Occupation manual 006 / Market research & marketing
A disciplined route from research question to source ledger, participant language, insight synthesis, positioning hypothesis, claim review, one approved campaign experiment, and an honest learning log.
Short answer: use AI as an evidence clerk, not a participant, customer, expert, analyst of record, advertiser, publisher, media buyer, or compliance approver. It may organize approved sources, group themes, surface contradictions, draft hypotheses, and prepare experiment briefs. It must not invent research, manufacture testimonials, claim causation, approve public claims, profile people unlawfully, publish, send outreach, launch ads, or spend.
The U.S. Census Bureau describes questionnaire design and evaluation as work that uses qualitative and quantitative methods to reduce measurement error. AI can help organize approved notes and test alternative wording, but it cannot create respondents, consent, or a representative sample.
FTC guidance says advertising claims and endorsements must be truthful and not misleading. Testimonials do not automatically substantiate objective claims, material connections need appropriate disclosure, and paid content must not masquerade as independent editorial work.
Record the decision, audience/population, current hypothesis, evidence needed, approved methods, privacy/consent owner, reviewer, and who controls publishing, outreach, ads, and spend.
Log the exact source, date, context, direct support, class, limitations, and bias. Separate primary evidence, participant statements, platform metrics, vendor claims, inference, and uncertainty.
Use approved, consented, minimum-necessary data. Preserve the question and faithful answer while keeping researcher observation separate. AI must never invent a participant, quote, response, demographic, session, or consent.
Group themes with supporting and contradicting evidence IDs, frequency only within the actual sample, confidence, alternative explanations, and the next evidence needed.
Connect audience, situation, painful workflow, existing workaround, proposed outcome, proof, reason to believe, and the claim you are not ready to make. Name what could disconfirm the position.
Tie every proposed public claim to evidence. Check objective proof, typicality, results, material connections, commercial context, and high-risk topics. Approve, revise, or reject.
Define one hypothesis, audience, useful asset, approved claim/disclosure, exact CTA, primary metric, guardrails, decision rule, timebox, privacy handling, budget authority, and publish/send owner.
Log exact version/channel, denominator, primary and guardrail results, data-quality limits, decision, and next test. Do not convert a conversion change into an unsupported causal explanation.
| Job | Approved input | AI output | Human gate |
|---|---|---|---|
| Source-control clerk | Approved sources and method notes | Evidence ledger with missing fields | Researcher verifies provenance, permission, and direct support |
| Language-coding clerk | Approved participant notes with identifiers minimized | Candidate themes and evidence IDs | Researcher checks fidelity, contradictions, and sample limits |
| Hypothesis clerk | Reviewed evidence matrix | Audience/positioning alternatives and disconfirming tests | Marketing owner selects what is worth testing |
| Claim-review clerk | Proposed copy and evidence ledger | Unsupported-claim, disclosure, and review flags | Qualified human approves, revises, or rejects |
| Experiment-prep clerk | Approved hypothesis, channel rules, and measurement plan | Draft asset and experiment packet | Authorized human controls launch, audience, budget, and account |
A participant can describe a real frustration without representing the whole market. Preserve exact context, sample, method, and limitations. Report “4 of 7 participants in this convenience sample mentioned…” when that is what happened—not “57% of customers” or “everyone wants…”
FTC resources distinguish genuine endorsements from fabricated or misleading proof, and objective claims need appropriate substantiation. Never create a fictional customer, expert, case study, quote, screenshot, review, result, or “typical outcome.” If an endorser has a material connection, the exact disclosure and placement require human review.
The FTC's dark-pattern work describes tactics such as disguising ads, burying key terms or fees, making cancellation difficult, and tricking people into sharing data. A higher click or conversion rate does not make a manipulative design acceptable. Put consent, comprehension, cancellation, privacy, and complaint signals into the experiment guardrails.
Google says helpful content should serve people and provide original information, reporting, research, or analysis. Use clear answer summaries, descriptive titles, visible dates, authorship/review, direct source links, structured data that matches visible content, and useful artifacts. Avoid scaled pages, scraped rewrites, doorway variants, hidden text, and thin pages aimed mainly at ranking or influencing generative search.
Act as a customer-evidence research clerk, not a participant, customer, expert, analyst of record, advertiser, publisher, media buyer, or compliance approver. Using only the supplied approved evidence, produce: (1) a source ledger, (2) themes with supporting and contradicting evidence IDs, (3) alternative explanations and limitations, (4) audience/positioning hypotheses labeled as hypotheses, (5) claims that are and are not currently supportable, and (6) a draft experiment brief that stops before publishing, outreach, targeting, ads, or spend. Do not invent people, quotes, sessions, survey responses, sample sizes, metrics, citations, testimonials, results, typicality, consent, or causation. Preserve uncertainty and private-data boundaries.
Download the desk, start with approved evidence, keep contradictions visible, and ship only the smallest human-approved experiment that can teach you something.