Analyse a long document with AI and keep the sources

Prepare your questions

Start with the structure of the document. Read the table of contents, the summary, the methodology and the list of annexes. Note where each topic is covered and on which pages. An annual report spreads its figures across the management report, the financial statements and the notes. A parliamentary inquiry report sets out its findings, its recommendations and the record of its hearings. A court ruling sets out the facts, the arguments of the parties, the reasoning and the decision.

Then write your questions. A broad question such as “What does the report say about safety?” produces a summary. A question such as “How much did the plant operator spend on maintenance in 2023?” produces a figure that can be checked on a page.

Group the questions by type: dates and events, figures, decisions and their authors, recommendations and their addressees. Add one question on what the document does not cover. This list guides the reading and makes the answers comparable from one document to the next.

Split the document

Every general-purpose tool has a length limit. Beyond that limit, part of the document may be left out of the answer. Splitting the document by chapter keeps each request within the limit and makes each answer easier to check.

Follow the structure of the document: one chapter, one section of the accounts or one part of the ruling at a time. Keep the original page numbers. The page number printed on a report often differs from the one shown by the PDF reader, because the cover and the front matter are counted differently. Choose one numbering and keep to it.

Treat the annexes separately. Tables, appendices and footnotes often hold the figures that the main text rounds or leaves out. Check that a scanned document has been converted into searchable text. Pages the tool cannot read do not appear in its answers.

Ask for the exact excerpt and the page

Ask for every answer in three parts: the answer, the excerpt copied word for word and the page. The excerpt lets you check the answer in seconds. Search for it in the PDF with the find function. An excerpt counts as a quotation only if it appears word for word in the document.

Ask one question at a time, so that each answer comes with its own excerpt. Ask the tool to say when the document does not answer the question. Without that instruction, a general-purpose assistant may fill the gap with what it already knows or with a plausible guess.

Keep a log of the questions, the answers, the excerpts and the pages. The log becomes the source file for the note, the article or the report you write next.

Cross-check the figures

Figures need a separate check. Check each one against the table it comes from, then against the other places where it appears.

• The unit: thousands, millions, percentages or percentage points.
• The period: calendar year, financial year or quarter.
• The scope: the group or one subsidiary, the country or one region.
• The version: a restated figure for the previous year can differ from the figure first published.
• The calculation: recompute totals, changes and ratios yourself.

Compare the figures in the summary of the document with those in its tables. A summary can round a figure or present it more favourably. Compare the document with its earlier editions as well. Note any figure that appeared in earlier editions and is missing from this one.

Look for what is missing

Compare the document with its terms of reference. An inquiry report answers the questions set by its mandate, and a question left unanswered deserves a note. A court ruling answers the arguments of the parties, and an argument dismissed without reasons should be recorded.

List the people and the documents the report cites, and those it does not cite. Check the list of people heard in an inquiry report and the list of exhibits in a ruling. Read the footnotes and the annexes, where reservations and dissenting contributions often appear.

Ask the AI tool directly about what is absent. “Which recommendations of the previous report are not mentioned in this one?” is a question a tool can answer with excerpts from both documents. The person doing the analysis decides which gaps matter.

What generic summaries miss

A one-click summary of a long document has known failings. Researchers have measured them.

Language models use the middle of a long text less well. In Lost in the Middle, accepted by the Transactions of the Association for Computational Linguistics, Nelson Liu and his co-authors found that performance is often highest when the relevant information sits at the beginning or the end of the text. Performance degrades significantly when the information sits in the middle.

Summaries leave out passages and misstate facts. The FABLES study, presented at the Conference on Language Modeling in 2024, examined 3,158 claims in AI summaries of 26 books. Most unfaithful claims concerned events and the states of characters. The authors also found omissions of key narrative elements and a systematic over-emphasis on events near the end of the book.

Answers misstate facts and sources. In October 2025, the European Broadcasting Union and the BBC published a study of over 3,000 answers to news questions from four AI assistants. Journalists from 22 public service media organisations in 18 countries reviewed them. 45% of the answers had at least one significant issue. 31% showed serious sourcing problems, and 20% contained major accuracy issues, including invented details and outdated information.

These results describe general-purpose tools at a given date. The checks in this guide apply to every tool.

The length limits of general-purpose tools

General-purpose assistants accept long files, within limits that vary by tool and by plan. According to their official pages on 17 October 2026:

• Google says that up to 10 files can be uploaded to Gemini in the same prompt, subject to availability, and up to 100 MB per file, videos excepted. Google compares the context window to a reading capacity. With the larger capacity offered by its Google AI Pro and Ultra plans, Gemini can understand up to 1,500 pages of text. When files exceed the context window, Google warns that responses may not take into account all the content provided.
• Anthropic’s developer documentation for Claude sets a limit of 600 pages per request, reduced to 100 pages when the request uses a smaller context window. It warns that dense PDFs, with small fonts, complex tables or heavy graphics, can fill the context window before reaching the page limit. It advises splitting the document into sections.

These limits apply to one conversation or one request. A file of several thousand pages, or a set of reports and annexes, has to be split, and each part analysed separately. The cross-checking between parts then falls to the reader.

Long documents with Verbatim

Verbatim reads document sets far beyond the limits of general-purpose assistants. Its agents read the whole file, keep every answer tied to its page and check the figures for you.

The Reading agent reads up to 40,000 pages: reports, scans, spreadsheets, emails and recordings, in most languages. It reads every page from start to finish. You ask your question, and the answer arrives with its excerpts. Each excerpt gives its page and lines, its minute in a recording or its cell in a spreadsheet. A click opens the document at that spot. When your documents do not hold the answer, Verbatim says so and offers to search the web. Read more about research and reading.

Excerpts appear exactly as they are in the document. A quotation is tied to a document only if the text appears there word for word. A finding taken from a summary counts as checked only once it has been found in its original document.

Every figure computed by the agents points back to its original cells. When an agent calculates a change, a ratio or a total, it keeps a record of the calculation, and a click opens the source cells. The Checking agent flags figures that appear in no source, such as an amount announced at a press conference that no annual report mentions. See how data analysis works.

Several agents work on the same project in parallel. One reads the annual reports, another reads the inspection report and a third compares their figures. A long reading becomes a task that runs for hours, and you follow its progress. The agents build a dated timeline, a table of figures or a summary, each item linked to its page, and the result exports to Word or PDF. The page on Verbatim for analysts shows the workflow for reports, accounts and filings.

Your documents are hosted and processed in Europe and are never used to train a third-party AI. VerbatimAgent starts at €49.95 a month, VAT included, with a 14-day free trial.

Sources

• arXiv, Lost in the Middle: How Language Models Use Long Contexts (Liu et al., Transactions of the ACL, 2023)
• arXiv, FABLES: Evaluating faithfulness and content selection in book-length summarization (Kim et al., COLM 2024)
• European Broadcasting Union, Largest study of its kind shows AI assistants misrepresent news content 45% of the time
• Google, Upload & analyze files in Gemini Apps
• Anthropic, PDF support (Claude developer documentation)