Meet the assistant
The assistant is the front door of the platform. Ask a question in your own words and the answer arrives with a panel of evidence beside it.
When you sign in you land on the assistant. There is a composer, a few suggested questions with live counts from your own workspace, and nothing to configure. You type a question and the answer arrives as prose with a panel of evidence open beside it.
What you'll learn
- What is on the screen when you arrive, and what each part is for
- What happens between your question and the answer
- The four keyboard shortcuts worth knowing on day one
What you land on
The composer. One box. The placeholder rotates through questions that already work on your workspace, so the example is never a question your account cannot answer.
Suggested questions. Each pill carries a count that came from your own data, for example "11 lost" or "46 open". The pills rotate through alternates, and a pill never offers a question that would open an empty panel. If you have no competitors on file, you are not offered a head-to-head question.
The last run. If an overnight agent run left something readable behind, it appears as a card you can open. See the last run.
The top bar. Two menus. Track is what you monitor: Prompts, Competitors, Changes and Shopping. Analyse is what it means: Intelligence, Trace and Reports. Rows that have a panel open it beside the conversation instead of navigating away.
What happens when you ask
An answer is not a database query dressed in prose. The turn runs in phases, and the app shows you which ones ran:
- Ground. Read your workspace and work out what you are asking about. If you named a prompt, a competitor or an engine, that becomes the focus of the turn.
- Retrieve. Fetch the counted rows that bear on it, for example the stored answers, the lost prompts, or one rival's head-to-head.
- Analyse. Work out the finding. For some questions this is arithmetic rather than a model call, because computing per-engine rates is the analysis.
- Decide. Choose what to say and which panel to open.
Each step reports what it did once it has done it, with the count attached: "Checked 6 platforms, 1,203 answers". A phase that does not run shows nothing, so a greeting does not wear a four-step trace it never earned.
The assistant answers from your workspace's stored, verified answers. It is not asking ChatGPT about you live while you wait. The monitoring runs on its own cadence, and the assistant reads what that monitoring produced. That is why the answer arrives with counts and dates attached.
The panel beside the answer
Every class of question opens a panel suited to it: Prompts, Sources, Proof, Platforms, Head to head, This week, Readiness, or the last run. The panel is not a picture of the answer, it is the evidence, and you can descend it: landscape, then one subject, then the proof behind that subject.
Each panel is also a page. The arrow in the panel's corner opens the full-page version of the same thing, which is what the Track and Analyse menus link to. Two hosts, one instrument. See panes and the ladder.
Naming your own things with @
Type @ in the composer to mention one of your prompts, competitors or engines. The mention travels as
an id rather than as text, which is what lets the answer scope itself exactly to the row you meant.
Without it the assistant has to find your entity by name in the sentence, and a rival whose name you
abbreviated will be missed.
Four shortcuts
| Keys | What it does |
|---|---|
⌘K | Shortcuts: search pages and prompts, or ask anything |
⌘↵ | Send the question |
⌘\ | Hide or show the evidence panel |
@ | Mention a prompt, competitor or engine |
Try this in Outercite
Open the assistant and ask "What did you find overnight?" If an agent run has completed, you get the brief and the run's own evidence. If it has not, the assistant says so rather than inventing a summary.
Related
What you can ask
The classes of question that work, and the panel each one opens.
Panes and the ladder
Landscape, subject, proof: the same three steps in every lens.
When a number is missing
Why the assistant says "not measured" instead of showing you a zero.
Last updated on
Why AI search matters
Buyers ask AI before they open a search engine, and an AI answer names one or two businesses instead of listing ten. Here is what that changes.
How verification works
What happens between an AI engine's answer and a counted citation: the prefilter, the judge, the verifier that audits it, and the rules that decide a tie.