Deep research source quality and citation policy

Quick answer

Deep research systems get better when they are judged on source quality before report quality.

If the workflow accepts weak sources, duplicate sources, low-authority summaries, or uncited claims, the final report can look sophisticated while still being unreliable. A healthy deep research policy defines:

what counts as an acceptable source,
how many independent sources are required for major claims,
which source types need to be down-weighted or excluded,
and how the final report must expose uncertainty.

Why this matters

The failure mode in deep research is rarely “the model wrote bad prose.” The more expensive failure is that the system:

overuses secondary summaries,
cites SEO pages instead of primary material,
repeats one claim across many low-value sources,
or produces confident synthesis from weak evidence.

That is a workflow problem, not only a model problem.

The minimum healthy source policy

For most production research systems, define at least four source classes:

Primary sources Regulatory filings, official documentation, earnings materials, standards, vendor technical docs, direct transcripts, original datasets.
High-quality secondary sources Established trade publications, credible analysts, technical explainers with clear sourcing, reputable journalism.
Low-confidence secondary sources Thin listicles, affiliate content, SEO rewrite pages, anonymous summaries, unsourced aggregators.
Disallowed sources Scraped duplicates, AI-generated summaries without provenance, pages with no identifiable publisher, and any source the team cannot defend in review.

The system should prefer class 1, selectively use class 2, heavily constrain class 3, and reject class 4.

The citation rule that prevents fake depth

Do not let the report cite “many sources” if those sources are just the same claim restated.

For material claims, require:

at least one primary source when possible,
or at least two independent high-quality secondary sources when a primary source does not exist,
and explicit uncertainty language when evidence remains incomplete.

This prevents the common failure where five weak web pages look like consensus.

The source-quality rubric

Before accepting a source into the research context, score it on:

Authority Is the publisher directly responsible for the information or clearly credible in the domain?
Specificity Does the source contain concrete facts, dates, numbers, or technical details?
Freshness Is the information recent enough for the topic?
Independence Is it genuinely independent, or just a rewrite of another source?
Traceability Can a reviewer easily find the original claim and inspect it?

If a source fails two or more of these dimensions, it should not anchor an important conclusion.

The policy for missing evidence

A deep research system should be allowed to say:

evidence is thin,
primary confirmation was not found,
or the answer is conditional on limited sources.

Teams often weaken research quality by forcing every run to end as if the question has a clean answer. Better systems are comfortable returning constrained conclusions.

What the final report should show

A healthy deep research report usually includes:

a short statement of scope,
the main conclusion,
the source basis behind each major claim,
the strongest known uncertainty,
and the next missing source that would improve confidence.

This is more useful than burying source ambiguity under polished structure.

Practical workflow rule

If the report quality is bad, inspect the source log before changing the model.

In many teams, the real fix is:

better source filters,
better prompt guidance,
better citation requirements,
and better review of evidence quality.

The model is often downstream of the real problem.

Implementation checklist

Your deep research source policy is probably healthy when:

source classes are defined before report generation starts;
primary sources are explicitly preferred;
duplicate-summary inflation is treated as a failure;
major claims require traceable support;
and the report can say “uncertain” without being treated as broken.