Opening the chat
From the Agents list, click on any agent to open its chat interface. The chat is where you interact with your agent in real time — ask questions, explore results, and follow up with more questions.Asking questions
Type your question in plain language at the bottom of the screen and press Enter (or click Send). Good questions to start with:- “How many sessions did we have last week, broken down by channel?”
- “Which campaigns had a ROAS above 3 in the past 30 days?”
- “Show me the top 10 highest-spend ad groups from last month.”
While the agent is working
A complex question can take a while — the agent explores your schema, runs queries, and builds visualisations before it answers. Three things help you follow along:- Thinking — where the agent reasons before acting, you’ll see a collapsible Thinking block that becomes Thought for Ns once it’s done. Expand it to see how the agent approached your question. Reasoning is kept with the message, so it’s still there when you reopen the chat.
- Tool activity — each query, chart and report shows up as a step in the tool panel, with progress while it runs.
- Stop — the send button becomes a stop button while a turn is running. Stopping keeps whatever the agent had already produced, so a partial answer isn’t lost.
A dropped connection no longer kills the turn. If you close your laptop, lose Wi-Fi, or navigate away, the agent keeps working server-side and the chat reconnects and picks up where it left off — including anything it finished while you were away.
Reading results
Depending on your question and the agent’s configuration, results can appear in several formats:Tables
Data is presented in a scrollable table with sortable columns, opened in the right-side panel. Columns can be styled by type — currency, percentage, change vs. previous period, badges, tags, ratings. Column headers show the field names from your BigQuery data — if these aren’t clear, add descriptions to your agent context.Charts
When you ask for a visualisation (or when the agent decides a chart would be more useful), results are rendered as an interactive chart directly in the chat. Three chart types are available:- Time series — a metric over time, as lines, bars or both, optionally with two Y-axes, stacked bars, or a 100% stacked share-of-total view
- Ranked bar — horizontal bars ranked by value, for top-N comparisons where the category names need to stay readable
- Scatter — two numeric metrics against each other, with optional bubble size, colour per group, and a trendline
“Can you show that as a ranked bar chart instead?”Which visualisation to use, and how to ask for it →
Reports
For multi-part answers, the agent can build a structured report — sections of narrative interleaved with KPI tiles, charts and tables — which opens in the right-side panel and exports to PDF. Reports are worth asking for explicitly when you want a layout you can share:“Turn that into a report with an executive summary, a KPI row, and a chart per market.”
Markdown text
For summaries, insights, and explanations, the agent responds in formatted text with bullet points, bold highlights, and recommendations.Following up
Every response is a starting point. You can ask follow-up questions that build on the previous result:“Great — now filter that to campaigns where spend was over €1,000.”
“Can you add a column showing the change vs. the previous period?”
“What’s the average CPA across all those campaigns?”The agent maintains context within a chat session, so you don’t need to repeat yourself.
Rating a response
Any agent response can be rated, and the rating is what feeds back into improving the agent.- Thumbs up marks the answer as a good example.
- Thumbs down opens a short What went wrong? form — pick a category (wrong table, wrong metric, missing filter, wrong date range, made-up data, or other) and optionally describe what the agent should have done instead.
Agents record their own learnings as they work — you’ll see it happen as a step in the tool panel. There’s nothing for you to save by hand. Note that neither a learning nor your feedback changes the agent’s behaviour on its own: to make a correction permanent, it has to go into the agent’s context or instructions.
Chat history
Every chat session is stored permanently and accessible from the chat history tab. If the chat was created by a scheduled job, it’s also linked to that job’s run history. Job run chats are read-only — they are a fixed snapshot of what the agent returned for that run. To continue the conversation from a job result, use the Start Follow-up Chat button in the input area. This creates a personal copy of the chat that you can continue freely. See Personal follow-up chat for details.Tips for better results
Be specific about time periods
Be specific about time periods
Agents use
CURRENT_DATE as a reference point. Phrases like “last month”, “this quarter”, and “last 7 days” are interpreted relative to today. If you mean a specific date range, say so explicitly.Name your campaigns or segments exactly
Name your campaigns or segments exactly
If you’re asking about a specific campaign, use the exact name or ID as it appears in BigQuery. You can ask the agent first: “What campaigns are currently active?” to get the exact names.
Say how you want it shown
Say how you want it shown
The agent picks a visualisation from what your question implies. If you already know what you want to look at, say so — “as a ranked bar”, “revenue in euros on the left axis”, “as a share of total”, “with the values on the bars”. See Charts, tables & KPIs.
Ask for one thing at a time
Ask for one thing at a time
Complex multi-part questions sometimes produce one part well and miss others. Break big analyses into steps for more reliable results.
Correct the agent when it's wrong
Correct the agent when it's wrong
If the agent misunderstands something, correct it in plain language: “Actually, the
revenue column is in euros, not cents.” The agent applies it for the rest of the conversation, and may record it as a learning. To stop it happening in future chats, add the correction to the agent’s context.