How to analyze a CSV with ChatGPT, and when SQL is better

Attach the file to a conversation and tell ChatGPT exactly what to calculate.

1. Drop your .csv file into the ChatGPT message box, or pick it with the attachment button.
2. Name the exact columns you care about, how to group or filter them, and the output you want.
3. Ask ChatGPT to show the Python code it runs and to print the total row count first.

For data analysis, ChatGPT writes and runs Python in a Jupyter notebook environment. That environment cannot make web requests, so upload the file itself instead of pasting a link.

As of October 2026, OpenAI limits CSV and spreadsheet files to roughly 50 MB, depending on row size. The 2,000,000 token cap for text documents does not apply to spreadsheets. Upload counts are capped too:

• Every plan: up to 80 files every 3 hours, and failed attempts can count toward it.
• Free plan: 3 file uploads per day.
• Projects: 5 files on Free, 25 on Go and Plus, 40 on Pro, Business, Enterprise and Edu.
• Custom GPTs: 20 knowledge files.

For analysis, ChatGPT reads CSV, TSV, XLS, XLSX, JSON, XML, YAML, Markdown, TXT and PDF. Parquet is not on the list, but rows.page opens it.

OpenAI's own advice is to review the generated code, outputs and assumptions before you rely on a result. It also notes that ChatGPT may pick a method that doesn't match your intent on the first try.

• Ask for the code behind every step, so you can check the exact filters and aggregations.
• Have it print the row count right after loading, and compare that with the source file.
• Open the same CSV in rows.page. The Columns panel shows each column's type, distinct count, share of empty values, min and max, and top values, so wrong numbers stand out.

Attaching a CSV sends it to OpenAI's servers. Whether OpenAI may use it to improve its models depends on your plan and data settings; business offerings such as ChatGPT Enterprise and the API are excluded. If the file holds customer emails or financial records, check your company's data rules first.

In rows.page, DuckDB-WASM reads the file inside your browser. Nothing is uploaded while you scroll, filter or run SQL, only when you press Share.

If you already know the count, filter or group-by you need, write it yourself and get an exact answer you can rerun. For quick checks, use the filter language: country:NL, price:>20, name:* (has a value) or -name:* (empty).

For aggregations, type one DuckDB SQL statement that starts with SELECT, WITH or FROM. The table is called data:

SELECT country, count(*) AS orders FROM data GROUP BY country ORDER BY orders DESC

The Columns panel skips mean, median and standard deviation, so use avg(), median() and stddev(), or profile every column at once:

FROM (SUMMARIZE data)

SQL results show up to 1,000,000 rows. Window functions, QUALIFY and unnest() work, and column names with dots need double quotes, like "user.city".

When the CSV is over ChatGPT's ~50 MB cap, or you only want to send aggregates or a few columns to OpenAI, use both tools.

1. Drop the large CSV or Parquet file onto rows.page.
2. Filter the rows, or run a SELECT that keeps only the columns and rows ChatGPT needs.
3. Open Export and choose CSV. A filtered view exports every matching row, a SQL result up to 1,000,000 rows.
4. Or export Markdown, a table of the first 1,000 rows you can paste straight into a prompt.
5. Upload the smaller file to ChatGPT, so you stay under the cap and spend fewer uploads.