ChatGPTData analysisLevel 2: Context

How Do You Analyze a CSV File in ChatGPT?

Upload the file with the paperclip, ask ChatGPT to describe what it sees without drawing conclusions, correct anything it got wrong, and then ask one question per message. OpenAI's Academy says ChatGPT writes and runs Python code to do this, so answers come from calculation on your data.

Prompt to copy
First message, with the file attached:

This is [what the data is, such as a CRM export of 2026 deals]. Don't analyze it yet. Tell me:
1. How many rows and columns you see, and what each column seems to hold
2. Which columns look like IDs, dates, categories or amounts
3. Any quality problems: blanks, duplicates, mixed formats, summary rows mixed in with the data
4. Five rows you find strange

Then stop and wait for my corrections.

Next message:

Question: [one specific question, such as which region's average deal size fell most from Q2 to Q3]. Say how you're defining [the key metric], which rows you excluded and why, and show a small table of the numbers behind your answer.
  1. 1Upload the file
  2. 2Ask what's in it
  3. 3Check it read correctly
  4. 4Ask one question
Next question, in its own message
Based on OpenAI's dataset use case: inventory first, conclusions later. One question per turn is a rule of thumb.

Try it now

  1. 1In New chat, click the paperclip, choose a real CSV or XLSX file, and send the first message above.
  2. 2Compare its row count and column types with what you see when you open the file. Correct anything wrong.
  3. 3Ask one question. It worked if you can match a total in its answer to a figure you already trust, like a monthly total from your finance report.

Use it when

Exports you need to understand quickly: sales by region, survey results, campaign numbers. OpenAI's Academy gives role-based examples, such as comparing email campaign open rates over time or charting employee attrition by quarter and department.

Skip it when

Data you're not allowed to share, such as personal records with no approval from whoever owns them. OpenAI's pricing page lists content on personal plans as used for training unless you opt out. And for a repeating, fully automated pipeline, OpenAI's Academy points to the API instead.

Why it works

A misread column at the start spoils every later result, so check the reading first. OpenAI's dataset use case tells you to ask what each file contains, which columns look like identifiers, dates or locations, and where the quality problems are. It says not to ask for conclusions yet.

Asking one question at a time keeps each result checkable. If you ask for five analyses at once and one total looks off, you can't see which step caused it.

The feature varies by plan. OpenAI's pricing page lists **Data analysis and File uploads as limited on Free, and included on Go, Plus and Pro. Interactive tables and charts** start at Plus. Older guides call this tool Advanced Data Analysis or code interpreter. OpenAI's Academy adds that reasoning models give the best results.

Example

Say you upload a 2026 sales export. ChatGPT reports 12,400 rows and notes that "Date" mixes two formats and the last row is a grand total. You tell it to drop that row and read dates as MM/DD/YYYY. Only then do you ask: "Which three products lost the most revenue between Q2 and Q3?" Without the correction, that summary row would have been counted as a sale and inflated the numbers.

Going further

Big or complicated files fail in quiet ways. OpenAI's Academy says to split very large files or focus your prompt on specific sections. No file size or row limits appeared on the pages that could be opened on 6 November 2026, because OpenAI's Help Center refuses automated reads, so check there if a file won't upload.

For several files, OpenAI's use case says to profile the merge before joining: check whether candidate keys are unique, look at null rates and formatting differences, and run trial joins that report match rates. A join with a low match rate can quietly drop rows. To get an interactive chart, type @Visualize in the composer if it's available on your account.

Common mistakes

  • Opening with "analyze this". You'll get general patterns from a file ChatGPT might not have read correctly.
  • Asking five questions in one message, then being unable to trace which step produced the odd total.
  • Uploading a screenshot of a table when the spreadsheet is available. The spreadsheet gives it real values to compute with.
  • Ignoring summary rows. A grand total at the bottom can be counted as data and double your sums.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course