# How to Research Documents with AI and Verify Citations

> Turn a set of documents into a defensible AI research brief with a source register, conflict table, citation checks, and clear unanswered questions.

Start AI document research with a defined question and a source register. Extract evidence before drafting, preserve conflicting versions, and open each citation that supports the decision. A citation is a pointer to check, not proof that a summary is complete or correct.

To research documents with AI, define one question, register the sources, and extract evidence before asking for a finished brief. Open the references supporting the answer and preserve unresolved conflicts. A citation helps you find a passage; you still need to check that the passage supports the claim beside it.

This workflow produces three reusable artifacts: a source register, an evidence table, and a reviewed brief. It works best when someone on the team can identify which documents govern the decision. The assistant should not invent that authority from filenames or confident wording.

## Ask a question that can be answered from the documents

“Summarize these files” can produce a readable overview without resolving your actual decision. Ask something bounded, such as: “Which customer onboarding commitments are approved for the launch brief?” That question tells you which claims need evidence and which details can stay outside the output.

Write down the intended reader and use of the answer. A customer-facing brief requires approved promises; an internal planning memo may also include tentative targets. If you blur those uses, a provisional date can become a commitment during drafting.

Use company-approved material and an approved workspace. Include only the documents needed for the question. If several people maintain the set, the [shared AI Projects guide](/guides/comparisons/chatgpt-projects-vs-claude-projects) provides a source-owner template and a teammate access check.

## Make a source register before uploading

The register is a simple inventory, not a special research standard. Give each source a stable ID so a reference remains understandable even when the output moves to another document. Record who can clarify its meaning and what it is authoritative for.

**Illustrative example:** all companies, documents, owners, and commitments below are invented. A fictional onboarding team supplies these three sources:

| ID | Document and version | Owner | Authoritative for | Location and limitation |
| --- | --- | --- | --- | --- |
| S1 | Sales draft, September 20, 2026 | Sales lead | Historical draft wording | “Onboarding” section; superseded for delivery commitments |
| S2 | Approved delivery policy, October 1, 2026 | Delivery lead | Current onboarding commitments | Section 2; timing depends on completed prerequisites |
| S3 | Meeting note, October 3, 2026 | Onboarding lead | Discussion and tentative targets | “Next launch” section; not an approval of new promises |

The authority rule is supplied by the fictional team: **S2 governs customer-facing delivery commitments**. Newer is not automatically more authoritative. A recent meeting note may discuss a goal without replacing an approved policy.

For your own register, add the original file location, review status, and known gaps. If a source refers to an appendix you do not have, record that absence. Do not treat a partial document as a complete account.

## Choose a tool by the operation you need

| Needed operation | Suitable starting point | What you still check |
| --- | --- | --- |
| Navigate references inside selected documents | Gemini Notebook offers source selection and citations you can open to inspect context. [Google](https://support.google.com/gemininotebook/answer/16179559?hl=en) | Whether the selected passage supports the exact claim |
| Turn source files into a finished document | ChatGPT Work accepts source files and supports reviewing, downloading, and revising generated documents. [OpenAI](https://learn.chatgpt.com/docs/artifacts-viewer) | Evidence, output structure, and corrections before use |
| Work in another approved assistant | Use it if it accepts the supplied material; request source IDs and locations manually. | Whether each reference can be found in the originals |

See the [Gemini Notebook overview](/tools/notebooklm) for broader product context. This walkthrough uses selected documents for a bounded analysis. Google's current documentation also describes experimental actions beyond uploaded-source chat; a scope instruction is not a technical guarantee that every mode stays inside your files. [Google's chat guidance](https://support.google.com/gemininotebook/answer/16179559?hl=en)

Choose around your required artifact and available access, rather than an unsupported accuracy ranking. The [ChatGPT](/tools/chatgpt) and [Claude](/tools/claude) entries provide broader product context; the evidence checks here remain the same.

When the corpus is academic papers you still need to select, the [AI literature-review workflow](/guides/workflow/ai-literature-review-workflow) adds eligibility criteria, study-level evidence tables, and a screening record. [Elicit](/tools/elicit) provides workflows for that research stage.

## Inspect the imported sources separately

Before synthesis, ask for the relevant claims from each document individually. Compare the result with the original file. Did a timing table arrive? Are headings and page references usable? Can you find the section the assistant names?

Google documents several import caveats: web URL imports extract page text, not nested pages or embedded images and videos; Google-file imports omit footnotes and comments. Current Drive imports can sync, so check the version actually in use rather than assuming it is permanently static. [Google's source guidance](https://support.google.com/gemininotebook/answer/16215270?hl=en)

If an import missed evidence, supply the missing table or passage with its original source ID and location. Label it as an extract. Keep a link to the full original so a later reviewer can inspect surrounding context.

## Extract a claim table before drafting

Use this prompt after supplying the register and documents:

```text
Question: Which onboarding commitments are approved for the launch brief?
Use only the three supplied documents for this pass.

First make an evidence table: claim, source ID, section/page,
supporting passage, condition or caveat, conflicting source,
and proposed treatment in the brief. Do not draft the brief yet.

Apply the authority rule in the source register. If it does not
resolve a conflict, mark Unresolved and name the owner to ask.
If evidence is absent, write Not found. Do not use outside knowledge.
Preserve relative dates until the relevant calendar is confirmed.
```

These are instructions for the desired behavior. Inspect the output to see whether it followed them. In particular, do not accept a source reference merely because its format looks plausible.

The invented documents contain these short passages:

```text
S1, Onboarding: “Onboarding within two business days.”
S2, Section 2: “Onboarding within five business days after
prerequisites are complete.”
S3, Next launch: “Aim for Tuesday, if the customer sends the access list.”
```

A checked evidence table should preserve their different meanings:

| Claim | Source/location | Qualifier and authority | Verified? | Treatment in brief |
| --- | --- | --- | --- | --- |
| Onboarding within two business days | S1, Onboarding | Superseded sales draft | Yes, passage matches invented source | Flag stale wording; omit from promises |
| Onboarding within five business days | S2, Section 2 | After prerequisites are complete; approved policy | Yes, including the condition | Use the complete conditional commitment |
| Tuesday onboarding target | S3, Next launch | Depends on the access list; discussion only | Yes, as a tentative target | Keep in internal planning, not customer promises |
| Owner of the access-list request | Not found | No source names an owner | No supporting evidence | Ask onboarding lead; do not assign someone |

The first row is a real statement in the fictional draft, but it is not the approved answer. The last row has no evidence and should stay incomplete. Those distinctions prevent a smooth summary from hiding the team's actual uncertainty.

## Open and check the citations

For each claim used in the decision, open the cited source location and check five things:

1. **Existence:** the file and referenced section actually exist.
2. **Exact support:** the passage supports this claim, not merely the same topic.
3. **Version:** the register and cited file identify the intended version.
4. **Qualification:** conditions, exceptions, and scope remain in the answer.
5. **Interpretation:** the assistant distinguishes document wording from its own inference.

In the example, citing S2 beside “onboarding takes five days” drops both “business” and the prerequisite condition. Correct the claim before drafting. Citing S3 beside “launch is Tuesday” turns a target into an approved date. The citation exists, but the interpretation fails.

This is the practical difference between [grounding](/glossary/grounding) an answer in sources and assuming that source references eliminate [hallucination](/glossary/hallucination). If the brief uses calculations too, apply the separate [AI data-analysis check](/guides/analytics/check-an-ai-data-analysis).

## Draft a brief from the checked evidence

Only after review, ask the assistant to use the accepted evidence rows. Keep the register and claim table attached so someone can re-check the result after a policy update.

```text
Question/decision:
Supported answer:
Evidence: source ID + section + verified claim
Conditions and exceptions:
Conflicts and superseded wording:
Unresolved questions: question + person to ask
Reviewer and review date:
Source register and original-file locations:
```

For the invented example, the supported answer is: “Use S2's commitment: onboarding within five business days after prerequisites are complete.” The brief flags S1's two-day statement for correction and keeps S3's Tuesday target out of customer-facing copy. It asks the onboarding lead to name the access-list owner and confirm whether the prerequisites are complete.

If a meeting note is central evidence, use the [meeting notes workflow](/guides/workflow/ai-meeting-notes-action-items) to check whether the record establishes agreement. Do not let the final prose erase an unresolved question just because it interrupts the narrative.

For vendor agreements, carry this evidence discipline into a more specific packet. The [vendor-contract review workflow](/guides/workflow/ai-vendor-contract-review-packet) preserves clause references, conditions, missing documents, and unresolved dates in an obligation register for a qualified reviewer.

## Repair common evidence failures

| Failure | Next action |
| --- | --- |
| Imported text omitted a table | Supply the table separately, preserve its source location, and repeat the affected checks. |
| A reference names no precise location | Request a section or passage, then find it manually; omit the claim if it remains unsupported. |
| Two versions appear current | Ask the owner which governs this claim and update the register. |
| A URL points to a summary | Locate the underlying document before treating the claim as primary evidence. |
| The final brief loses a caveat | Restore the condition and re-check every sentence derived from that row. |

Finish with a reviewable answer, not a forced complete answer. A brief that says exactly what the documents establish, what they contradict, and what still needs an owner gives the team a useful basis for its next decision.

## Sources and further reading

- [Use chat in Gemini Notebook](https://support.google.com/gemininotebook/answer/16179559?hl=en) — Google
- [Add or discover new sources for your notebook](https://support.google.com/gemininotebook/answer/16215270?hl=en) — Google
- [Work with files](https://learn.chatgpt.com/docs/artifacts-viewer) — OpenAI

---

_Source: https://agentscamp.com/guides/workflow/ai-document-research-with-citations — Guide on AgentsCamp._