one. Its written reply sits above that and says what the numbers mean — what moved, what it implies, what to do next. This split is deliberate. A model asked to restate a forty-row table in prose does it at length, and occasionally retypes a number wrong. Handing it the rows and drawing them separately means the figures you read came out of the measurement, not out of a sentence.What it can draw
The agent does not write its own markup. It names a component from a fixed set, and the app draws it:
Two things follow from this. The agent can only produce layouts that have been designed, so a generated answer is held to the same visual standard as the rest of the app. And an app that has not shipped a component yet falls back to a plain-text description of what it would have said, rather than breaking the conversation.
Gaps are gaps, not zeros
A day nobody sampled shows as a break in the line, never as a drop to the floor. This is the same rule the dashboard follows. A mention rate ofnull means the prompt was not read that day; a mention rate of 0 means it was read and you were not mentioned. Those are different facts, and drawing the first as the second would invent a collapse that did not happen. The caption under a chart says how many days in the window were not measured.
If a chart looks sparse, that is usually sampling depth rather than a bad week. See Reading the numbers.
Copying and downloading
Every table and chart the agent draws carries Copy and Download.- Copy puts the rows on your clipboard tab-separated, which pastes straight into Sheets or Excel and still reads correctly if you paste it into a message instead.
- Download saves a
.csv. Charts save the rows behind the line, not a picture of it — a picture is not something you can pivot.
Seeing what it did
While the agent works, each step appears as it happens — reading your data, searching the web, fetching a page, changing what you track — each with its own mark, so a lookup is distinguishable from a crawl at a glance. Once the steps have finished they collapse into a single line, and the answer sits directly under the question. Click that line to open them again. Two exceptions never collapse: a step that failed, and a step waiting on your approval.Chats
How to ask a good question, and which agent model answers it.