Working with files, spreadsheets, and data in ChatGPT
PDFs, spreadsheets, charts, calculations, parsing failures, and reproducible checks.
ChatGPT guide · as of July 10, 2026 · 4 minutes read · details change — confirm current specs on chatgpt.com
ChatGPT can read uploaded documents, inspect structured data, run calculations in a code environment, create charts, and produce downloadable files. This is one of its most useful capabilities—and one of the easiest places to accept a wrong answer because the result looks analytical.
The safe habit is to verify what was loaded before discussing what it means.
What it handles well
| Input | Useful work | Common failure |
|---|---|---|
| CSV or spreadsheet | Cleaning, grouping, calculations, charts | Wrong sheet, range, type, or filter |
| Text PDF | Summary, extraction, comparison | Missed footnotes, columns, or appendices |
| Scanned PDF | Limited extraction after visual or OCR processing | Misread text and tables |
| Word or text file | Outline, rewrite, terminology analysis | Formatting and comments may be lost |
| Images | Read labels and inspect visible patterns | Tiny text and spatial detail can be wrong |
| Multiple files | Compare versions and synthesize | Conflicting versions or retrieval gaps |
OpenAI's data analysis guide recommends clean structured data with clear column names and one record per row.
Prepare the file first
For spreadsheets:
- Use one header row.
- Give columns unambiguous names.
- Keep one type of value per column.
- Remove merged cells where possible.
- Explain units and date formats.
- Identify the authoritative sheet.
- Preserve a read-only original.
For long documents, name the version and date. Tell ChatGPT whether appendices, footnotes, tracked changes, and comments matter.
Do not upload private data simply because analysis is convenient. Remove unnecessary names, account numbers, and identifiers.
Begin with an intake report
Before asking for conclusions, request:
List every uploaded file, sheet, row count, column count, date range, missing-value count, and field type. State anything you could not parse. Do not analyze trends yet.
This exposes many silent failures.
| Validation | What it catches |
|---|---|
| File and sheet names | Wrong attachment or tab |
| Row count | Truncated or filtered data |
| Column types | Numbers imported as text |
| Date range | Missing months or wrong date parsing |
| Null and duplicate counts | Quality problems hidden by averages |
| Units and currency | Invalid comparisons |
| Sample rows | Shifted columns or broken headers |
Ask for reproducible analysis
A useful prompt defines the population, filters, calculation, grouping, and output.
Example:
Using only the 2025_Data sheet, calculate commission revenue by department. Exclude rows whose status is Completed, Canceled, or Uncollectible. Show the exact status filter, row count before and after filtering, formula, and a table sorted by revenue. Save the cleaned data and summary as separate files.
Ask ChatGPT to keep calculation and interpretation separate. First establish the numbers. Then discuss possible meaning.
Inspect code and intermediate results
ChatGPT's data-analysis environment can use code, which is usually more dependable than mental arithmetic. Code can still implement the wrong logic.
Ask to see:
- The executed code or calculation steps.
- Filtered row counts.
- Grouped intermediate tables.
- Treatment of blanks and duplicates.
- Any assumptions about dates, signs, units, or categories.
- Reconciliation totals against the source.
A chart is not validation. A visually plausible chart can be based on a wrong denominator.
PDFs need special caution
Text extraction can scramble multi-column pages, tables, headers, and scanned content. Ask ChatGPT to cite page numbers and quote the exact source text for important claims.
When a table or figure matters, inspect the page image rather than relying only on extracted text. For contracts and financial reports, compare the answer against the original page.
Creating files
ChatGPT can create spreadsheets, documents, presentations, PDFs, charts, and other artifacts in supported modes. Download and open the result.
Check:
- Formulas rather than only displayed values.
- Broken links or missing images.
- Sheet names and hidden rows.
- Fonts, page breaks, and print area.
- Accessibility and alt text.
- Whether the file opens in the software your audience uses.
Store the final file somewhere durable. A chat attachment is not your records system.
When this is the wrong tool
ChatGPT is the wrong tool when the analysis must be formally audited, the source cannot leave a controlled environment, or a silent error could trigger payroll, tax, legal, clinical, or investment action.
Use a validated spreadsheet model, database, statistical package, or qualified analyst. ChatGPT can help write and review the process, but it should not be the only control.
