A reader asked ChatGPT for every tree marked for removal in an arborist report. It kept missing entries across the multi-page table, according to their January 23, 2025 forum post. That older case is a useful reason to check what a PDF answer leaves out before trusting the summary.
The short answer
Open the PDF beside the answer and check one small section first. Write down the headings and details you expect, then ask your assistant to attach a source passage to each claim. Compare those passages yourself. A correct answer about the last page does not prove the rest was covered. Keep missing or unreadable material on a separate list.
What the reports show
A separate July 9, 2026 Reddit report describes repeated requests to re-upload PDFs and a summary containing information the reader said was absent from the file. The poster later reported normal behavior returning. I would treat this as a reported episode, rather than proof of a permanent product fault.
OpenAI’s current upload guide supports asking questions about attached documents. It also distinguishes direct chat attachments from stored project or GPT files, where embedded images may be handled differently. Work from the actual file you want checked.
Start with a public manual or report you are allowed to share. Remove private details from a working copy. Reading the relevant pages yourself, using the document’s search box and bookmarks, is the lower-risk option when you only need one answer.
Try one small task
Choose one answerable question
Set aside “summarize everything.” Pick a question such as “What must I bring to the appointment?” or “Which tasks need finishing before setup?” Find the section that answers it in your PDF reader before asking AI.
Write down the PDF viewer page number and the printed page number if they differ. A cover and contents page can make them different. Decide which numbering you will use.
Make a small coverage list
List the relevant heading, subheadings, table and any footnote. For a practice task, choose two or three pages. Beside each item, leave space for “checked,” “missing” or “unreadable.”
If you cannot select text, the page may be a scan. Try the file directly in the chat, or attach a clear image of the relevant page if that feature is available. Check the recognized words against the page before relying on them.
If you use a chatbot that accepts PDF files, ask it to identify pages by both PDF page number and printed heading. If you paste an excerpt, label the pages yourself. Claude’s file guide explains when visual PDF elements are available to the model.
Try selecting a sentence in the PDF. If all you can select is the whole page as an image, use a legible, non-sensitive page image if your assistant supports it, or read the page yourself. For a table, count rows and check the column labels on the original.
Ask for evidence beside each answer
Attach the prepared excerpt and use the prompt below. Ask for the original wording as well as the explanation. This gives you something concrete to compare with the page.
If the answer refers to a document you did not attach, stop and identify the file it used. Repeating the question without fixing the source can produce a different confident answer.
Use only the attached excerpt to answer: [my question]. Identify the file and the page-numbering convention. For each answer, give the section heading, viewer page number and a short exact supporting passage. Separate explicit statements from your interpretation. Check these items: [my coverage list]. List anything missing, unreadable or not answered. Do not fill gaps from general knowledge.
Compare the answer with the pages
Read every supporting passage in the PDF. Check that it says what the explanation claims and includes any exception nearby. Tick your coverage list yourself.
A three-page check is a check of those three pages. For a longer report, repeat by section and keep the unchecked sections visible. I would rather have a short answer with one honest gap than a complete-looking answer that quietly skips the appendix.
This is the kind of check I help people practice one-to-one.
Check the result before you use it
| Illustrative output | What I would do |
|---|---|
| “Bring photo ID.” The cited passage says “Photo ID is required unless already verified online.” | Keep the exception. The shorter version changes who needs to act. |
| “All requirements checked.” No appendix is mentioned. | Mark the appendix unchecked and inspect it separately. |
| The final heading is correctly named. | Use that as one spot check, then check the middle and the details that matter. |
Watch for answers that combine a correct heading with an unsupported detail. A page reference helps you find evidence; it does not make the claim true.
Questions you may have
Does asking AI to read every page guarantee coverage?
No. Compare the output with your own section list and source passages. Check the parts that affect your next action.
Does this check work with other AI assistants?
Yes. Matching a claim to the original passage is a reader-performed check. File-upload and document-image support depend on the service you use; check that service’s documentation separately.
Make this work for your own situation
The harder task is building a routine for the reports you read every month: deciding which sections need a complete check, which can be skimmed, and how to carry exceptions into an action list. That depends on what you will do with the answer.
For health, legal or financial decisions, take the original document and unresolved questions to the relevant professional. A clearer summary does not replace their judgment.
Work through your next step with me
Bring a public report or a version you are permitted to share to a 60-minute virtual session. We can practice one section together, build a coverage list and compare the answer with the original. You’ll have a checking method you can repeat and a clear record of what still needs reading.
See session options and book →
Still exploring? Browse practical AI questions.
Last checked November 6, 2026.