AI Prompts for Ecommerce: Product Copy That Actually Converts

Why most AI product descriptions make a store worse

The standard ecommerce AI pitch is volume: point a tool at your catalogue, get five hundred product descriptions in an afternoon. Stores that do this and then look at their analytics three months later usually find nothing improved, and sometimes find things got worse.

The reason is not that the copy is badly written. It is that it is identical. Every store using a generic prompt on a similar product gets copy built from the same handful of moves — "elevate your everyday", "crafted with precision", "whether you are heading to the office or out for the evening" — because that is the average of the training data, and the average is what a vague prompt requests.

For search this is close to fatal. A product page that says nothing a hundred other pages do not say has no reason to rank, and if your descriptions came from the manufacturer's feed to begin with, you now have thin AI-rewritten duplicate content instead of thin duplicate content. For conversion it is worse: a shopper comparing three tabs cannot tell your product apart from the other two, and defaults to price.

None of that is an argument against using AI for ecommerce copy. It is an argument against the prompts nearly everybody uses. What follows assumes you would rather have forty differentiated pages than five hundred identical ones.

The rule that changes everything: feed it facts it cannot guess

A model writing about a product knows the category and nothing about your item. Ask it to describe a "ceramic pour-over coffee dripper" and it will write about ritual, morning light and rich aroma, because that is what the category writes about.

Give it the flow rate, the rib geometry, the fact that the cone angle is 60 degrees and that this matters for extraction time, the two most common complaints in your reviews, and the fact that the handle was redesigned after the first production run because customers burned their knuckles — and it writes something no competitor page contains.

This is the whole game. Every prompt below is a mechanism for getting real facts into the model and keeping generic vocabulary out. If you have no facts, no prompt will save the page — go and get them from the supplier spec sheet, the reviews, and the returns log.

• Specifications: exact dimensions, weight, materials, capacity, compatibility, care requirements, warranty.
• Reviews: the three things buyers praise most and the two they complain about most, quoted.
• Returns data: the actual reason people send this back. This is the objection your page has to answer.
• Competitive delta: what the two products a shopper is comparing this to do differently.
• Provenance: where it is made, by whom, what changed between versions and why.

Prompt 1 — The product description that is not interchangeable

The core prompt. Note the banned-words block — it is not decoration. Every ecommerce category has a set of words that appear on every page and therefore carry no information, and banning them forces the model to reach for the specifics you supplied instead.

• Prompt: "Write a product description for an online store. Product facts (use only these; do not add features, materials or claims not listed): [PASTE SPECS]. What buyers say they like: [PASTE]. What buyers complain about: [PASTE]. Who this is for: [SPECIFIC BUYER]. Structure: one opening sentence that states what it is and who it is for with no adjectives; one short paragraph on the single thing that makes it different from the alternatives; a bulleted specs block; one short paragraph that directly addresses the most common complaint above rather than hiding it. Banned words: elevate, curated, crafted, seamless, effortless, game-changer, perfect for, whether you are, look no further, unleash, indulge, luxurious. Under 180 words. Do not invent a benefit that is not derivable from the facts."

Prompt 2 — Write the objection into the page

The highest-converting paragraph on most product pages is the one that names the reason people do not buy. Stores avoid it out of instinct — why raise a doubt? — but the doubt is already in the shopper's head, and the page that addresses it is the page that gets the sale.

Your returns log tells you what the objection is. So does your support inbox and the two-star reviews. Feed it in and ask for an honest answer, not a deflection.

• Prompt: "The most common reason people hesitate before buying this product is: [OBJECTION]. The most common reason for returns is: [RETURN REASON]. Write two short paragraphs for the product page that address these directly. Rules: acknowledge the concern in plain terms rather than minimising it; give the actual fact that resolves it or, if it is a genuine limitation, say so and say who the product is not right for; do not use the words 'don't worry', 'rest assured' or 'peace of mind'. Facts available: [PASTE]"

Prompt 3 — Turn reviews into copy without fabricating them

Your reviews contain the language your buyers actually use, which is almost never the language your marketing uses. Mining them is one of the few genuine shortcuts in ecommerce copywriting.

The hard boundary: never generate reviews, never invent a quote, never attribute a sentence to a customer who did not write it. That is fabricated content, it is illegal in most markets to present it as genuine, and platforms remove stores for it. The prompt below extracts patterns and phrasing from real reviews for use in your own copy — it does not manufacture social proof.

• Prompt: "Below are real customer reviews for this product. Do not write any new reviews and do not invent quotes. Analyse them and output: THE WORDS BUYERS USE (the specific nouns and phrases they reach for that our marketing copy does not use); TOP THREE THINGS PRAISED, each with the count and one verbatim quote; TOP TWO COMPLAINTS, same; THE USE CASE WE ARE NOT ADVERTISING (anything buyers mention using it for that is not on our page); THE COMPARISON THEY MAKE (what they say they bought this instead of). Reviews: [PASTE]"

Prompt 4 — Collection and category pages

Category pages are where most stores leave the largest amount of search traffic on the floor. They typically carry a two-line introduction and a grid, which gives a search engine almost nothing to work with for a query like "cast iron pans for induction hobs".

The useful category page answers the buying question the category implies: how to choose. That content is genuinely helpful, it is what people search for, and it is impossible to duplicate across stores because it depends on the range you actually stock.

• Prompt: "Write the introductory content for an ecommerce category page for [CATEGORY]. We stock: [PASTE PRODUCT LIST WITH KEY DIFFERENCES]. Structure: one sentence stating what the category covers; a section called 'How to choose' with three to five decision criteria specific to this category, each explained in two sentences and each tied to which of our products fits it; one short section on the mistake first-time buyers of this category make. Do not write a general essay about the category. Every criterion must map to a real difference between the products listed. Under 350 words. Banned words: wide range, something for everyone, high quality, best-in-class."

Prompt 5 — Product titles that survive both search and the grid

Product titles do three jobs at once: they carry the search terms, they have to be scannable in a grid at small size, and on marketplaces they have to fit a character limit. Generating them one at a time by hand is slow; generating them without a format rule produces inconsistency across a catalogue, which looks amateurish at grid level.

• Prompt: "Write five title options for this product. Format: [BRAND] [PRODUCT TYPE] — [KEY DIFFERENTIATOR] [KEY SPEC]. Rules: the product type must be the words a buyer would search for, not our internal name; include the single spec buyers filter on in this category ([SPEC TYPE]); maximum [N] characters; no marketing adjectives; no ALL CAPS; no repeated words. Product facts: [PASTE]. After the five options, say which one you would choose for search and which for the grid, and why."

Prompt 6 — Meta titles and descriptions at catalogue scale

This is where AI genuinely does scale, because the job is mechanical and the constraint is a character count. The trap is the same as everywhere else: without differentiating facts you get four hundred meta descriptions that all say "Shop our range of X with free delivery".

• Prompt: "Write a meta title and meta description for this product page. Meta title: maximum 60 characters, must contain [PRIMARY SEARCH TERM], no brand name unless it is what people search. Meta description: maximum 155 characters, must contain one specific fact from the specs below that a competitor page would not contain, and must state the one thing that would make someone click this result over the others. No 'shop now', no 'free shipping' unless we actually offer it, no ellipses. Product: [PASTE]"

Prompt 7 — The abandoned cart sequence

Cart abandonment email is the highest-revenue-per-send email most stores run, and the default version — "you left something behind!" followed by a discount — trains customers to abandon carts deliberately.

A better sequence separates the reasons for abandonment and addresses them in order, holding any discount until last if it appears at all.

• Prompt: "Write a three-email abandoned cart sequence for [PRODUCT]. Email 1 (1 hour): assume they were interrupted. No discount, no urgency. One line, a clear link back, nothing else. Email 2 (24 hours): assume a specific hesitation — the most common one for this product is [OBJECTION]. Address it with facts: [PASTE FACTS]. Still no discount. Email 3 (72 hours): assume they are comparing to [COMPETITOR TYPE]. Give the honest comparison including where we are not the right choice. [INCLUDE / DO NOT INCLUDE] a discount. Rules for all three: subject lines under 45 characters, no emoji, no false urgency, no 'don't miss out'. Under 90 words each."

Prompt 8 — Ad copy variants that test something

Most AI-generated ad variants are the same message in different words, which produces a test with no information in it. To learn anything, each variant needs a different underlying claim, so that whichever wins tells you something about your buyers.

• Prompt: "Write four ad variants for [PRODUCT] on [PLATFORM]. Each must lead with a genuinely different angle, not different wording of the same one: (1) the specific problem it solves; (2) the comparison against the obvious alternative; (3) the specific buyer it is designed for; (4) the single most surprising fact about it. Facts available: [PASTE]. Constraints: [CHARACTER LIMITS]. No superlatives we cannot support, no invented statistics, no claims about results. After the four, state in one line what we learn about our buyers depending on which one wins."

Prompt 9 — Product FAQs from your actual support inbox

Product page FAQs do three jobs: they reduce pre-sale support tickets, they answer the questions that block purchase, and they are eligible for FAQ rich results in search. Inventing them wastes all three — the questions have to be the ones people actually ask.

• Prompt: "Below are real pre-sale questions we received about this product. Group them into the distinct underlying questions — the same question asked five ways is one entry — and write an FAQ for the product page. Rules: phrase each question the way a customer asked it, not the way we would describe the feature; answer in under 40 words; every answer must be factual and derived from the specs below; where the honest answer is 'no', say no and then say what we do offer. Order by how often the question was asked. Questions: [PASTE]. Specs: [PASTE]"

Prompt 10 — Size, fit and compatibility copy that reduces returns

For apparel and for anything with compatibility constraints, the single most profitable copy on the page is the paragraph that stops the wrong person buying. A prevented return is worth more than a marginal sale — you keep the shipping, the restocking cost and the customer's goodwill.

• Prompt: "Write the sizing and fit section for this product. Data: measurements [PASTE]; what reviewers say about fit [PASTE]; our actual return reasons and their frequency [PASTE]. Rules: state plainly if it runs small or large and by how much, based on the review data; give the specific instruction that would prevent the most common return reason; name the body type, use case or setup this product does not suit. Do not write 'true to size' unless the review data supports it. Under 120 words."

Prompt 11 — Bundles and cross-sells that make sense

Cross-sell copy generated without order data produces arbitrary pairings that shoppers ignore. With order data, the model can articulate why two products belong together, which is what actually moves attach rate.

• Prompt: "Here are the products most frequently bought together with [PRODUCT], with frequencies: [PASTE]. For the top three, write a one-sentence cross-sell line that states the concrete reason they go together — a functional dependency, a consumable that runs out, a gap the first product leaves. Do not write 'customers also loved' or 'complete the look'. If a pairing has no plausible functional reason and is likely just category co-occurrence, say so and exclude it rather than inventing a justification."

Prompt 12 — Rewriting a manufacturer feed without duplicating it

Resellers face a specific version of the duplicate content problem: every store selling this item has the same supplier copy. Rewriting it with a generic prompt produces a paraphrase, which is still substantially duplicate and adds nothing.

The way out is to add something the feed does not contain. Your own testing, your own photography notes, your customers' reviews, the questions your buyers ask. The prompt should force that addition rather than allow a paraphrase.

• Prompt: "Here is the manufacturer description for a product we resell: [PASTE]. Here is information the manufacturer does not have: our customer reviews [PASTE], our pre-sale questions [PASTE], and our own notes from handling the product [PASTE]. Write a new product description in which at least half the substantive content comes from the second set of information, not the first. Keep only the specifications from the manufacturer text and treat everything else in it as unusable. Do not paraphrase the manufacturer's marketing sentences — cut them. Under 200 words."

Prompt 13 — The buying guide that earns the search traffic

Product pages capture demand that already exists. Buying guides create it, and they rank for the "how to choose" and "best X for Y" queries that sit above the purchase in the funnel — which is also where AI assistants increasingly pull their answers from.

The guide has to be genuinely useful, which means it has to include the case where the answer is not your product.

• Prompt: "Write a buying guide for [CATEGORY] aimed at someone buying their first one. Structure: what the category is for and who does not need one; the three or four specifications that actually change the experience, each explained in terms of what the buyer will notice rather than the number; the specs marketing emphasises that do not matter much, and why; three common use cases with the right choice for each; the mistake that causes most returns in this category. Our range: [PASTE]. Rules: recommend our products only where they genuinely fit the use case, and name the case where a buyer should choose something we do not sell. No conclusion paragraph. Around 1,200 words."

Prompt 14 — Seasonal and campaign copy without the clichés

Seasonal copy is where generic vocabulary is worst, because the category has a fixed set of phrases and every store reaches for the same ones in the same week. Differentiation costs almost nothing here: a concrete, product-specific angle stands out sharply against a page of identikit campaign copy.

• Prompt: "Write campaign copy for [OCCASION] for [PRODUCT]. Requirements: the connection to the occasion must be specific to this product's actual use, not a generic gifting angle; state who on the recipient list this is right for and who it is not; include one concrete detail from the specs. Banned: 'the perfect gift', 'for the [X] lover in your life', 'treat yourself', 'spoil someone special', '\u2019tis the season', any countdown urgency. Product facts: [PASTE]. Produce a subject line, a 60-word email body and a 25-word banner line."

Prompt 15 — The pre-publish check

Ecommerce copy has legal exposure that most other marketing copy does not: product claims, health and performance assertions, comparative statements and price claims are all regulated in most markets. A check prompt costs seconds and catches the sentence that a generative model added because it sounded right.

• Prompt: "Check this product copy before publishing. List only problems. Flag: any factual claim about materials, dimensions, performance, compatibility or origin not present in the specs below; any health, safety, medical or environmental claim; any comparative claim about a competitor; any implied guarantee of results; any statistic; any placeholder left unfilled. Then list any phrase that appears generic enough to appear on a competitor page unchanged. Specs: [PASTE]. Copy: [PASTE]"

How to work through a catalogue without wrecking it

The practical question is sequencing. Nobody has time to hand-feed facts for two thousand SKUs, and doing the top forty properly beats doing all two thousand badly by a wide margin.

Sort by revenue and by traffic, take the overlap, and work down. For the long tail, a specs-driven template with the banned-words rule is fine — it will not win any awards but it will not be actively harmful either, which is more than can be said for a generic prompt run at volume.

• Tier 1 — your top 20 to 50 products by revenue: full treatment. Facts, reviews, objection paragraph, FAQ, custom meta. This is where the return is.
• Tier 2 — products with search traffic but weak conversion: run the objection and reviews prompts only. The traffic is already arriving; the page is failing to convert it.
• Tier 3 — the long tail: a consistent template with real specs and the banned-words rule. Do not spend hours here.
• Never: bulk-generate the whole catalogue from titles alone. That is how a store ends up with two thousand pages that say nothing.
• Measure per tier. If tier 1 pages do not improve on conversion or search impressions within a few weeks, the facts you fed in were not differentiating enough.

What AI search means for ecommerce pages

A growing share of product research now happens inside AI assistants rather than a results page. Someone asks which pour-over dripper suits a coarse grind, and the assistant answers from pages it can parse — then may or may not cite where the answer came from.

This changes what a good product page looks like in one specific way: the facts have to be extractable. Specifications in a structured block get read; the same numbers dissolved into a paragraph of marketing prose often do not. Explicit answers to explicit questions get quoted; implication does not.

The stores that do well here are the ones that were already doing the unfashionable thing — publishing real specifications, honest limitations, and direct answers to the questions buyers ask. That content is what an assistant can use. Marketing prose about elevating your morning ritual is not usable by anything, human or machine.

• Keep specifications in a labelled list, not buried in prose.
• Answer the buying question explicitly on the page, in a sentence that stands alone if quoted.
• Use Product and FAQPage structured data so price, availability and answers are machine-readable.
• State limitations plainly — "not compatible with X" is one of the most quotable and most useful sentences you can publish.
• Keep the comparison honest. An assistant asked to compare will find the contradictions between your page and reality, and the store that reads as trustworthy is the one that gets recommended.

Building this into a repeatable system

The prompts above are worth roughly nothing as a one-off exercise and quite a lot as a saved workflow. The difference is whether the facts block is a habit or a heroic effort.

Set it up once: a spreadsheet column for each product's differentiating facts, review summary and top return reason, filled in as part of the product setup process rather than as a copywriting task later. Then the description prompt becomes a paste rather than an investigation, and the marginal cost of a good page drops to a couple of minutes.

Save the prompts with your banned-words list baked in — that list is store-specific and grows every time you catch a phrase that sounds like everyone else. Version them when a prompt produces something you had to heavily rewrite. Within a season you have a copy system that produces pages your competitors' generic prompts cannot match, using the same models they are using.

Related articles