How to Write Product Descriptions With AI (Without the Mush)

Why AI Product Copy Fails: The Input Problem, Not the Model

The mush everyone complains about is not a model failure; it is an input failure. A general model asked to describe a camp chair with no evidence has exactly one strategy available: adjectives. Durable, premium, crafted, versatile, the same words it would write for every other camp chair, delivered with total confidence. To write product descriptions with AI that sell, you have to change what the model is given before you change which model it is.

The second failure is governance, not generation. Stores that let the model ship directly to the page discover invented specs, borrowed press mentions and confident nonsense after customers do, and one invented burn-time or load rating costs more than the entire quarter's copy savings. The pipeline below fixes both failures with documents, not prompts: a voice guide that survives scale, a proof bank that feeds evidence instead of vibes, and an audit that catches the model's signature lies before they ship.

The Pipeline at a Glance

The full assembly line has five stations, and the order is load-bearing: documents before drafting, drafting before humans, humans before audits, audits before shipping. Each station produces a checkable artifact, which is what makes the system governable by one part-time editor instead of a content team. The stations:

• The voice guide, ten checkable rules, a banned list, three golden samples
• The proof bank, one row per SKU of specs, evidence and objection answers; no row, no draft
• Structured drafting, one master prompt per page type, with placeholders that route judgment to humans
• The human pass and fact audit, ten minutes of editing, three verdicts on every claim
• The QA gate and two-week read, banned-list search, voice check, then the PDP-to-cart gauge

Step 1: The Voice Guide (Ninety Minutes That Save a Catalog)

The voice guide is ten checkable rules, a banned list, and three golden samples, one hero page, one mid-tier, one long-tail, written to those rules. Rules that translate to any store: lead with what the product does for the buyer's life, numbers over adjectives, second person, one metaphor per page, no exclamation marks, every claim needs its proof tag, end with the guarantee. The point of checkable is literal, an editor can verify each rule in seconds, which is what makes the guide govern a thousand pages instead of decorating three.

Ten rules and a banned list

The banned list is half the document's value: the six words your category's clichés hide behind, enforceable by search. Every catalog has them, the words that sound like marketing because they are the sound of marketing, and naming them converts a mood into a rule. Enforce the list with a thirty-second search per page during QA, and add every offender the QA catches to the list with a date, so the guide learns at the speed the catalog ships.

Three golden samples

The golden samples anchor the whole system. When a reviewer asks does this sound like us, the answer is a comparison against a named artifact, not a mood, and when the master prompt is built, the samples ship inside it. Write them yourself or sharpen your best existing page; either way they become the definition of good that every later draft is judged against, which is why they are worth their ninety minutes.

Step 2: The Proof Bank (No Row, No Draft)

The bank's discipline is absolute: no SKU gets drafted until its row is complete. The fill is faster than it sounds, specs live in your product data, test results in spec sheets, the best quotes in your review export, the pre-emptive answers in your support tickets. One afternoon of exports fills most of a mid-size catalog's bank. What the exercise surfaces along the way is equally valuable: the claims in your current copy that no evidence supports, which is the meeting where marketing and legal become friends.

The bank also fixes the tier question. Hero SKUs, the pages carrying most of your traffic, get filled first, drafted first and measured first, so the pipeline's earliest outputs land where the money is. Traffic-weighted sequencing is what buys organizational patience: a measured win on the first tier funds the long tail, and the long tail is where the model's speed advantage compounds into catalog-scale economics.

Step 3: Structured Drafting With Judgment Routed to Humans

Drafting runs on one master prompt per page type, and the design principle is routing: the model handles structure and evidence, and every genuine judgment call comes back to a human as a placeholder. The prompt carries the voice guide, the proof-bank row and the page anatomy, and returns a structured draft rather than prose soup. The stations:

• One master prompt per page type, voice guide, proof-bank row and page anatomy in; structured draft out: fold argument, proof stack, objection answers, spec strip, guarantee line
• Placeholders like [CHOOSE: which quote belongs above the fold] are deliberate, they route judgment to the human instead of letting the model flatten decisions into mush
• One draft per SKU per pass, regeneration loops are how voice drifts; if the draft misses, fix the row or the prompt, not the dice
• Tier order, hero SKUs first, measured first, then the long tail once the gauges have spoken
• The fold argument is the contract, if the first screen cannot state what the product does for the buyer's life, the draft goes back before the human pass starts

Step 4: The Human Pass and the Fact Audit

Two gates stand between draft and page, and they are deliberately different kinds of attention. The human pass protects voice; the fact audit protects truth. Neither is optional, both take minutes, and together they are why the pipeline ships pages you would sign your name to.

The ten-minute pass

The human pass has four moves, in order: read the fold aloud, trace every number to the bank, delete the two weakest adjectives, and replace one abstraction with a specific image. Ten minutes per page, and it is where voice survives, the model produces competent structure, and the human produces the sentence a customer remembers. Editors who skip the pass to save time are buying page count with brand consistency, and the trade never nets positive.

Three verdicts: confirmed, contradicted, unsourced

The fact audit scores every claim against its proof-bank row with three verdicts. Confirmed means the evidence is in the row; contradicted means the row says otherwise, and both versions get shown to whoever owns the claim; unsourced means no evidence exists, and the claim is flagged for the bank or deleted. Hunt the confident-adjacent language first, tested in, proven to, featured in, because models repeat specific hallucination patterns: invented counts, overstated specs, borrowed press. Your first ten audits produce the pattern list that becomes the reviewer's checklist.

The QA Gate and the Two-Week Read

The QA gate is a four-item checklist that takes two minutes per page: banned-list search, golden-sample voice check, reading-level hold, and no surviving placeholders. Then the page ships, and the system's honesty begins, every shipped tier gets a two-week read against the gauges, because copy earns its keep on the ladder or it does not earn it at all.

The gauges are PDP-to-cart on the rewritten pages, store-wide conversion as the slow read, and the support-ticket count as the quiet bonus, pages that answer the real objections cut the tickets that ask them. A tier that underperforms gets the pipeline diagnosis rather than despair: usually a thin proof row or a failed fold argument, both fixable in the documents. That is the compounding logic, every failure makes the machine smarter, as long as the documents, not the model, are the source of truth.

The read also needs honest ranges: compute your own PDP-to-cart baseline from exports before the first tier ships, then judge every tier against its own baseline rather than a generic average, as the 2026 benchmark table recommends for every funnel number. A two-point lift on your own traffic beats a five-point claim on somebody else's.

The Evidence: What Product Content Is Worth

The investment case for disciplined product content is unusually well documented, because the failure case is too. Shoppers abandon purchases over incomplete or incorrect product information at near-universal rates, they consult reviews as a trust step on the way to checkout, and pages that display real review evidence convert dramatically better than pages that assert. The figures below are the ones to quote when the pipeline asks for an afternoon of setup time.

• The review-lift figure is why proof-bank rows carry quotes keyed to objections, not just specs

The Production Math: Speed, Cost, and Where Humans Still Own It

The pipeline's economics are why stores adopt it and never go back: the model drafts at machine speed, the documents govern quality, and the human spends ten minutes per page on judgment instead of ninety on invention. Independent research on AI-assisted knowledge work puts measurable numbers on the speed side of that trade, and, just as importantly, on its boundary. The figures below define both the return and the limit.

• Product copy sits comfortably inside the frontier, structured, evidence-fed, verifiable, which is why the gains hold
• The 19-point error rate outside the frontier is the fact audit's budget: it exists because confidence and correctness diverge

Common Mistakes That Turn Catalogs Into Mush

Every mistake above is a governance failure, which is the consistent finding of catalogs that worked and catalogs that did not. The fix is procedural in every case: documents first, placeholders instead of dice, verdicts instead of vibes, and a write-back habit that turns each week's catches into next week's rules. The model is the least of it, the tools article makes the same point about the whole stack.

• Drafting before the documents, no voice guide, no proof bank, so the model defaults to category clichés at machine speed
• Voice by vibes, rules nobody can check, so every editor rewrites every page and scale dies at editor number two
• The regeneration loop, re-rolling drafts until one feels right, which trades voice consistency for dice luck
• The unaudited ship, confident inventions reaching the page because the fact audit felt optional
• Uniform tiering, drafting the long tail before the hero pages are measured, so the pipeline's first impression lands where nobody is looking
• Set-and-forget prompts, the master prompt never updated from QA catches, so the machine never gets smarter

Rollout: Tiering the Catalog Across 30/60/90 Days

The rollout is traffic-weighted and stage-gated, and it resists the temptation to boil the catalog. The first month proves the machine on pages that matter; the second scales it; the third makes it boring, which is the goal.

• Days 1-30: voice guide, golden samples and proof bank for the hero tier; draft, audit and ship the top pages; measure PDP-to-cart against the baseline
• Days 31-60: fill the bank for the mid tier; draft and ship at pipeline speed; write QA catches back into the guide and the audit checklist
• Days 61-90: long-tail drafting with variant architecture; the support-ticket gauge joins the read; the master prompt and banned list get their quarterly sharpen

Ship the System, Not Just the Pages

If you want the full pipeline with worked examples, the voice guide template, the proof-bank columns, the master prompt and the audit protocol, it is Day 7 of the E-Commerce Growth Engine program, and the free Day 1 builds the founding artifacts on your own store first. For the writing-side craft of keeping machine drafts human, the humanizing system pairs with this pipeline at the editing stage, and the free marketer prompt library covers the non-product copy the catalog still needs. And for teams rolling the pipeline out across brands, the e-commerce case-study books and ready-made training kits in the ZLearning store include voice-guide worksheets and audit checklists you can run in a single workshop. AI writes the drafts; the system writes the brand.

Frequently asked questions