Ask your first question

Arcus answers business questions, not SQL. The art is in the question shape — a sharp question gets a sharp answer.

This guide covers what to ask, how to ask it, and how to read the answer.

Ask the way you'd ask an analyst

The fastest way to get a useful answer is to ask the question you'd ask a smart analyst sitting next to you. Specific. Time-bounded. With a clear unit.

Good questions:

• What was revenue last week, by country?
• Which Meta campaign drove the most conversion value in March?
• Compare ROAS on TikTok vs Meta in the UK over the last 60 days.
• What's our best-performing creative for women 18-24?
• Why did revenue drop between week 10 and week 12?

These work because they specify:

1. A metric (revenue, conversion value, ROAS).
2. A time window (last week, March, the last 60 days).
3. A dimension to break out by (country, campaign, audience).

Less-good questions:

• How are we doing? — too vague. Arcus will ask for clarification.
• Tell me everything about Q1. — too broad. The agent will narrow it itself, but you'll get a less-targeted answer.
• Show me the SQL for revenue. — Arcus is for the decision-maker, not the SQL writer. Ask for the answer instead.

Use plain business language

You don't have to know your warehouse schema. Arcus's semantic layer maps business words to columns. If your team calls something "GMV" instead of "revenue," that's in the synonyms — the agent picks it up. If a "campaign" is campaign_name in one table and meta_campaign_id in another, the relationships are mapped — you don't have to think about it.

If a synonym is missing, the agent says so. You can edit the semantic layer at any time. See Define a metric.

What you get back

Every answer renders as three layers, top to bottom.

Layer 1 — Narrative

Two to four sentences in plain English, with a confidence tag (high, medium, or low). Every number cites its source — hover over a figure to see the column, filter, and aggregation behind it.

Example:

Revenue last week was $184k, up 12% week-over-week. The lift came from UK (+$28k) and DE (+$11k), driven by the Hydrostars product line. US was flat.

Confidence: high. Source: mart.orders — 7 days, 47k rows.

Layer 2 — Chart

Auto-selected based on the shape of the answer:

• Single number → big-stat card with delta.
• Time series → line chart.
• Comparison across categories → bar chart.
• Cross-cut (e.g. country × channel) → stacked bar or grouped bar.
• Anything else → table.

You can swap the chart type if the auto-pick isn't right.

Layer 3 — Raw data

A collapsible section showing the result rows. The SQL is one click away — click "view query" to see exactly what Arcus ran. This is the audit trail: if the narrative is wrong, the chart is wrong, or you suspect a calculation, the SQL shows you.

Follow up

Arcus has memory of the conversation. After the first answer, you can ask:

• Why? — the agent investigates.
• Show me the same for last month. — same query, different window.
• Break that out by channel instead. — same metric, different dimension.

Each follow-up is a turn in the same thread. Threads are saveable, shareable, and pinnable. See Pin to Watch for how to keep an answer fresh.

Confidence tags

The narrative is tagged with a confidence:

• High — the answer is supported by multiple days of data, the metric is well-defined, the SQL is straightforward. You can act on it.
• Medium — there's a caveat. Maybe the time window is short, or the metric definition has a known edge case, or one source is sparse. The narrative explains.
• Low — best-effort. The agent ran out of confidence and is telling you so. Treat as a starting point for a deeper question, not a decision.

When confidence is low, the agent typically points at what would unblock it — "would you like me to widen the window to 30 days?" or "this requires the marketing semantic layer to define audience — please add it in Settings."

When the agent says "I can't answer this"

Sometimes the data isn't there. Maybe the warehouse only has 3 days of history. Maybe the question requires a metric the semantic layer doesn't define. Maybe two warehouses need to be joined and only one is connected.

In those cases, the agent calls flag_insufficient_data and tells you why. We never fabricate. If we can't answer, you'll know — and the answer card explains what would need to change.

What's next

• Understanding the answer card — the three-layer format in detail.
• Branch and edit-rerun — fork a thread to compare alternatives.
• Pin to Watch — turn an answer into a refreshing brief.