Guide
How to make AI cite the exact page a number came from
Ask for page-level citations explicitly and make each page identifiable, but the reliable way is to give AI a version of the document where every value already carries a reference to its source page, slide or cell. Then the citation is data the AI reads, not a location it has to reconstruct from a PDF.
Why AI cites pages badly
When an assistant reads a PDF, it usually works from extracted text split into chunks. Page boundaries may or may not survive, and a value can be pulled from a chunk that spans two pages. The model then names a page from memory of the chunk, which is why citations are often near the right place but not on it.
Long files make it worse: OpenAI notes that large or image-heavy files may not be fully analysed and suggests asking about specific sections. And a number found in a chart or table may be extracted unreliably, so the cited page can hold a different figure than the one quoted.
Prompts that help
- Ask for the page with every figure: “For each number, give the page number and quote the sentence or table row it comes from.”
- Require a verbatim quote — it forces the model to locate the text and lets you check it fast.
- Ask it to say “not found” rather than estimate.
- Narrow the scope: “Using only pages 12–18…”.
- Ask it to list period, currency and qualifier with each value, so you can spot a figure from the wrong column.
Document habits that help
- Print page numbers in the text layer, not only as images in a footer design.
- Number slides and keep the number visible.
- Label tables and figures (“Table 4: Revenue by segment”) so a citation can name them.
- Keep each key figure in one place and refer to it, rather than repeating it with small differences.
- For spreadsheets, use named ranges and clear sheet names so cells can be cited meaningfully.
Make the reference part of the data
Prompts improve the odds; they do not guarantee a correct page. The structural fix is provenance: store, with each value, where it came from. In a structured version of a document that looks like this:
| Metric | Value | Period | Qualifier | Source |
|---|---|---|---|---|
| Revenue | $4.2M | FY2025 | unaudited | page 7 |
| Gross margin | 62% | FY2025 | — | page 7 |
| Revenue | $6.0M | FY2026 | forecast | slide 14 |
| Headcount | 38 | Sept 30, 2026 | — | sheet “Team”, cell B4 |
Who needs this most
- Finance and investor relations: an analyst’s assistant must quote the right period from the right page.
- Legal: amounts, dates and parties in a contract must be traceable to a clause and page.
- Research: figures from reports must be traceable to the table they came from.
- Sales and procurement: a price or spec quoted from a proposal must point to the exact line.
How The Papers handles it
In The Papers, every value in a document’s AI Version keeps a reference to its source page, slide or cell. The JSON version (/{room}/{doc}.json) carries page provenance alongside each metric’s period, currency, classification and qualifiers, so an agent can cite the location instead of guessing it. Before publishing, each number is checked against the source; values the source cannot settle go to you and are recorded as owner-confirmed. The link stays the same when you update the document, so citations point to the current version.
FAQ
- Why does ChatGPT give a page number that is off by one or two?
- Extracted text is split into chunks that may cross page boundaries, so the model approximates. Asking for a verbatim quote makes errors easy to spot.
- Can I make citations work for slides and spreadsheets too?
- Yes, if the reference is stored with the value: slide number for decks, sheet and cell for spreadsheets. The Papers keeps those references for PowerPoint and Excel sources.
- Does a citation mean the number is right?
- It means you can check it. Provenance shows where a value came from; fidelity check confirms it matches that source.