Summarising documents with citations.

Goal: Produce a summary of a long document in which every point carries a page reference and an exact quote, then verify a sample of those citations. Who it is for: Anyone who briefs others on documents they have not all read: managers, analysts, consultants, board secretaries and project leads. Time: 25 minutes, plus a long PDF (the exercise uses a free MBIE guidance document) and any AI assistant that accepts file uploads.

There is a cautionary tale inside the document you will practise on. MBIE’s Responsible AI guidance for businesses (July 2025) describes a fictional New Zealand policy consultancy whose AI-drafted report cited a compelling statistic from a report that did not exist. The client checked, the rework was expensive, and the firm changed its process so that every citation is verified against the original source. This lesson builds that process into the prompt.

Why a summary needs citations

A cited summary can be checked in minutes; an uncited one can only be trusted or redone. Citations do not make the AI more accurate. They make its errors visible, which is what you actually need.

The exact quote is the important part. A page number can be wrong by one, but a quote either appears on that page or it does not.

Exercise: summarise a guidance document with page-cited points

1. Get the document. Download MBIE’s Responsible AI guidance for businesses (PDF, 43 pages), or use a long document from your own work such as a lease, supplier agreement or policy manual.
2. Decide who the summary is for. Write one line: the reader and the decision they face, such as “our operations manager deciding whether to buy an AI tool”. The summary should serve that decision, not the whole document.
3. Upload and run the prompt. Use ChatGPT, Claude, Copilot or Gemini with the worked example prompt below. Adjust the reader line and the number of points.
4. Spot-check the citations. Pick three points at random plus any point you would act on. Open each page, search for the quote, and confirm it says what the point claims.
5. Ask what was left out. Ask the assistant which sections it did not cover and why, then compare against the document’s contents page.
6. Record the result. Note how many citations you checked and how many failed. If any random check failed, verify every point before sharing.

Worked example: the MBIE guidance, for a buying decision

I have uploaded "Responsible AI guidance for businesses" (MBIE, July 2025).
Reader: our operations manager, who is deciding whether to buy an AI tool
for our 20-person business.

Summarise the points most relevant to that decision in 5 to 7 bullets.
For EVERY bullet give:
- the PDF page number and the printed page number, if they differ
- an exact quote of no more than 20 words that supports the bullet
If you cannot find a supporting quote, leave the point out.
Finish with "Not covered": topics the reader might expect that this
document does not address.

Example output (abridged, checked against the PDF on 19 October 2026):

- Before buying, document customer needs, business needs, costs, risks,
  benefits and evaluation criteria, and consider a trial.
  [PDF p.15 / printed p.14] "consider requesting a trial (isolated from
  other technical systems)"
- Human involvement should scale with risk; decisions about money,
  health or the law need a person involved.
  [PDF p.30 / printed p.29] "a person should always be involved"
- Human review can fail through automation bias.
  [PDF p.30 / printed p.29] "accept outputs as correct without critical
  evaluation"
Not covered: specific product recommendations or prices.

Notice the page offset. In this document the printed page number is one lower than the PDF page, because the cover is unnumbered. An assistant that reports only one of them will send you to the wrong page about half the time.

Where cited summaries go wrong

• Fabricated citations. References to pages or reports that do not exist. Exact quotes are the control: a missing quote is fast to confirm.
• Real quote, wrong point. The quote exists but is about a different situation, or is qualified by the next sentence. Read a line either side when you check.
• Silent omissions. The clause you most needed, such as a termination fee, is simply absent. The “not covered” request and a glance at the contents page are the controls.
• Page drift. Cover pages, scanned documents and pasted text shift numbering. Ask for both page numbers, or for section headings.
• Very long documents. Quality can drop for material in the middle of long files. Split by section and summarise each, or see our comparison of RAG versus long context.

The same habit scales into systems. Lucent, Levy Street’s agentic investment research platform, grounds its outputs in citations for exactly this reason, and our contract clause review workflow shows the pattern applied to legal documents. When volumes are high, our AI document processing work builds the checking into the pipeline. For tools designed around source-grounded answers, see our guide to NotebookLM, now Gemini Notebook. The wider skill of verifying claims is covered in checking AI answers against sources.

Self-check

• The summary names its reader and the decision it supports.
• Every point has a page reference and an exact quote.
• You opened at least three random citations and every point you will act on.
• Each checked quote exists on the cited page and supports its point.
• You asked what was left out and compared with the contents page.
• You recorded how many citations you checked and how many failed.

Next lesson: Building a searchable business knowledge base.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.