Conduct business research

Conduct Business Research

Desk/secondary research done as a discipline: find what is already answered before paying to answer it again, and convert what is NOT answered into the primary-research agenda. The gap list is the deliverable — the reading is not.

When to use / when NOT to use

Use when:

• A project, discovery phase, or client engagement is starting and the market/domain is unfamiliar
• Anyone is about to plan primary research (interviews, surveys, tests) — this pass runs FIRST
• A strategy, pitch, case study, or recommendation needs external claims and none are cited yet
• The ask is a market scan, domain scan, industry overview, or "what's out there on X"

Do NOT use for:

• Tearing down competitor products or comparing their usability → run-competitive-analysis (this skill only records competitors as market facts, then hands off)
• Framing what problem to solve → write-problem-statement (run it first; its discovery goals seed the research questions here)
• Planning the primary study itself → write-research-plan and choose-research-method (they consume this skill's gap list)
• Sizing opportunities or picking which problem to attack → identify-business-problems
• Building the differentiation story → position-product
• Answers that live in a person's head, not a document → conduct-stakeholder-interviews
• Deep analytics interpretation or A/B design → use-quantitative-evidence

The method

Step 0 — Load craft-critique. Every external claim this skill produces is handled per its evidence protocol (cite / get / flag). That protocol is defined there and only there — do not improvise a local version.

Step 1 — Anchor to a decision. Write one sentence: "This research informs [decision], made by [who], by [when]."

• No decision named → no research. Push back and get one. Research without a decision anchor becomes the "black hole of interesting data" (NN/g's phrase) — collection with no action.
• Then write 3–5 research questions: method-agnostic, answerable, each traceable to the decision. These are NOT search queries; they are the questions the decision needs answered. If write-problem-statement already produced discovery goals, derive the questions from those.

Step 2 — Timebox the pass. Desk research is a bounded sweep, not a reading project. Default: 0.5–2 days (NN/g industry data: most full discoveries run under 2 weeks — the desk pass is a fraction of that). Declare the timebox before opening the first source. Stop early at saturation: when new sources repeat what the inventory already holds.

Step 3 — Inventory INTERNAL sources first. Cheapest, most specific, most ignored:

• Prior research reports and study notes (check the research repository / archive first — NN/g atomic-research logic: past studies answer future questions)
• Analytics and product data already collected
• Support tickets, sales notes, client emails, CRM records
• Past decision records and archived workpacks
• Anything a teammate/client already wrote about this space

If an internal answer exists, the question is ANSWERED — do not re-research the known.

Step 4 — Scan EXTERNAL sources down the quality ladder. Start at Tier 1; only descend when a tier is exhausted for the question.

Tier	Source type	Status in output
1	Primary data: government statistics, regulators, census, company financial filings, peer-reviewed studies	Citable as-is
2	Named-methodology industry research: established research firms, academic industry reports, NN/g-grade practitioner research	Citable with date + methodology noted
3	Journalism and trade press: named author, named outlet, dated	Citable for events/facts; opinions labeled as opinions
4	Vendor content, company blogs, SEO listicles, "State of X" marketing reports, social posts	Leads only — never citations. Use them to find the Tier 1–2 origin

Scan rules:

• Chase claims to origin. If source B cites source A, cite A. If A cannot be found, the claim drops to the tier of B — or gets flagged.
• Triangulate load-bearing claims. Any claim the decision rests on needs 2+ independent sources (independent = not tracing back to the same origin). A single-source load-bearing claim gets flagged per the evidence protocol.
• Set a recency window matched to the market's rate of change (AI products: ~12 months; regulated/slow industries: 2–3 years). Note the window in the output; stats outside it get flagged as stale.
• Separate fact from opinion. "Analyst predicts X by 2030" is an opinion about the future, never a market fact. Record it labeled as a prediction, with the predictor named.

Step 5 — Extract into an evidence inventory. Atomic findings — one claim per row, saved with enough context to be reused without re-reading the source:

Claim	Source (linked)	Date	Tier	Status
one specific, checkable statement	origin source, not the article that quoted it	publication date	1–4	CITED / UNDER-EVIDENCED

Then mark each research question: ANSWERED / PARTIAL / OPEN, pointing at its inventory rows.

Step 6 — Build the gap list → primary-research agenda. This is the payload. Every PARTIAL or OPEN question, plus every load-bearing claim still under-evidenced, becomes a gap entry:

• Gap: what exactly is unknown
• Why it matters: which decision breaks if this stays unknown
• Cheapest way to close it: named method (hand to choose-research-method; the agenda feeds write-research-plan)

A gap the decision doesn't need → cut it. Not every hole gets filled.

Step 7 — Synthesize with a recommendation. State what the evidence supports NOW, what it cannot support yet, and one recommended next action. A scan without a recommendation fails — presenting data without insight is not finished work.

Worked example

(Structure is the lesson — rows illustrate the discipline, not verified findings.)

Context: designing a conversational shopping experience for voice-first commerce aimed at first-time internet users in India (his Amazon Echo Show concept project).

Decision anchor: "This research informs which shopping tasks the concept designs for first, decided by me, before concept exploration starts next week." Timebox: 1 day. Recency window: 24 months.

Research questions:

1. How do first-time internet users in India currently discover and buy products online?
2. What share of that population prefers voice/regional-language input over typed English?
3. What has already been tried in voice commerce there, and where did it stall?

Evidence inventory (excerpt):

Claim	Source	Date	Tier	Status
Regional-language internet users outnumber English-language users in India	Telecom regulator annual report	2025	1	CITED
"Voice search is the primary input for next-billion users"	Vendor blog citing its own survey; original methodology not published	2024	4	UNDER-EVIDENCED — need a Tier 1–2 origin or own survey
Two major platforms launched, then scaled back, voice-shopping features	Trade press, named reporters, two independent outlets	2024	3	CITED (event fact)

Question status: Q1 PARTIAL · Q2 OPEN (only Tier 4 sources found) · Q3 ANSWERED for "what was tried," OPEN for "why it stalled."

Gap list → primary-research agenda:

1. Gap: actual input-mode preference (voice vs. typed) among target users. Why: the core interaction model rests on it. Close: 6–8 user interviews with first-time internet users; hand to write-research-plan.
2. Gap: why prior voice-shopping efforts stalled. Why: avoids repeating a known failure. Close: teardown of the two scaled-back products via run-competitive-analysis + practitioner postmortems if published.

Recommendation: Evidence supports designing for regional-language-first users now (Q1/Tier 1). It does NOT yet support voice as the default input — that is the highest-risk assumption; close gap 1 before concept selection.

Anti-patterns / red flags

• Primary research before the desk pass. Commissioning interviews to learn what a Tier 1 report already says is the exact waste this skill exists to prevent.
• Citation laundering. Five articles that all trace to one vendor press release counted as five sources. Independence means independent origins.
• The scan as the deliverable. A 30-source reading list with no gap list and no recommendation is homework, not research.
• Market-size numbers pasted bare. Any TAM/market-size figure without date + methodology + origin gets flagged, not repeated. Wildly divergent market-size numbers across sources is itself a finding.
• Opinion laundered as fact. Analyst predictions, founder quotes, and trend pieces recorded as market facts.
• Internal context recycled as external evidence. Synthesizing "the market wants X" from the team's own beliefs or prior positioning — the exact failure craft-critique's protocol exists to stop.
• No timebox. "Still reading" past the declared window means saturation was never checked or the decision anchor was never real.
• Tier 4 in the citations. Vendor content may open a trail; it never ends one.
• Filling every gap. Gaps get closed in decision-priority order; a gap no decision needs gets cut, not scheduled.

Output format

## Desk Research — [decision it informs]

**Decision anchor:** [decision, decider, deadline]
**Timebox:** [declared] · **Recency window:** [window]

### Research questions
1. [question] — ANSWERED / PARTIAL / OPEN

### Evidence inventory
| Claim | Source (linked) | Date | Tier | Status |
|---|---|---|---|---|

### Gap list → primary-research agenda
1. **Gap:** … **Why it matters:** … **Cheapest close:** [method → hand to write-research-plan]

### What the evidence supports now
[Only CITED, triangulated claims]

### What it cannot support yet
[Under-evidenced items, named per craft-critique's protocol]

**Recommendation:** [one next action]

Sources

• Secondary research in UX (desk research as the step before primary): https://www.nngroup.com/articles/secondary-research-in-ux/
• Discovery phase (problem-space research, evidence over opinion, discovery timeboxes): https://www.nngroup.com/articles/discovery-phase/
• Triangulation — 2+ methods/metrics on the same question, single-source claims are weak: https://www.nngroup.com/articles/triangulation-better-research-results-using-multiple-ux-methods/
• Analytics for UX — the "black hole of interesting data," decision-anchored measurement: https://www.nngroup.com/articles/analytics-user-experience/
• Research repositories (inventory internal knowledge before researching anew): https://www.nngroup.com/articles/research-repositories/
• Atomic research — one evidence-backed insight per item, saved for reuse: https://www.nngroup.com/videos/atomic-research-small-insights-big-impact/
• Competitive usability evaluations (the sibling this skill hands competitor work to): https://www.nngroup.com/articles/competitive-usability-evaluations/
• UX research plans (what the gap list feeds): https://www.nngroup.com/articles/pm-research-plan/

Boundaries

• craft-critique is the single source of the evidence-discipline protocol. This skill applies it to every external claim; it never restates it.
• run-competitive-analysis owns competitor teardowns and competitive usability evaluation. This skill records "competitor X exists / launched / retreated" as market facts and hands deeper analysis off.
• write-problem-statement owns problem framing and runs before this skill; its discovery goals seed the research questions.
• write-research-plan and choose-research-method own primary research. This skill's gap list is their input — it never plans the study itself.
• identify-business-problems owns opportunity sizing, risk-reward quadrants, and milestone anchoring of what the research surfaces.
• conduct-stakeholder-interviews owns answers that live in people. Desk research inventories documents and data; when the trail ends at "ask the person who knows," hand off.
• use-quantitative-evidence owns analytics interpretation depth and A/B design; this skill only inventories analytics as an internal source.