DeepResearch Output Quality Control: Model Selection, Prompt Tuning, and Citation Verification

DeepResearch output quality varies - sometimes you get a 5-minute fluffy report, sometimes a 20-minute deep dive. This article turns "luck" into "process".

1. Model selection

Experience: DeepResearch runs on the main model tier - switching to o-series changes model capability (slower, pricier); don't use it for light research.

2. The 5-part prompt - the QC core

Strictly five parts. Miss one and quality drops:

1. Background (why research)
   "I'm [role], currently in [context] facing [specific problem]."

2. Scope (include / exclude / time window / source preference)
   "Include A, B, C; exclude D, E. Time window 2024-2026. Sources: official announcements + media reports + third-party analysis."

3. Goal (reader / decision)
   "Output to the CEO as input for the decision on whether to launch localization in H2 2026."

4. Format (sections / tables / citations)
   "Markdown with 5 sections (market state / key players / user profile / growth drivers / risks), 3 comparison tables, citations as [link format]."

5. Depth (specific data / comparisons / recommendations)
   "Include specific user counts, pricing, growth rates; not a fluffy compilation; with actionable recommendations."

3. Citation verification - the most important step

Every link the model gives must be opened and checked - AI hallucination can fake citations.

Three verification steps:

1. Every key conclusion - open the citation and check against the original: the model can misread, distort, or quote out of context
2. Key data - go back to the primary source: first-hand statistics (company financials / SEC filings), policy dates (government announcements), medical data (peer-reviewed journals)
3. Cross-verify across sources: a fact only earns trust when two independent sources agree

Common errors:

• Model cites a paper that doesn't exist
• Model cites a real paper but misinterprets it
• Model fabricates plausible-looking statistics
• Model attributes 2024 events to 2026

Citation verification takes ~15 minutes per report - but that's cheaper than acting on wrong conclusions.

4. Iteration feedback loop

DeepResearch isn't one-shot - the model is a lazy learner and will replicate wrong patterns next time:

Round 1: 5-part prompt → 80% accurate output
Round 2: Rewrite prompt based on feedback (emphasize exclusions + data format) → 90%
Round 3: Tune again → 95%

Feedback forms:

• ChatGPT's 👍 / 👎 buttons - influence long-term model behavior
• Explicit continuation: "redo based on the following feedback: [content]"
• Save good outputs as templates - reuse directly on similar tasks

5. Structured output - explicitly request it

The model's default output can be scattered - explicitly require structure:

Output: Markdown with these sections
1. Market state (with 2024-2026 trend data)
2. Key player comparison (table: player / users / pricing / growth rate)
3. User profile (demographics + behavior)
4. Growth drivers (≥3 specific drivers)
5. Risks (≥3 specific risks)

Each section ends with a citation list (Markdown link format).

Structured output lets you systematically verify, without missing key points.

6. Cost estimate

Cost is mainly time, not money. Block off an async window (5-30 min) up front; don't leave the user waiting.

7. Practical SOP

Scenario: Competitive product research (output for PM)

1. Triage the tier: Pro default is sufficient (5-part + explicit format)
2. Write the prompt: full 5 parts; emphasize "include specific user counts, pricing, growth rate, max 5 pages markdown"
3. Launch + wait 15 min: close the tab, do other work
4. Citation verification: 5 key conclusions in the report must open and verify - this takes 15 min
5. Feedback: errors / omissions get reinforced next time; save the good output as a template

8. Common errors and troubleshooting

• Output reads like an "AI overview" → 5-part's "goal" or"depth" section wasn't specific; the model doesn't know how to go deeper
• Citations are mostly 404 / irrelevant → "scope" section didn't specify source preference; emphasize "official / academic / media" category
• All data is "approximately"" → "depth" section didn't say "specific numbers"; the model defaults to vague
• Conclusions don't match the citations → citation verification wasn't done; the model reported 2024 as 2026 - go line by line
• Multiple runs give unstable results → same prompt run multiple times is model behavior - pick the best one

9. What's Next

• Deep Research Prompt Patterns and Quota Strategy - 5-part prompt templates
• GPT-5.6 Selection Guide - choosing the model tier
• ChatGPT Subscription Plans Compared - choosing the plan

Update log

• 2026-08-08: Initial publish