A confident summary with no source labels is not research. It is a draft you now have to fact-check from scratch. The fix is not a better model — it is a shorter brief that forces every claim to carry a source tag, and a review pass that deletes anything you cannot trace.

Quick answer

Write a brief that does four things before you run anything: names the questions you need answered, names the sources you will allow, requires a source label on every claim, and requires the tool to write "not in sources" instead of filling a gap. Then spot-check at least two claims against the original text yourself, downgrade or delete anything untraceable, and keep the source list attached to the summary so a colleague can repeat the check.

That is the whole method. The rest of this article is the wording, the review checklist, and the failure modes.

Why summaries arrive confident but untraceable

Most research prompts ask for an answer. They do not ask for provenance. When you write "summarize what's happening in this industry," you get fluent prose with no way to tell which sentence came from which document — or whether a sentence came from a document at all.

Three things go wrong in that gap:

Blending. Two sources say slightly different things. The summary merges them into one sentence that matches neither.
Gap filling. A question has no answer in your sources. The summary answers it anyway, because the prompt implied an answer was expected.
Citation drift. A source label appears next to a claim, but the label points at a document that does not actually contain that claim.

None of these are exotic. They are the normal result of asking for output without asking for traceability. The brief is where you prevent them, because once the summary exists, you are reverse-engineering it.

Step 1: Before you run anything, write down two lists

Do this in a scratch file, not in your head.

List A — the questions. Keep it to three to five. Each one should be answerable from a document, not from opinion. "What do the sources say about pricing pressure in this category?" is answerable. "Is this category growing?" invites a guess unless a source states it.

List B — the allowed sources. Name them specifically: a report, three competitor pages, a set of interview notes, a folder of PDFs. Anything not on the list is out of scope. This is the single most useful constraint in the brief, because it gives the tool a boundary and gives you a finite set of documents to check against.

If you cannot name your sources, you are not ready to brief. You are ready to search.

Step 2: Write the brief so every claim carries a label

The brief has five parts. Keep the whole thing under about 200 words — length is not what makes it work.

Task — one sentence on what you want back.
Questions — your List A, numbered.
Sources — your List B, named.
Labeling rule — every claim gets a source label; unsupported claims get not in sources.
Output shape — bullets, one claim per bullet, label at the end of each bullet.

Copyable example brief

text

Task: Summarize what the attached sources say about [industry topic].
Do not add outside knowledge.

Questions:
1. What is described as the main demand driver?
2. What is described as the main constraint?
3. What numbers are stated, and by which source?

Sources (use only these):
- Source A: [document name]
- Source B: [document name]
- Source C: [document name]

Rules:
- End every bullet with a source label in brackets, e.g. [Source B].
- If a question is not answered in the sources, write exactly:
  "not in sources" and stop. Do not infer, estimate, or fill the gap.
- Do not state any figure that is not written in a source.
- Do not create citations. Only label with the source names above.

Output: one bullet per claim, grouped under the three questions.

Three lines in that brief do the heavy lifting. "Do not add outside knowledge" closes the gap-filling door. "not in sources" gives the tool a permitted way to fail. "Do not create citations" tells it that a plausible-looking label is worse than no label.

What a compliant output looks like

text

1. Main demand driver
- Buyers are described as consolidating vendors to cut admin overhead. [Source A]
- The shift is attributed to procurement review cycles, not product features. [Source C]

2. Main constraint
- Budget approval is described as the bottleneck for new tooling. [Source B]
- not in sources

3. Stated numbers
- A 12% figure appears in Source A's summary of survey responses. [Source A]
- not in sources

Notice what the not in sources lines do. They are the most valuable part of the output, because they tell you where your source set is thin. A summary with no gaps is usually a summary that guessed.

Step 3: Spot-check two claims yourself

Do not read the summary for quality first. Read it for traceability.

Pick two bullets — ideally one with a number and one that surprised you — and open the labeled source. Find the sentence. If you cannot find it in a couple of minutes, the bullet fails, regardless of whether it sounds right.

Two checks are enough to calibrate the run. If both pass, you have reasonable grounds to trust the labeling behavior on that document set. If either fails, stop spot-checking and treat the whole output as unlabeled — the labeling rule did not hold, so the labels carry no information.

This is a judgment call, not a measurement. You are not scoring accuracy; you are deciding whether the output is checkable.

Step 4: Delete or downgrade what you cannot trace

Sort every bullet into one of three buckets:

Bucket	What it means	What to do
Traceable	You found the claim in the labeled source	Keep as-is
Partly traceable	The source supports a weaker version of the claim	Rewrite to match the source, or mark it as your own interpretation
Untraceable	No source contains it, or the label is wrong	Delete it, or move it to a separate "open questions" list

The middle bucket is where most damage happens. A source says "several respondents mentioned cost." The summary says "cost is the top barrier." That is a real difference, and it survives review unless you compare wording. Downgrade it.

Deleting is not losing work. An untraceable claim in a summary that goes to a client is a liability. An untraceable claim in an "open questions" list is a research lead.

Step 5: Keep the source list with the summary

Attach the brief — questions, source names, and the labeling rule — to the summary itself. Not in a separate folder, not in your memory.

This is what makes the check repeatable. A colleague who receives the summary plus the brief can re-run the same questions against the same sources and see whether they get the same labels. Without the brief, they can only re-read the summary and take your word for it.

If you are working inside a research workspace, keeping the brief and the source set in the same project as the summary is the simplest version of this. The AI research agent capability page describes the workspace where sources, prompts, and outputs live together, which is the practical version of "keep the source list with the summary."

Output review checklist

Run this before the summary leaves your hands.

Every bullet ends with a source label, or reads not in sources.
No bullet contains a figure that is absent from its labeled source.
At least two bullets were checked against the original text by a human.
Every not in sources line is still there — none were quietly deleted to make the summary look complete.
Partly traceable claims were rewritten to match source wording, or labeled as interpretation.
The brief and source list travel with the summary.
No citation appears that you did not verify exists in the named source.

Common mistakes

Asking for a summary instead of answers. "Summarize these documents" produces prose. "Answer these three questions using only these documents" produces checkable bullets. The second is barely harder to write and far easier to review.

Leaving the source list vague. "Use relevant industry sources" gives the tool permission to draw on anything. Name the documents.

Not giving the tool a way to say no. If the only acceptable output is an answer, you will get an answer. not in sources is a feature, not a failure.

Treating labels as verified. A label is a claim about provenance, not proof of it. It becomes evidence only after you check it.

Editing the summary before checking it. Rewriting for tone first makes it much harder to tell which sentences were yours and which came from the run.

Dropping the gaps. Gaps are the finding. A summary that says "two of three questions are not answered in these sources" has told you something true and useful about your source set.

FAQ

How long should the brief be?

Short enough to read in under a minute. Five parts: task, questions, sources, labeling rule, output shape. If your brief is longer than your summary, you have probably turned it into a spec for a different task.

What if the tool ignores the labeling rule?

Treat the run as failed rather than partially successful. Re-run with the source list shortened to two documents and the output shape made stricter — one claim per bullet, label required. If labels still do not hold, the output is not checkable and should not be circulated.

Is "not in sources" a sign the tool did something wrong?

No. It usually means your source set does not cover the question. That is worth knowing before you present the summary, and it is a much better outcome than a confident sentence with no basis.

Can I use this brief for a summary I will publish?

The workflow makes claims traceable; it does not grant you rights to the underlying material. Check the terms attached to each source before republishing anything, and keep quotes attributed. See pricing if you need to confirm what your current plan covers.

How many sources should I allow?

Start with three. A three-source brief is small enough that you can spot-check the whole thing when something looks off, and large enough to show you where sources disagree. Add sources once the labeling behavior is holding.

Do I need to re-check every run?

You do not need to re-check every bullet. You do need to re-check when the source set changes, when the questions change, or when a bullet contains a number. Those are the conditions where labeling tends to slip.

The takeaway

A verifiable summary is a brief-writing problem, not a model problem. Name your questions, name your sources, require a label on every claim, permit not in sources, check two claims yourself, delete what you cannot trace, and keep the brief attached. The summary gets shorter and less impressive. It also becomes something you can defend.

How to Write a Short Brief for an AI Research Summary You Can Actually Verify