Three AI research modes: auto, broad, and deep

Executive overview

Most people use AI for research the same way regardless of the task, wasting time or getting shallow results. A simple two-axis framework — stakes vs. time available — determines which of three research modes to use.

Auto covers ~50% of queries, broad ~40%, and deep ~10% — matching the mode to the task cuts research time by up to 80%.

The auto mode: fast fact-finding

• Use for basic questions: what, where, when — the kind previously handled by Google.
• Leave ChatGPT on its default (auto) setting; no model changes needed.
• Dictation or advanced voice mode speeds this up further when on the move.
• Best for low-stakes questions where speed matters more than depth.

The broad mode: analysis and comparison

• Use when you need confidence in the response and have slightly more time.
• Requires a high-reasoning model (e.g. ChatGPT "thinking" with extended reasoning) plus web search enabled.
• Prompt structure matters: include your role/context, what you want compared or shortlisted, and the exact output format (e.g. a scoring matrix or comparison table).
• Comparison example: ask ChatGPT to compare QuickBooks vs. Xero, but add your org type, size, and service category — then request a table of pricing and three "choose this if" recommendations.
• Shortlisting example: provide 10 tools, state your weighting criteria (accuracy, CRM sync, price, security), and request a scored matrix with links to pricing and security pages.
• Real-world use: upload a photo of a broken spigot, enable thinking + web search, ask what's broken and which nearby stores carry the part — the model returns local store options open right now.
• "Thinking" mode is usually as good as "Pro" and much faster; use extended reasoning for broad searches.

The deep research mode: high-stakes reports

• Use for complex, high-stakes questions where a 10–30 page report is warranted.
• Expect 8–20 minutes wait time; ask for a one-page summary or checklist in the same conversation if needed.
• Enable via the "+" menu in ChatGPT → "deep research"; external web sources are the default, but internal data sets (Google Drive, OneDrive) can also be targeted.
• The model will ask 3–5 clarifying questions before starting — answer all of them in a single message using Shift+Enter between answers, not Enter (which triggers early research).

Writing an effective deep research prompt

Structure the prompt around three components:

1. Scope — narrow by time period (e.g. 2019–2025), region, or market segment.
2. Intent — state your goal and what you will do with the output; this shapes the report's focus.
3. Sources — optionally specify source types (government, academic); the model will weight toward them but cannot guarantee exclusivity.

A strong prompt also includes: a role/persona at the top, a context-rich task description, an explicit list of what to include in the report, and output format instructions (e.g. benchmark table, key trends, recommendations).

Bonus: multi-model deep research for critical decisions

• For the highest-stakes decisions, run the same deep research prompt across every available model: ChatGPT, Claude, Gemini, Perplexity, Grok.
• Read each report, then feed all of them into a separate AI and ask it to synthesise a single holistic report pulling the strongest elements from each.