Research a Topic with AI Without Getting Misled

Tier: Core · Hands-on time: ~45 minutes

This playbook turns those into a concrete workflow you can finish in under an hour.

When to use

• •
  Technology evaluations ("should we adopt X?"), market/vendor comparisons, "what's the current state of Y" questions, unfamiliar-domain onboarding.
• •
  Any research that will inform a decision someone else relies on.
• •
  Not for questions where a primary source is one search away — go read the official doc directly.

Prerequisites

• •
  An AI assistant with web search/browsing enabled (most current chat tools have it — verify it's on; a model answering from training data alone is months stale).
• •
  A second, independent route: a different vendor's assistant, or a traditional search engine.
• •
  A destination document for findings (your notes, a wiki page).

The workflow

1. 1.

   Frame the question with decision context.

   Copy

   I need to research: <question>.
   Decision this informs: <what you'll do with the answer>.
   My current assumption: <state it — so you can notice confirmation>.
   
   Rules for this whole conversation:
   - Use web search; do not answer from memory alone.
   - Every factual claim gets a linked source and the source's date.
   - Distinguish clearly: ESTABLISHED FACT / VENDOR CLAIM / OPINION / YOUR
     INFERENCE. Label each claim with one of these.
   - If sources conflict, show the conflict instead of resolving it.
   - Tell me what I should be asking that I haven't.
   

2. 2.

   Interrogate the weak points. After the first answer, push on the claims that matter most to your decision:

   Copy

   For claims <2> and <5>: open the actual sources. Quote the exact passage
   supporting the claim, with the page's publication date. If the source
   doesn't directly support the claim, say so plainly.
   

3. 3.

   Click through yourself. Open the 3–5 sources behind your decision-critical claims. Check: the page exists, actually says what was claimed, is dated within a useful window, and is what it appears to be (a vendor blog is a VENDOR CLAIM regardless of how it was labeled).

4. 4.

   Triangulate via an independent route. Ask the same core question of a second assistant (different vendor) or run a plain web search yourself — without sharing the first answer, so the second route can't just agree. Diff the results:

   • •
     Agreement on a claim from independent sources → promote to trusted.
   • •
     Disagreement → the interesting part; investigate manually, this is often where the real answer lives.
5. 5.

   Write findings with confidence labels. Record each key finding as: claim → source(s) with dates → confidence (High / Medium / Low) → what would change your mind. Findings without a surviving source get dropped or marked "unverified, needs follow-up."

Tool variants

• Search-enabled chat assistants (any two major vendors): the workflow as written; using two different vendors for triangulation guards against shared blind spots.
• Dedicated deep-research modes: several assistants offer a long-running research mode producing a cited report — good first drafts, but they average sources rather than adjudicate them; steps 3–5 still apply in full.
• Traditional search + AI summarization: find sources yourself, then have AI summarize each (pb-01) — slowest, but strongest provenance; right choice for high-stakes decisions.

Verification

• •
  Every decision-critical claim: source opened by you, passage found, date checked.
• •
  Independence check: your two routes cited at least some different sources (identical source lists mean you triangulated one echo).
• •
  Labels audit: no VENDOR CLAIM silently upgraded to fact in your write-up.
• •
  Recency: sources for fast-moving claims (pricing, model capabilities, market share) are months old, not years — check dates, not vibes.
• •
  Your original assumption is explicitly confirmed, refuted, or marked unresolved — if the research just agreed with you everywhere, suspect confirmation bias in your prompts.

Pitfalls

• Fluent averaging. AI blends conflicting sources into one smooth, wrong paragraph. The "show conflicts" rule counters this — enforce it.
• Leading the second witness. Pasting route 1's answer into route 2 turns triangulation into agreement theater.
• Citation laundering. Blog A cites blog B cites a tweet. Follow decision-critical claims to a primary source (spec, paper, vendor doc, filing).
• Stale training data as fresh news. If search wasn't actually invoked, "current" answers can be a year old. Ask "what is the date of your most recent source?"
• Over-researching. Triangulation is for load-bearing claims; timebox the rest. Not every fact deserves the full treatment.

Restricted-environment variant

• •
  Research questions leak strategy ("comparing vendor X and Y for a 200-seat rollout" reveals plans). Neutralize prompts: ask about the category generically, keep volumes, budgets, and client names out.
• •
  Never paste client RFP text or internal strategy docs into a hosted tool as "context" for research; summarize the need in your own generic words.
• •
  Fully restricted setting: local models (pb-07) can't browse, so invert the flow — gather sources manually in your browser, save as text, and use the local model for the summarize-and-compare steps only.

Success criteria

• •
  A findings document where every key claim carries a source, a date, and a confidence label.
• •
  At least one claim from the first AI pass was corrected or downgraded by your verification — if nothing failed, your checks weren't real.
• •
  Two independent routes agree on the claims you'll act on.
• •
  Total time under an hour for a typical technology-evaluation question, with reusable prompts saved to your library (pb-05).