Question types
Specific metric questions
This is the most common kind of question. One query runs, and you get an answer you can check against the SQL. The best questions name three things: the metric, the time range, and the breakdown. “Revenue last month by week” is a complete question. “How are sales?” is not.Raw data pulls
Use pull or export language when you want the data itself rather than an interpretation. Name the fields, the time range, and any filters such as product, campaign, store, or channel. Data pulls always run as a single query. They skip Deep Analysis and never produce a chart. The reply confirms what was pulled and attaches the rows as a CSV. Asking for several dimensions does not make a question a data pull. “What is revenue by UTM source and medium?” asks what the data means, so it gets an analysis. “Pull order ID, UTM source, UTM medium, and revenue for last week” asks for rows. Data pulls return at most 100 rows. If you get exactly 100, assume the result was cut off and narrow the request.Open-ended questions and Deep Analysis
Questions such as “Why are our metrics declining?” or “What should we focus on to grow?” have no single number as an answer. For these, the analyst runs Deep Analysis. Deep Analysis runs two analyses: your question, plus one strategic question the analyst writes to add a second angle. Each shows up in the thread with its own title. The final reply combines both into one answer with recommendations, and it carries the reasoning and feedback actions. Deep Analysis takes longer than a single query. If you want a fast, specific answer, phrase the question specifically. “What was our Meta ROAS last month?” answers faster than “How is Meta performing?” A single question never runs longer than 7 minutes.Definitions and schema
Questions about what a metric means, how it is calculated, or where data lives are answered without running SQL. Definitions come from SourceMedium documentation. Schema questions are answered from your warehouse’s table metadata, and the reply includes Schema docs links to the table pages. “What is LTV?” gets a definition. “What was our LTV last month?” gets an analysis.Data and attribution health
“How is my data health?” reports which tables are fresh, which are empty, and which domains are ready to analyze. “How is my attribution health?” audits UTM coverage on your acquisition orders. See Data and attribution health.Follow-ups
In the same thread you can ask for a new breakdown, a summary, or a recommendation. Summaries and recommendations use what the thread already found and do not run new SQL. If you ask for one before any analysis has run, the analyst asks you to run an analysis first. If your request is ambiguous, it runs a fresh analysis instead. Thread rules, including when to mention the bot, are on Reading results.Example prompts by domain
Copy one, then swap in the metric, time range, and breakdown you care about.Orders and revenue
Customers
Marketing and ads
Email and SMS
Products, subscriptions, and returns
Web analytics and attribution
Raw data pulls
Definitions and schema
Health checks
Tips for better answers
State a time range
State a time range
Say “last 30 days,” “in January,” or “Q4 2024.” If you leave it out, the query picks one, and you will see it in the SQL. “Recently” is not a time range.
Name the metric you mean
Name the metric you mean
“Revenue” is clearer than “sales.” “Orders” is clearer than “transactions.” Use the terms from your dashboards. See Metrics.
Ask one thing at a time
Ask one thing at a time
“What was revenue and how did AOV change?” works better as two questions in the same thread.
Use top N for rankings
Use top N for rankings
“Top 10 products by revenue” is specific. “Best products” is not.
Use pull or export language for rows
Use pull or export language for rows
“Pull order ID, date, UTM source, and revenue for last week” returns rows and a CSV. “What are the trends in orders by UTM source?” returns an analysis.
Add context to open-ended questions
Add context to open-ended questions
“How can we improve Q1 performance given the new product line?” gives Deep Analysis something to work with.
Check data health first when it matters
Check data health first when it matters
A “last 7 days” answer is wrong if the last 3 days have not synced. Ask “How is my data health?” before an analysis you plan to present.
Related
Reading results
What each part of a reply means, thread rules, and how to validate a number.
Table schemas
Column-level documentation for every table the analyst can query.