Data Analysis with ChatGPT

Start by describing your data

Before asking for analysis, describe what you have. State how many rows and columns there are, what each column means, and what time period the data covers. If you are working with a file you can upload, check the app's current options for uploading and analysing files, since features and file handling change over time. If you cannot upload, paste a small sample of rows and describe the rest.

Clean-looking questions come from clean descriptions. Saying "analyse my sales" is too vague to produce a useful answer. Saying "this file has monthly sales by product from January to December, with columns for product, region, and revenue" gives ChatGPT something concrete to work with, and it lets you spot immediately if it has misunderstood a column.

I have a CSV of online orders with columns: order_date, product, category, quantity, price, and region. Summarise total revenue by category for the last 6 months, list the top 5 products by revenue, and flag any month where revenue dropped by more than 20% compared with the previous month.

Ask one question at a time

Good data questions are specific. Instead of asking for "insights," ask for a total, a comparison, a ranking, or a change over time. Each of those has a clear answer you can check. For example, ask which category grew fastest this quarter, which region has the highest average order value, or which products appear most often in orders over a certain amount.

Work through questions in order. Start with totals and averages to get your bearings, then look at breakdowns by category or region, then investigate anything unusual. This order matters because a surprising breakdown is easier to judge once you know the overall picture. If a breakdown contradicts the total, that is usually a sign to check the data or the calculation, not to accept the surprising number.

Clean messy data before you trust the results

Real data is rarely tidy. Dates may appear in several formats, product names may be spelled inconsistently, and some rows may be blank or duplicated. These problems distort every calculation built on top of them, so cleaning comes first. Ask ChatGPT to help standardise a column, for example by converting all dates to one format or grouping spelling variants of the same product under one name.

Keep a note of every cleaning decision. If you merged two product names or removed duplicate rows, write it down, because it affects how the results should be read. A total that excludes duplicates is not the same as a total that includes them, and anyone reading your report should be able to tell which one they are looking at.

Turn results into charts and plain summaries

Numbers alone are hard to act on. A short summary in plain language, such as which category drives most revenue and whether the trend is rising or falling, is often more useful than a long table. Ask ChatGPT to write that summary for a specific audience, for example a manager who needs the headline and one supporting detail, not a full breakdown.

Charts help when the story is about change over time or comparison between groups. Ask for a simple chart of one thing at a time, such as revenue by month, rather than a chart that tries to show everything. A chart with too many lines or categories is harder to read than the table it replaced. Check that the chart matches the numbers in the table before sharing it.

Verify before you decide

ChatGPT can make arithmetic mistakes and can misread a column, especially in large or messy files. Treat every result as a draft. Recalculate key totals in your spreadsheet, spot-check a few rows by hand, and be suspicious of numbers that look too round, too large, or inconsistent with what you already know about the business.

This verification step is the difference between exploration and reporting. Exploration can be fast and rough; a report that will be shared or used for a decision needs checked numbers and a clear note of the data source and date range. If you cannot reproduce a number from the source data, do not include it, no matter how plausible it sounds.

Finally, remember that analysis answers what happened, not why it happened. A drop in sales in one month might reflect a holiday, a price change, a stock problem, or a data gap. ChatGPT can suggest possible explanations, but confirming the cause means checking outside the data, such as talking to the people involved or reviewing what changed that month.

Frequently asked questions