Data analysis is most reliable when the table is structured and the question is measurable. Treat generated conclusions as analysis to review, not an automatic source of truth.
1. Prepare the table
Use clear column headers, one record per row, consistent units and dates, and remove decorative empty rows or merged cells when possible.
2. Ask for a schema check first
Before conclusions, ask ChatGPT to list columns, types, missing values, date coverage, suspicious duplicates and unit ambiguities.
3. Define the metric in plain language
Write exactly how the KPI should be calculated and which rows or periods should be excluded. Do not assume the model shares your business definition.
4. Request the analysis and the method
Ask for the table/chart plus the calculation logic, assumptions and any code-backed steps that matter to the result.
5. Spot-check the result
Recalculate a few important numbers manually or in the source spreadsheet, and investigate any result that conflicts with the raw data.
First inspect this dataset. Report columns, data types, date range, missing values, duplicates, unit issues and anything that could invalidate analysis. Do not draw business conclusions yet. Then wait for my metric definition.
Jumping directly to “what insights do you see?”; mixing percentages and raw counts; ignoring missing rows; interpreting correlation as causation; trusting a chart without checking its aggregation.
Finish check
The final output should include the metric definition, the result, the method used, limitations, and a small set of numbers you independently verified.