An AI answer cites your PDF. Check the sentence before you reuse it

You ask an assistant whether a device can be used outdoors. Its answer says yes and links to the manual. Before copying that answer into a purchasing note, you open the citation. The highlighted paragraph discusses indoor installation. This is an illustrative situation, but it gives the verification task a useful starting point: compare the answer with the passage it names.

Keep the PDF open beside the response. Work on the sentence you intend to reuse, rather than asking the assistant to defend its entire summary. The practical question is whether the source supports the subject, conditions and conclusion in that sentence. Google’s grounding documentation uses this standard of full support for a claim; partial support does not satisfy it.¹

Establish which page you are looking at

Save the answer and identify the exact document version used in the conversation. Record its title and revision date if available. Begin verification with the copy the assistant received. Check the publisher’s latest edition separately. Otherwise you cannot tell whether a discrepancy belongs to the answer or to a different edition.

Then record the page’s position in the file and the number printed on it. Adobe explains that navigation labels can differ from printed page numbers.² Covers and introductory material deserve a look before you assume a consistent offset. For a working note, write both locations in words, such as “file page; printed page,” and fill in the values you observe.

A citation may expose a technical range instead of a reader-facing page label. In Anthropic’s citations API, PDF page ranges start at one and exclude the end page; a document index identifies the supplied file.³ Interpret those fields using that API’s documentation. Do not apply its counting convention to another product.

Check the text the system could extract

Try selecting a short sentence on the cited page and copying it into a plain text editor. Compare the pasted result with the visible page. Treat this as a diagnostic of your copy, not a demonstration of what an unknown service ingested.

Adobe’s OCR creates a searchable text layer from a scan and calls for reviewing the recognized text.⁴ If a word or symbol matters to your conclusion, compare it directly with the image. Look especially at a negative word in a quotation, a decimal separator in a value, or a marker directing you to a note.

Search failure needs investigation. Search a shorter distinctive phrase, then inspect the cited page visually. A sentence that you cannot find by text search remains unverified until you locate it or determine what happened. Conversely, successful copying establishes only that text is available in your reader. It does not establish that the assistant retained the same context.

Follow the citation beyond the retrieved fragment

Google’s document parsing documentation distinguishes extracting digital text from recognizing document layout and retrieving chunks.⁵ A chunk reference identifies retrieved material. Check separately whether it supports the conclusion, using the full passage.⁵¹

Ask the assistant to expose the cited passage and its location if the interface permits it. Then read the original paragraph, including the sentence before and after. If the answer combines a condition with a conclusion from elsewhere, find both passages. Keep the connection as your inference unless the document makes it explicitly.

Use the following troubleshooting guide to decide what to inspect next. The actions are editorial recommendations based on the documented mechanisms, rather than measured failure rates.⁴²³⁵

What you find	Inspect next	What you can establish
The link opens a plausible but different page	File position, printed number and section heading	Whether page numbering explains the mismatch
Copied words differ from the image	Original scan at readable magnification	Whether the extracted text is dependable for this passage
A short fragment lacks a condition	Full paragraph and nearby notes	Whether the answer broadened the source’s meaning
A table value appears without its heading	Row, column, unit and attached note	Which quantity the value describes
A citation points only to a file or chunk	Exact supporting sentence in the original	Whether the claimed support can be reproduced

Verify a quotation without letting it drift

Start with a distinctive phrase from the quoted sentence. Search inside the PDF, inspect the match, and compare the wording all the way through. If line breaks interrupt the search, shorten the phrase. Keep searching and reading separate: a partial match gives you a place to investigate.

As an invented wording example, suppose a manual says “For indoor use only.” A response presenting that sentence as permission for outdoor use contradicts the passage. Changing the answer to a paraphrase cannot repair the contradiction. Correct the conclusion and attach the passage that supports the correction.

A less obvious case is a quotation that drops a condition introduced earlier in the paragraph. Read enough context to identify the device or configuration the writer means. If you remove words from a quotation, ensure the remaining text preserves its meaning. If you translate it, label the translation and retain the original wording in your verification note.

Anthropic’s PDF citation feature requires extractable text and does not support image citations.³ That specific limit concerns the cited feature. When a scan is involved, check which reading and citation features your actual application uses.

Verify a numeric cell as a complete statement

For a claimed value in a specification table, open the whole table before checking the digits. Find the named product row, follow it to the relevant column, and copy the value with its unit. Include any qualifier in the heading, such as “typical” or “maximum.” Read the attached note before deciding that the number answers your question.

Do this manually once, without looking at the assistant’s value while you transcribe. Then compare your note with the response. This method is a recommendation for making the comparison explicit; it carries no promised accuracy rate.

If a table continues on another page, inspect both parts and confirm the applicable heading. For a calculation, preserve the original input values and show your arithmetic separately. A derived result needs enough information for another reader to repeat it. A citation to the table does not tell the reader which operation you performed.

Google documents visual processing of PDF tables and charts in the Gemini API.⁶ If the evidence is a chart or diagram, inspect its labels and legend in the original image. Still inspect the original cell when the number matters to the sentence you will reuse, and check the application configuration before assuming it supplied the visual content.

Stop when the source and answer will not align

If the citation opens the wrong edition, the quoted words remain missing, or the cell cannot be identified, mark that sentence as unverified and leave it out of the note you are preparing. Preserve the mismatch so someone else can inspect it. “Unable to verify against the supplied PDF” is more useful than a repaired answer with an unexplained citation.

You can ask for a focused retry: identify the supporting passage, give its location, and state any missing condition. Treat that response as a new candidate for inspection. If the source still does not support it, return to the original document and narrow or remove the claim.

Keep a short record of the final wording, the passage or table context, both page locations and your assessment. Distinguish full support from partial support and contradiction; keep unreadable or missing evidence unresolved. This records what you checked. It does not establish that the document itself is correct or current.

References

Footnotes

1. Google Cloud — Grounding requires the whole claim to be supported by supplied facts: https://docs.cloud.google.com/generative-ai-app-builder/docs/check-grounding ↩ ↩²

2. Adobe Acrobat — Page navigation labels and printed page numbers can differ: https://helpx.adobe.com/acrobat/desktop/edit-documents/organize-pages/renumber-pages.html ↩ ↩²

3. Anthropic — PDF citation locations, exclusive end ranges and image citation limits: https://platform.claude.com/docs/en/build-with-claude/citations ↩ ↩² ↩³

4. Adobe Acrobat — OCR creates searchable text and requires review of the result: https://helpx.adobe.com/acrobat/desktop/create-documents/scan-documents-to-pdfs/recognize-text.html ↩ ↩²

5. Google Cloud — Digital parsing, layout recognition and document chunks: https://docs.cloud.google.com/generative-ai-app-builder/docs/parse-chunk-documents ↩ ↩² ↩³

6. Google AI for Developers — Visual and textual processing of PDFs in the Gemini API: https://ai.google.dev/gemini-api/docs/document-processing ↩