::industry track · marketing ai

Marketing AI. The variant engine — not the strategy oracle.

Marketing is the field where AI feels most useful and gets misused the fastest. You can spin twenty subject lines in a minute. You can A/B/C/D ad copy until you forget what you were testing. You can ghostwrite a newsletter that reads exactly like every other newsletter on the internet — which is the problem. Volume without voice is how brands quietly die. The honest framing: AI is a variant engine. It is excellent at producing the seventh, eighth, and ninth version of an angle you already understand. It is terrible at deciding which angle matters, what your audience actually wants, when to spend, what to price, or how to position against a competitor. Those are operator decisions. They always will be. The marketers winning with AI in 2026 share three habits. First, they keep a strict brand-voice primer — a documented rubric the model gets fed before every generation, so output sounds like them and not like beige LinkedIn. Second, they never cite an AI-generated statistic, study, or quote without independent verification — because the FTC is now actively enforcing against AI-fabricated marketing claims, and "the AI made it up" is not a defense. Third, they treat AI as a faster draft, not a smarter strategist. Decisions stay human. Variants get automated. This track gives you the workflows, the do-not list, the regulatory context, and a real case study of an indie marketer who survived the brand-voice-dies-in-volume trap by building a Skill Primer she actually maintains.

::the safety rules · non-negotiable

What you never do.

• 01Never publish an AI-generated statistic, study citation, or quote without independent verification from a primary source — the FTC actively enforces against fabricated marketing claims under Section 5 of the FTC Act.
• 02Never paste customer PII (emails, names, billing data, support tickets with identifying info) into a public AI chat — use anonymized excerpts or local/enterprise-tier AI only.
• 03Never paste unreleased pricing, product roadmap, or M&A-adjacent strategy into public AI chats — assume training data exposure risk.
• 04Never paste competitor confidential material (leaked decks, NDA'd interviews, scraped internal docs) — legal exposure plus model-poisoning risk.
• 05Never publish AI-generated testimonials, reviews, or fake user quotes — illegal under the FTC's 2024 endorsement rule revisions, fines start at five figures per instance.
• 06Never send to EU/UK audiences without verified GDPR consent, regardless of how the list was assembled — AI-generated lookalike lists do not bypass consent law.
• 07Never auto-publish AI output to live channels without a human editorial pass — every state-level AI disclosure law emerging in 2026 assumes human review.
• 08Never use AI to mimic a real person's voice (a real journalist, a real customer, a competitor's exec) in published marketing material — defamation and right-of-publicity exposure.
• 09Always disclose AI involvement in any context where a reasonable audience would assume human authorship and the deception would be material — FTC deceptive-practices standard.
• 10Never train your brand-voice primer on copyrighted material you do not own or have license to use — corpus contamination is a real legal risk.

::the workflows · 8 named plays

The plays. Each with the exact prompt.

You are running a subject-line lab for [BRAND NAME]. Brand voice primer is attached. The email is going to [SEGMENT — e.g., active trial users on day 5 who have not yet imported data]. The email body is: [PASTE BODY]. The single goal of this email is [GOAL — e.g., get them to book a 15-min onboarding call]. Produce FOUR subject lines, each testing a distinct angle: (A) Curiosity gap — implies missing information. (B) Specific benefit — names the concrete outcome. (C) Loss framing — names the cost of inaction. (D) Pattern interrupt — unexpected phrasing or formatting. For each, give: subject line (under 50 chars), preview text (under 90 chars), the angle being tested, and one risk (e.g., 'might read as clickbait to enterprise buyers'). Do not produce four variations of the same angle. Use the brand voice primer strictly — no exclamation points unless the primer allows them, no emoji unless the primer allows them.

Attached is the [BRAND NAME] brand voice primer (tone, banned phrases, signature rhythms, approved register). Below is a draft of [ASSET — e.g., a launch announcement, a Twitter thread, a sales page section]. Audit the draft against the primer. For each section, identify: (1) phrases that match the brand voice — quote them and say why. (2) phrases that drift toward generic/beige — quote them, say what's wrong, propose a voice-locked replacement. (3) banned phrases (per the primer) that appear — flag and replace. (4) signature rhythms (per the primer) that are missing where they should appear. Do not rewrite the whole draft. Surgical edits only. End with a one-sentence verdict: 'ship as edited' / 'one more pass needed' / 'voice is too far off, restart.' Draft: [PASTE DRAFT]

I'm running [PLATFORM — Meta/LinkedIn/TikTok] ads for [PRODUCT] targeting [AUDIENCE — be specific: role, company size, pain]. The single conversion event is [EVENT — free trial / demo book / lead magnet download]. Brand voice primer is attached. Produce 8 ad copy variants in 4 hook categories (2 per category): (1) Problem-aware — names the pain in their language. (2) Outcome-aware — names the after-state they want. (3) Mechanism-aware — explains the unique-how. (4) Social-proof — anchors on real customer evidence. For each variant: hook (first 125 chars — what shows above the fold), body (under 200 chars), CTA. Flag any variant that makes a quantitative claim (percentages, dollar amounts, timeframes) so I can verify before shipping. Do not invent customer names, stats, or testimonials. If a social-proof variant needs evidence, write '[INSERT REAL CUSTOMER QUOTE HERE]' as a placeholder.

I'm running a retro on [CAMPAIGN NAME] which ran [DATE RANGE]. Below is the raw data from each funnel stage: (1) Top of funnel — impressions, CTR, CPC per channel: [PASTE]. (2) Mid-funnel — landing page conversions, asset downloads, demo books: [PASTE]. (3) Bottom of funnel — trial-to-paid, demo-to-close, deal size: [PASTE]. (4) Qualitative — sales team notes, support tickets from new users, NPS comments: [PASTE]. Produce a structured retro: (a) What worked — three claims, each grounded in a specific number from the data. (b) What did not work — three claims, same standard. (c) What surprised us — anomalies in the data, hypothesis for each. (d) What we still don't know — gaps where the data cannot tell us yes/no. (e) Three concrete bets for the next campaign, each with a hypothesis and a metric to test against. Do not invent numbers. Do not extrapolate beyond what the data shows. If a claim requires data I didn't paste, say so explicitly.

Scan publicly available sources for trend signals in [CATEGORY — e.g., B2B vertical SaaS for the construction industry] from the last [TIMEFRAME — e.g., 30 days]. Produce three signals, each with: (1) The signal — one sentence, concrete and falsifiable. (2) Evidence — at least two primary sources (news article, public company filing, industry report, regulator filing) with URLs. I will verify every URL before using any of this. (3) Why it matters for [BRAND] specifically — one sentence. (4) Confidence — high/medium/low and what would lower the confidence. Do not pull from social media trend lists. Do not invent sources. If a signal only has one source, flag it as low confidence. If you cannot find evidence for a signal, do not include it. Three real signals beat ten speculative ones.

Track publicly visible pricing and positioning changes for these competitors: [LIST 3-5 COMPETITORS WITH URLs]. For each, produce: (1) Current published pricing — exact tiers, prices, what's included. Cite the pricing page URL with timestamp. (2) Positioning — the headline on their homepage and the first sentence of their about/manifesto page. Cite URLs. (3) Recent visible changes — anything that has changed in the last 90 days (new tier, removed tier, renamed plan, new positioning line). Only include changes you can evidence with a current page + a Wayback Machine archive showing the prior version. URL both. (4) What this might signal — one cautious sentence per competitor, framed as a question not a conclusion. Do not invent prices. Do not extrapolate from screenshots I have not seen. If you cannot verify a change with two sources, omit it.

Below are [N] anonymized win/loss interview notes from [QUARTER]. Each interview is tagged WIN or LOSS and includes: deal stage, ICP segment, primary competitor encountered, top-3 buyer concerns, why-we-won or why-we-lost in the buyer's own words. [PASTE NOTES]. Produce a synthesis: (1) Win patterns — three patterns that appear in WIN notes but not LOSS notes, each cited with interview IDs. (2) Loss patterns — three patterns that appear in LOSS notes but not WIN notes, same standard. (3) Mixed signals — patterns that appear in both, requiring deeper investigation. (4) Quotes — pull six exact quotes (three wins, three losses) that capture the strongest patterns. Do not paraphrase quotes. Do not invent quotes. If a quote needs editing for anonymity, mark the edit with brackets. (5) Three concrete recommendations — one for sales enablement, one for marketing messaging, one for product. Each recommendation must cite the pattern it addresses.

I am preparing a public response to [SITUATION — describe the facts neutrally, no spin]. Brand voice primer is attached. Produce THREE distinct response postures, each as a 2-3 sentence statement: (A) Accountability-forward — acknowledge directly, name the action, no hedging. (B) Information-forward — clarify the facts that may have been misrepresented, calm and specific. (C) Restraint-forward — minimal statement, decline to amplify, redirect to action. For each posture: the statement, what audience it serves best, the risk if we pick this one, what factual claims (if any) need legal review before shipping, what we'd need to verify before publishing. Do not pick a winner. Do not editorialize. This is a decision for me and legal — your job is to make the three options clearly distinct so the choice is real, not theatrical.

::case study · composite anonymized

One human. One year. The pattern.