Five AI Research Tricks to Get Dramatically Better Results

Executive overview

Most people use AI research tools the way they use Google — single queries, single models, no strategy. The real gains come from matching the right model and tool mode to the type of question being asked, then layering in techniques like multi-model synthesis and structured prompting to reduce bias and surface more sources. This framework covers five concrete tricks that move AI research from basic search to a systematic, high-quality workflow.

The core insight: model choice, question type, and synthesis strategy matter more than prompt wording alone.

Choosing the right model

• Grok excels at real-time X/Twitter data and less-restricted responses on controversial or geopolitical topics.
• Perplexity offers toggleable source categories (academic, social, financial, general web) and consistently pulls from a more diverse source pool than ChatGPT or Claude alone.
• Claude uses a parallel multi-agent architecture: a lead agent orchestrates sub-agents researching simultaneously, typically pulling 300–400 sources versus 60–100 from other models.
• ChatGPT is reliable and broadly capable but rarely best-in-class for any specific research task or source type.

Matching the tool to the question type

Three categories of research questions require different tool configurations:

• Fact finding — basic lookups, quick answers. Use any model (GPT-4o or 4o-mini recommended for speed). Always toggle web search on explicitly; without it, models may use internal knowledge instead of live data. Adding "ensure info is up to date as of [today's date]" forces live retrieval.
• Breadth — trend analysis across many topics and sources, but a response needed quickly. Use high-end reasoning models (o3, Sonnet 4, Opus 4) with basic web search enabled, not deep research. Good for comparison tables across products or broad landscape questions.
• Depth — specific, well-scoped questions requiring comprehensive synthesis. Use deep research mode. Claude is the preferred model here due to parallel agents and higher source count. Example: "Summarise all studies since 2022 on the long-term metabolic effects of GLP-1 drugs in healthy adults; surface open questions."

Abundance over scarcity

• Advanced users ask the same question to multiple models simultaneously rather than relying on one output.
• For critical research, run deep research with the same prompt across ChatGPT, Claude, Perplexity, Grok, and Gemini.
• Okay approach: use one of the input models to synthesise the others — creates strong self-bias toward its own output.
• Better approach: use a third-party model not in the input set (e.g., feed GPT, Claude, Perplexity, and Grok outputs into Gemini) to reduce bias.
• Ideal approach: third-party synthesiser plus a specific aggregation instruction. Example: "If a study is cited in one report but missing from the others, include it in the final report as long as it was published after 2022." Vague instructions like "find the best insights" still produce biased outputs even with a neutral model.

Right source for the right job

• Perplexity source toggles (academic, social, financial) are the simplest way to constrain research to relevant source types.
• For models without toggles, use inclusion or exclusion prompts:
  • Inclusion: "Research X as of [date]. Focus primarily on Reddit, X, and Hacker News. Only consult other sources after exhausting these."
  • Exclusion: "Research X. Avoid forums and social media; focus on mainstream outlets and academic sources."
• Inclusion prompting is more reliable and easier to control than exclusion prompting based on practical experience.

The AI interview

• When the research goal is clear but the constraints and specifics are not fully articulated, use a reverse interview to externalise implicit knowledge before running deep research.
• Use a fast model (o4-mini) so the back-and-forth is quick.
• Prompt it: "Ask me one question at a time; each answer informs the next question. Goal: help me build a detailed deep research prompt on [topic]."
• The model asks targeted questions; at the end it produces a fully contextualised prompt ready for a deep research model.
• This converts vague intentions into specific, constrained research briefs — improving output quality significantly.