You may have a 60-page contract, report, invoice packet, meeting record, or research document and need only a few facts from it. Reading the entire file and copying every name, date, and number by hand is slow. Asking ChatGPT to “summarize this document” is not the best solution either, because a summary may leave out details that look minor to the AI but matter to you.
Extract names, dates, amounts, and other important details from your documents in a structured format, making it easier to review, verify, and export the data.
The better method is to request structured data extraction. Tell ChatGPT exactly which items to find, how to organize them, what wording to preserve, and what to do when information is missing or uncertain. Then verify the result against the source before using it in a spreadsheet, report, contract review, or business decision.
Quick answer: Upload the document, name the exact fields you want, request a table or CSV, require page references, and add: “Do not guess. If a value is not clearly stated, write Not found.” This produces a much more useful result than asking for a general summary.
Quick Overview
| Goal |
Best Output |
Key Instruction |
| Read the result on screen |
Markdown table |
Use clear column headings |
| Move the result into Excel or Google Sheets |
CSV in a code block |
Return one record per row |
| Preserve legal or official wording |
Verbatim table |
Copy the exact wording without paraphrasing |
| Check where each fact came from |
Table with page or section reference |
Include the source location for every item |
| Reduce invented or uncertain values |
Verified extraction |
Write “Not found” instead of guessing |
Why a Summary Often Misses the Details You Need
A summary is designed to compress a document into its main ideas. That makes it useful when you want the big picture, but it can be unreliable for tasks such as collecting every deadline, invoice number, employee name, contract amount, or phone number.
Data extraction has a different goal. Instead of asking, “What is this document about?” you ask, “Which specific facts appear in this document, and where do they appear?” In AI terminology, names, organizations, dates, addresses, and identification numbers are often called entities. Pulling them into a consistent format is sometimes called entity extraction or structured extraction.
Simple distinction
- Summarize: Reduce the document to its main points.
- Extract: Find selected facts and place them into a defined structure.
- Analyze: Interpret, compare, calculate, or draw conclusions from those facts.
For better accuracy, extract first. Analyze second.
Start by Defining Exactly What You Want
“Extract the important information” is too vague. Different readers consider different facts important. A contract manager may need party names, effective dates, renewal terms, and payment amounts. An office administrator may need names, email addresses, phone numbers, and appointment dates.
Before uploading the file, write a short field list. Use the same labels you want in the final table.
Extract the following fields from this document:
- Full name
- Organization
- Date
- Amount
- Reference number
- Page number
Return the results as a table.
Use one row for each separate record.
Do not guess.
If a field is not clearly stated, write "Not found."
This prompt gives the model a schema—a fixed set of columns that defines how the answer should be organized. A clear schema reduces messy output and makes missing information easier to spot.
Use the Right Output Format
Choose a Markdown table for quick reading
A Markdown table is usually the easiest format to review inside ChatGPT. It keeps the fields aligned and lets you scan the results before copying them elsewhere.
Extract all people, dates, and monetary amounts from the document.
Return a Markdown table with these columns:
Category | Extracted Value | Related Context | Page
Keep the wording concise.
Do not combine separate records.
Do not infer missing values.
Choose CSV for Excel or Google Sheets
CSV stands for comma-separated values. It is a plain-text format that spreadsheet programs can place into rows and columns. CSV is often the better choice when the result contains many records or when you plan to sort, filter, calculate, or share the data.
Extract the requested information and return it as CSV inside one code block.
Use these headers:
Full Name,Organization,Date,Amount,Reference Number,Page
Rules:
- Use one record per row.
- Put quotation marks around any field that contains a comma.
- Keep dates exactly as written in the document.
- Do not add commentary before or after the CSV.
- If a value is missing, leave the field blank.
- Do not guess.
After ChatGPT returns the code block, copy only the CSV content. Paste it into a blank text file and save it with a .csv extension, or import it directly into your spreadsheet program. Review the columns before relying on the data.
Important details can be extracted from a document, reviewed in a table, and then exported as a CSV file for use in Excel or Google Sheets.
Ask for Verbatim Extraction When Wording Matters
Verbatim means copying the original wording exactly as it appears. This matters for contract clauses, policy statements, legal names, product codes, quotations, and formal deadlines. Without that instruction, ChatGPT may shorten or normalize the text.
Extract every deadline and notice period from the document.
For each item, provide:
1. The exact wording verbatim
2. The normalized date or time period, only if it is unambiguous
3. The page or section
4. A short label describing what the deadline applies to
Do not rewrite the quoted text.
Do not calculate a date unless the document provides enough information.
If the meaning is uncertain, mark it "Needs review."
Keeping the original text and the normalized value in separate columns gives you both accuracy and convenience. You can sort the normalized dates while retaining the source wording for verification.
Copy-and-Paste Prompts for Common Extraction Tasks
Extract names
Extract every person's full name from this document.
Return a table with:
Full Name | Role or Relationship | Organization | Page
Preserve spelling exactly as written.
Do not merge people with similar names.
If the role is not stated, write "Not stated."
Extract dates and deadlines
Extract every date, deadline, renewal date, appointment date, and notice period.
Return a table with:
Original Text | Date Type | Normalized Date | Related Event | Page
Keep the original wording.
Use YYYY-MM-DD for the normalized date only when the date is unambiguous.
If the year or exact date is missing, write "Incomplete date."
Do not guess.
Extract numbers, IDs, and amounts
Extract all monetary amounts, percentages, quantities, invoice numbers, order numbers, tracking numbers, account references, and contract IDs.
Return a table with:
Type | Exact Value | What It Refers To | Page
Keep punctuation, currency symbols, leading zeros, and letter-number combinations exactly as written.
Do not perform calculations.
Do not treat page numbers as extracted data.
Extract contact information
Extract all contact information from this document.
Return a table with:
Name | Organization | Email | Phone | Mailing Address | Page
Keep the original formatting.
Do not create missing email addresses or phone numbers.
Place each person or organization on a separate row.
Include Source Locations for Every Result
Page references make the output auditable. They let you return to the source and confirm whether a value was copied correctly and whether nearby text changes its meaning.
For documents with reliable page numbering, request the page number. For web pages, exported chats, or documents without stable pages, ask for the section heading, paragraph opening, table name, or another identifiable location.
For every extracted item, include the most precise source location available.
Use:
- Printed page number when visible
- PDF page number when no printed page number exists
- Section heading when page numbers are unavailable
If the location cannot be determined, write "Location unavailable."
Do not invent page numbers.
✔ Check before moving on: Every important value should have a source location or an explicit note that the location could not be identified.
Use an Extract → Verify → Export Workflow
Even a well-written prompt does not make AI output automatically correct. The safest practical workflow has three stages: extract the facts, verify them against the document, and export only after the verification pass.
- Extract: Ask for a structured table with fixed columns.
- Verify: Ask ChatGPT to recheck each row against the source and flag uncertainty.
- Export: Convert the verified table to CSV for Excel or Google Sheets.
Recheck the extraction against the uploaded document.
For every row:
- Confirm that the value appears in the source
- Confirm that the page or section is correct
- Flag spelling, date, or number mismatches
- Mark uncertain items as "Needs manual review"
- Do not add any new facts during this verification pass
Return the corrected table only.
A second AI pass can catch formatting problems, but it is not a substitute for your own review. Check high-impact items such as payment amounts, legal names, deadlines, bank details, medical information, and contract numbers directly against the source.
A clear, specific prompt helps ChatGPT extract information in a consistent format, making the results easier to review and use.
Before and After: A Better Extraction Request
| Weak Request |
Better Request |
| Find the important information. |
Extract names, dates, amounts, reference numbers, and page numbers into fixed columns. |
| Summarize all dates. |
List every date verbatim, identify what it refers to, and normalize it only when unambiguous. |
| Give me the numbers. |
Separate amounts, percentages, quantities, IDs, and page numbers so unrelated numbers are not mixed together. |
| Put it in a spreadsheet. |
Return valid CSV with one record per row and no text outside the code block. |
| Fill in missing details. |
Write “Not found” or “Needs manual review.” Do not infer or invent missing values. |
The improved version controls both the content and the format. That is the main reason it produces a more reliable result.
What to Do With a Very Long Document
Very long files can be harder to process consistently, especially when important details are scattered across many pages. Do not assume that one response captured every occurrence.
For a large document, split the job into sections or page ranges. Use the same extraction schema for each batch, then ask ChatGPT to combine the tables and remove exact duplicates. Keep a batch or page-range column so you can trace each record.
Process pages 1–25 only.
Extract:
Full Name | Date | Amount | Reference Number | Source Page
Use the same rules as before.
Do not include information from other pages.
At the end, report the number of rows extracted.
Combine the extraction tables from all batches.
Rules:
- Preserve every source-page reference
- Remove exact duplicate rows only
- Do not merge records merely because the names look similar
- Flag possible duplicates in a separate "Review Note" column
- Return the final result as a Markdown table
This approach is sometimes called chunking, which simply means dividing a large task into smaller parts. The goal is not to change the document but to make coverage easier to inspect.
Common Mistakes That Reduce Accuracy
Asking for too many unrelated tasks at once
Extraction, summarization, comparison, calculation, translation, and opinion are different tasks. Combining all of them in one prompt makes the output harder to verify. Extract the source facts first, then start a separate analysis step.
Failing to define what counts as a number
A document may contain page numbers, footnote numbers, dates, quantities, percentages, account numbers, and monetary values. Tell ChatGPT which categories to include and which to ignore.
Allowing the AI to “complete” missing information
Some models try to be helpful by filling gaps. Prevent this by stating that missing fields must remain blank or be labeled “Not found.” Never request guessed contact details, dates, or identifiers.
Changing the format halfway through
If one batch uses different column names or date formats, combining the results becomes difficult. Reuse the same prompt and schema for every section.
Ignoring leading zeros and exact punctuation
Codes such as 00127, AB-0042, or 05-08 can be corrupted when treated as ordinary numbers. Ask ChatGPT to preserve exact formatting and import code-like values into a spreadsheet as text.
Troubleshooting Extraction Problems
| Problem |
Likely Cause |
Fix |
| Names or numbers are missing |
The request was too broad, or the file is very long |
Process smaller page ranges and request an item count |
| Dates are reformatted incorrectly |
The model normalized ambiguous dates |
Keep an Original Text column and normalize only unambiguous dates |
| Several facts appear in one row |
The record rule was unclear |
State “one separate occurrence per row” |
| CSV opens in the wrong columns |
Commas inside text were not quoted |
Request valid CSV with quotation marks around fields containing commas |
| Page references do not match |
Printed and PDF page numbers differ |
Specify which numbering system to use and verify manually |
| The document appears unreadable |
It may be a scanned image rather than selectable text |
Use OCR or export a searchable copy before extracting |
ChatGPT supports common document and data file formats, but available tools, limits, and behavior can vary by plan, workspace, device, and product updates. The OpenAI Help Center currently lists common formats such as PDF, DOCX, TXT, CSV, XLSX, and PPTX among supported file types.
Protect Sensitive Information Before Uploading
Documents may contain personal data, confidential business information, medical records, financial details, passwords, signatures, or private account numbers. Remove or mask anything that is not required for the extraction task.
Before uploading, consider removing:
- Social Security numbers or national identification numbers
- Bank and credit card details
- Passwords, access codes, and security answers
- Private medical details that are not needed
- Signatures and confidential client information
- Full account numbers when only the last four digits are needed
For personal ChatGPT workspaces, review Settings → Data Controls and the “Improve the model for everyone” option. OpenAI’s documentation explains that Data Controls let users manage whether conversations may help improve models. Business, Enterprise, Edu, and managed workspaces may follow different organizational controls and policies.
Do not upload material that your employer, client, school, healthcare provider, or contract prohibits you from sharing with an external AI service. When the document is regulated or highly confidential, use an approved workplace system or ask the responsible administrator.
Best Practices for Reliable Results
- Define the fields first. Use fixed column names.
- Separate extraction from analysis. Do not ask for conclusions in the first pass.
- Preserve original wording. Request verbatim text when precision matters.
- Request source locations. Add page or section references.
- Prevent guessing. Use “Not found” and “Needs manual review.”
- Use CSV for spreadsheets. Use Markdown tables for quick review.
- Process large files in batches. Reuse the same schema for every batch.
- Verify high-impact values manually. Check the source before acting on the result.
- Protect private information. Mask data that is not needed.
- Save the final prompt. Reuse it for similar documents to keep results consistent.
A quick final checklist helps ensure your document is extracted accurately before you review or export the results.
Final Copy-and-Paste Prompt
Extract structured information from the uploaded document.
Fields to extract:
- Full Name
- Organization
- Date
- Amount
- Reference Number
- Contact Information
- Source Page or Section
Output rules:
1. Return one separate occurrence per row.
2. Use a Markdown table first.
3. Preserve names, codes, amounts, and dates exactly as written.
4. Add a short context note explaining what each value refers to.
5. Do not summarize the document.
6. Do not calculate, interpret, or infer missing information.
7. If a field is not clearly stated, write "Not found."
8. If a value is ambiguous, write "Needs manual review."
9. Include the source page or section for every row.
10. After the table, report the total number of extracted rows.
Wait for me to review the table before converting it to CSV.
After checking the table, use this follow-up:
Convert the verified table to valid CSV inside one code block.
Keep the same columns and row order.
Quote fields that contain commas.
Do not add explanations before or after the code block.
Do not change any verified value.
Frequently Asked Questions
Can ChatGPT extract every name and date from a PDF?
It can extract many items from a readable PDF, but no AI result should be assumed complete or error-free. Use precise fields, smaller page ranges for long documents, source references, and manual verification.
Should I ask for a table or CSV?
Use a Markdown table when you want to review the result inside ChatGPT. Use CSV when you are ready to move verified records into Excel, Google Sheets, or another data tool.
What does “verbatim” mean?
Verbatim means copying the exact original wording without rewriting it. It is useful for legal clauses, official names, product codes, deadlines, and quotations.
How do I stop ChatGPT from inventing missing values?
Add explicit rules such as “Do not guess,” “Write Not found when the source does not state the value,” and “Mark uncertain items Needs manual review.” You should still verify important results yourself.
Why did ChatGPT miss information in the middle of a long document?
Long documents are harder to inspect consistently. Divide the file into page ranges, run the same extraction prompt on each range, and then combine the results while preserving source references.
Can I upload confidential documents?
Only upload documents you are authorized to use with the service. Remove unnecessary sensitive information, check your organization’s policy, and review the current ChatGPT Data Controls and workspace rules.
Related Articles
👉 How to Create a ChatGPT Account: A Step-by-Step Beginner Guide
👉 How to Use ChatGPT for the First Time: A Simple Beginner Guide
👉 ChatGPT Free vs Plus: Which Plan Do Beginners Need?
👉 How to Find the Official ChatGPT Website and App
📚 Professional References
Technology Information Notice
AI tools, file support, interface labels, usage limits, and privacy settings can change. Availability may differ by plan, workspace, country, device, and account. Check the current official documentation before relying on a feature, limit, or policy. AI-generated extraction may contain omissions or errors, so verify important names, dates, amounts, codes, and legal wording against the original document.
#ChatGPT #DocumentExtraction #DataExtraction #CSV #Excel #GoogleSheets #AIPrompts #Productivity
Comments
Post a Comment