Is AI copy better than yours, or just longer?
Usually longer. That is the honest answer for most shop owners, and it is not a criticism of the models. A language model asked for a product description will produce four paragraphs where you would have written one, because length is what the request pattern rewards, and because it has nothing specific to say so it says the general thing at greater volume.
Which leaves you with a question the tool reviews never answer: your own copy is flat, you can feel that, and the AI version is smoother, but is it actually better? The way to settle it is a rubric you decide before you look at the output, so you are judging against a standard instead of against your own tired first draft.
- Score AI copy on four things: factual specificity, one clear action, absence of the promotional tell, and whether a customer could distinguish it from the shop next door.
- The tell that makes a customer feel AI copy is not the vocabulary, it is the ratio of adjectives to facts. Fix the ratio and the tell disappears.
- Wikipedia's editors have documented the pattern in detail: promotional tone, participle padding, negative parallelisms, forced triplets, and words like vibrant, nestled and testament.
- The five kinds of shop copy need wildly different amounts of editing. A returns policy is nearly unusable from a model. An email subject line is nearly free.
- US Bureau of Labor Statistics data puts median pay for writers and authors at 72,270 dollars a year, or 34.75 an hour as of May 2024, which is the real number an assistant subscription is competing against.
- Google's spam policy treats production method as irrelevant. Scaled content abuse is about pages generated without adding value, whether a person or a model typed them.
- Editing AI copy is faster than writing from scratch only when you delete rather than rework. If you are rewriting sentences, you are slower than starting from a blank page.
The rubric, stated before the output
Four checks, in this order. The order matters because the first one usually disqualifies the draft and there is no point polishing something you are about to delete.
- Specificity. Count the concrete facts. A material, a measurement, a compatibility, a care instruction, a real use. If the paragraph contains no fact that could be wrong, it contains no information.
- One action. What is the reader supposed to do next, and does the copy make that single thing obvious? Copy that gestures at three possible actions produces none.
- The tell. Adjective-to-fact ratio. More than about one adjective per fact and a customer starts to feel sold at rather than informed.
- Substitutability. Swap your brand name for a competitor's. If the copy still reads as true, you have written the category, not the product.
Check four is the one that catches most AI drafts, and it catches plenty of human drafts too. A model writing about your candle has access to the general facts of candles. Substitutability is the direct measurement of that gap.
How much editing does each kind of copy need?
Very different amounts, and this is the practical finding. Treating all copy as one job is why people conclude either that AI copywriting tools are magic or that they are useless, depending on which task they happened to try first. The same split by content type decides how much review a translated catalogue needs before it goes live.
| Copy a shop needs | How much editing before it ships | Why | Verdict |
|---|---|---|---|
| Email subject lines | Almost none, but generate twenty and pick | Short, low stakes, and the model is genuinely good at variation. You are choosing, not writing | Use it every time |
| Category page copy | Light. Add your own range facts and cut a third of the length | Category copy is legitimately general, which is the one place the model's weakness is not a weakness | Use it, then trim |
| Product descriptions | Moderate. Every specific must be supplied by you and checked back | The model will invent plausible attributes if your brief was thin, and plausible is worse than blank | Use it as a structure, not a source |
| Paid ad copy | Heavy, and test rather than judge | Ads live or die on the angle, and the angle depends on things about your buyer no model can see | Draft with it, decide without it |
| Returns and shipping policy | Near total. Treat the output as a checklist of clauses to write yourself | Legal effect, jurisdiction and your actual operational practice. A confident wrong clause is a liability | Do not publish the draft |
| Replies to individual customers | Light for the routine, total for the exception | The routine ones are patterns it has seen. The exception is exactly what it flattens | Use it for the two hundredth answer, never the first |
Subject lines are the one row where the verdict is unqualified, and the wider case for that sits in our piece on which stage of an email programme AI genuinely improves. Read the Why column and one rule falls out. The editing cost scales with how much of the copy's value comes from something only you know. Subject lines carry almost none of that. A returns policy is nothing but that.
What exactly is the tell a customer notices?
Not the vocabulary, though the vocabulary is the symptom everyone points at. It is the adjective-to-fact ratio, and it registers as a feeling of being sold at by someone who has not seen the product.
The most thorough catalogue of the surface symptoms was not written by a marketing agency but by Wikipedia's editors, who had to learn to spot machine-written articles at volume. Their documented signs of AI writing name the specific habits: promotional adjectives like vibrant and nestled, participle phrases that add emphasis without adding a fact, negative parallelisms of the not just X but Y shape, forced triplets, and the substitution of serves as or boasts where the sentence wanted is or has. The page is written for encyclopedia articles, and it transfers almost unchanged to product pages, because both fail in the same way: they use the shape of authority to cover an absence of specifics.
Here is the same product written twice. Both are fictional, both describe a hypothetical ceramic mug, and the difference is entirely the ratio.
The second version is shorter, contains six checkable facts, and could not be reused by another shop. The first version could be pasted onto any mug on the internet, which is the exact definition of failing the substitutability check. Notice too that the second one is not better written in any literary sense. It is just informed.
This is why the fix is almost never a better prompt. The first draft was not short of instructions, it was short of facts, which is the same reason a generated social post fails on its angle rather than its grammar. Give the model the six facts and it will produce the second version happily. The bottleneck is that somebody has to know the mug is 9cm tall.
What does it cost against paying a writer?
The comparison is usually run dishonestly, in both directions, so it is worth putting real numbers on both sides.
On the human side, the US Bureau of Labor Statistics puts the median pay for writers and authors at 72,270 dollars a year, or 34.75 dollars an hour, as of May 2024, with employment projected to grow 4 percent from 2024 to 2034. Freelance rates for commercial copy sit above that hourly figure rather than below it, since a freelancer carries their own overheads and idle time.
On the tool side, a general assistant subscription runs at roughly the price of an hour of that writer's time per month. Anthropic's published plans list Pro at 17 dollars a month billed annually or 20 monthly. So the arithmetic is not close, and anyone comparing on price alone has already finished the argument.
Which is why price is the wrong axis. The real comparison is what each buys you. A writer who spends thirty minutes on your product line learns the six facts about the mug and then writes forty descriptions that all pass the substitutability check. A subscription writes forty descriptions that pass it only if you supply the facts, which takes you roughly the same thirty minutes. You have not eliminated the work, you have moved it from writing to briefing and kept the money. For most small shops that is a good trade, and it is a different trade from the one the marketing describes.
Does publishing AI copy put your rankings at risk?
Not because a model wrote it. Google's spam policies page defines scaled content abuse as generating many pages primarily to manipulate rankings rather than help users, and it lists using generative tools to make many pages without adding value alongside scraping and stitching. The policy states plainly that the production method is not the deciding factor.
Its companion page on generative AI content makes the same point from the other side, and adds one detail merchants miss: AI-generated titles, meta descriptions and alt text are held to the same standard as anything hand-written. So the risk is not authorship, it is volume without value.
Search is also not the surface where authorship gets noticed any more, which is the change covered in our report on what sharper AI content detection means for a shop's copy. Applied to a shop, the line is easy to locate. Forty descriptions drafted from your own six facts per product and corrected by you is normal work. Four hundred pages spun from one template to catch long-tail searches is the thing the policy exists to stop. If you are unsure which side you are on, the substitutability check answers it: if the page would be equally true of a competitor's product, it added nothing.
Which AI copywriting tool should you use?
The general assistant you already pay for, in almost every case. This is an unpopular answer because the category contains dozens of products built specifically for marketing copy, and several of them are good. It is still the right answer for a shop, for two reasons.
The first is the briefing problem. A dedicated copy tool's advantage is templates: a product description form, an ad headline generator, a subject line builder. A generated shop arrives with those fields already in place, which is the difference between writing the copy and generating the store that holds it. Templates help when you do not know what good output looks like. They do not help when the limiting factor is that nobody has typed the mug's dimensions anywhere, and that is the limiting factor for most shops. A template turns a thin brief into a well-structured thin brief.
The second is the tool count. A dedicated copy tool competes for one of your three slots against an assistant that also reads your reviews, drafts your supplier emails and summarises your tickets. On frequency of use, the general tool wins that comparison most weeks, and frequency is what turns a subscription into a habit.
The case for the dedicated tool is real in one situation: if you publish copy at volume across many channels and need consistency enforced rather than remembered, a purpose-built tool with saved brand rules earns its slot, provided the consistency stops short of identical text everywhere, since LinkedIn and Snapchat now cut the reach of machine written posts that show no first hand input. A shop with forty products and one channel is not that situation.
How to build the six facts list
Sit with one product and answer six questions in writing. Not to a model, to a text file. Dimensions or capacity in the units your customer thinks in. What it is made of, precisely enough to be wrong. One thing about how it was made or sourced that you would mention if a customer asked in person. The single most common question you get about it. What it does not do, or who it is not for. Where it comes from.
That last one, what it does not do, is the field everyone skips and the one that does the most work. Copy that names a limitation is instantly credible, because nothing else in a product page is. It also reduces returns, which is a margin effect rather than a marketing effect, and it is the kind of sentence a model will never generate for you because it has no idea what your product fails at.
Six fields, forty products, roughly two hours of unglamorous typing. Everything downstream gets better and stays better, including anything you generate a year from now with a model that does not exist yet. It is the highest-leverage two hours in the whole exercise and no tool sells it to you, because there is nothing to sell.
When is editing slower than writing?
When you find yourself reworking sentences rather than deleting them. This is the single most useful signal for whether a tool is helping you on a given task, and it takes ten seconds to read.
Deleting is fast. Cutting three padded paragraphs down to one informative one takes less time than writing the informative one from nothing, because the structure is already there and you are only removing. Reworking is slow. Once you are rewriting clause by clause to fix the meaning, you are doing the writing anyway, with the added cost of first reading someone else's version and the added risk of keeping a phrase because it was there rather than because it was right.
So the operational rule is: if the second pass is deletions, keep using the tool for that task. If the second pass is rewrites, that task belongs on the list of things you write yourself, and no amount of prompt tuning changes it. This is the same rewrite rate test we apply across the whole category in the piece on which AI tools for a business earn their subscription, and copy is where it bites hardest because the output always looks finished. Cold outreach is the extreme case of that, which is why the funnel arithmetic behind an AI sales agent matters more than how polished its drafts read.
A workflow that holds up
Write the six facts per product once, in a plain list, before you open any tool. That list is the asset. It is what makes every description, every ad and every email downstream of it specific, and it is the thing no model can generate for you.
Then draft from the list, edit by deletion, and run the substitutability check before publishing. If your product pages, category copy and policies all live in the same system as your catalogue, that briefing list stops being a separate document you maintain and starts being the product data itself, which is the arrangement we build toward with an AI generated storefront that owns its own product data. What that costs to run is on the credit pricing page, and it is worth comparing against an hour of the median writer's time rather than against a free trial.
None of this makes AI copywriting tools a replacement for knowing your product. It makes them a fast way to turn what you know into pages, which is a smaller claim and a true one. The shops that get the most out of them are not the ones with the best prompts. They are the ones who wrote down the six facts.