- Marketplace listings now write themselves from a photo, a URL or a sparse spreadsheet, on every large platform, at no extra charge.
- Amazon reports that sellers accept its generated content with little to no edits roughly 90 percent of the time, across more than 900,000 sellers.
- eBay reports that 30 percent of its US app sellers tried the description feature and more than 95 percent of those kept what it produced.
- When almost nobody edits, the copy stops being a differentiator. What separates listings is the structured detail the generator could not know.
- Four fields are worth overriding every time, and three of them are not prose at all.
- A generator working from your own website inherits every error already on it, then publishes that error somewhere you do not check.
Upload a photograph. Wait a few seconds. A title appears, a category is chosen, item specifics are filled in, bullets are written, and a description sits underneath it all reading like a competent product page. On some platforms you can do this to hundreds of photos at once. None of it costs extra.
The useful question is not whether to use these tools. Almost everyone already does, and refusing on principle mostly means listing more slowly than your competitors. The question is which of the fields they fill you should still be checking, given that the platform's own numbers say hardly anyone checks any of them.
How much do sellers actually edit?
Very little, and both large marketplaces publish the figure rather than hiding it.
Amazon states that sellers accept its AI generated content with little to no edits approximately 90 percent of the time, alongside adoption figures of more than 900,000 sellers using its tools and over 400,000 using the generative listing capabilities specifically. Its inputs are broad: a brief description, a URL from the seller's own brand website, a product image, or a spreadsheet with a sparse set of details.
eBay's numbers point the same way. When it introduced AI written descriptions, 30 percent of its US app sellers tried the feature and more than 95 percent of those who tried it used the result, including with edits. Its bulk version takes hundreds of photos per batch and returns drafts with suggested categories, titles and item specifics, currently limited to sports trading cards in the United States with more categories promised.
Sit with the 90 percent figure for a moment, because it carries the whole argument. If nine listings in ten go out as generated, then the words on a listing page are converging across every seller on the platform. Whatever advantage careful copywriting used to buy has been arbitraged away by a free button.
So what still separates two listings?
The parts a generator cannot know, which turn out to be the parts buyers filter and search on.
A model looking at your photograph can tell it is a wool coat. It cannot tell you the pit to pit measurement, whether the lining is intact, that the third button was replaced, or that it fits small. It can write that a cable is durable. It cannot know which port revision it terminates in or which devices it has actually been tested with. The fix is the same one that makes AI product descriptions on your own site readable, which is to hand the model your measured attributes instead of a product name.
Those specifics do two things at once. They answer the question that stops a purchase, and they populate the structured fields that decide whether you appear when somebody filters. That second effect is the one sellers underrate. A buyer narrowing by size, compatibility or condition is not reading your bullets at all; the filter runs on item specifics, and a listing with a blank field is simply absent from the results.
Which four fields should you always override?
These, in order of how much damage the default does.
The product category. Inferred from an image, and the highest risk item on the page because everything downstream depends on it. A wrong category means the wrong required specifics, the wrong filters and sometimes the wrong fee. It is also the field sellers are least likely to question, because a plausible category looks correct at a glance. The mechanics of why this cascades are the same ones covered in our piece on how product categorisation decides which attributes become mandatory.
Measurements and compatibility. Never guess these and never let anything else guess them. A generated dimension is a fabricated fact with a returns cost attached, and in some categories it is a safety claim. If you sell anything worn, fitted or connected, this field is your entire returns rate.
Condition, for anything not new. Generators are trained on product data, and product data describes the model rather than the item. The flaw, the wear, the missing accessory: none of that is in the photograph in a way a model will reliably surface, and all of it is what a dispute will later turn on. Write it yourself, plainly, including the bad part.
The first line of the title. Not for beauty. Because generated titles converge on the same phrasing from the same visual cues, so a page of search results starts to read as one paragraph repeated. The differentiating token, usually a model number, a size or a specific use, is the thing worth putting early and the thing a generator tends to bury.
Where each field comes from and what it costs you
| Field | How it is produced | Failure you will see | Where the cost lands |
|---|---|---|---|
| Category | Inferred from image or text | Plausible but wrong branch | Missing from every filtered search |
| Item specifics | Guessed, or left blank | Blanks nobody notices | Invisible to buyers who narrow |
| Measurements | Invented or omitted | Confident wrong number | Returns and disputes |
| Condition notes | Not produced at all | Silence read as perfect | Not as described claims |
| Description prose | Generated, rarely edited | Reads like every rival | No differentiation, low cost |
| Attributes from your site | Parsed from your own pages | Your existing errors, copied | Wrong data in a second place |
What happens when it reads your own website?
It inherits everything, including the parts you had forgotten were wrong.
Amazon accepts a URL from your brand website and parses it into a listing. That is genuinely convenient and it has a property worth thinking about: the generator has no way to distinguish your current specification from the outdated paragraph you never removed, or the shipping promise you changed last year and left in a template. Whatever is on the page is treated as true.
So a marketplace listing built this way is a copy of your site's data quality, published on a surface you check less often. If your own product pages carry a discontinued variant, an old warranty period or a measurement in the wrong unit, that error now exists twice and diverges from there. The remedy is not to stop using the feature. It is to fix the source page first and generate afterwards, which is the reverse of the order most people work in.
Bulk listing generation deserves a stricter rule than single listings, because the review step scales badly. Reviewing one draft is a minute. Reviewing four hundred drafts is a day nobody has, so in practice they go out unreviewed. If you are generating in bulk, spot check a random sample of twenty and check the category on every single one, since that is the field where a systematic error repeats across the whole batch.
Does disclosure apply to a written description?
It depends on the platform, and the rules are moving fast enough that this is worth checking on your own marketplace rather than trusting a summary. Handmade and creative marketplaces have taken the strictest line, with disclosure requirements and attribution settings that distinguish an item you made from one you designed with AI assistance. General marketplaces have so far treated generated text as one of their seller tools rather than a disclosable fact.
The practical position for a seller is to separate two questions that get muddled. Whether AI helped you write the words about a product is a copywriting matter. Whether AI made the product itself, or the image representing it, is a provenance matter and is what the strict policies are actually aimed at. Those are different risks with different consequences, and a policy written for the second sometimes catches the first in its wording.
Where an image is generated rather than photographed, the ground shifts again. We went through what a generated image can honestly claim to show in our note on how AI content detection now reads shop pages, and the same caution applies here with the added weight of a marketplace policy behind it.
What the generator is optimising for
Not your margin, and this is worth stating because the incentive gap is easy to miss when the tool is free and helpful.
A marketplace wants complete listings, because complete listings are searchable, filterable and comparable, which makes the catalogue better for buyers and lifts platform sales overall. Amazon frames its Enhance My Listing feature exactly this way, offering recommendations for titles, attributes, descriptions and missing details. Filling gaps is genuinely the point. Whether the gap gets filled with a fact or with a plausible sentence matters much less to a platform than whether it gets filled at all.
Your interest is narrower and sharper. You want to be found by the buyers who will keep the item, and you want to repel the ones who will return it. Those two aims diverge in one specific place: any field where an ambiguous or optimistic answer widens your reach. A slightly generous size band, a compatibility list that includes the probable rather than the tested, a condition described as good rather than good with a scratch. Each of those brings more clicks and a worse return rate, and a generator with no view of your returns has no reason to be conservative.
The rule that follows is simple enough to apply without thinking. Where a generated field is a claim you would have to defend in a dispute, write it yourself. Where it is description, let it stand. That single test sorts most of a listing in about thirty seconds.
What to do with the listings you already published
Audit by symptom rather than by going through them one at a time, because most sellers have more listings than review capacity.
Start with returns. Pull your last three months of returns by reason, and look only at the ones logged as not as described or wrong item. Those map directly onto listing fields, and the products appearing twice are telling you exactly which listing is lying. Fix those first, since each one is currently costing you shipping in both directions plus a metric that affects your standing on the platform.
Then look at the products that get views but not sales. On most marketplaces that report is free. A listing with traffic and no conversion usually has an unanswered question rather than a price problem, and the unanswered question is almost always a specific the generated draft skipped.
Last, sample by category. Take ten listings from each category you sell in and check whether the required specifics are actually filled. Blank fields cluster by category rather than scattering randomly, because they follow the shape of what the generator could infer from your typical photograph. Finding one blank usually means finding forty.
Is the time saving real?
Yes, and it is the honest reason to use these tools. Listing is tedious, the tedium is why sellers under list, and a draft that arrives in seconds removes the blank page problem entirely. eBay's bulk tool exists precisely because photographing a stack of trading cards is fast and typing them up is not.
What the saving buys you is the interesting part. If the hours freed go back into listing more inventory, the tool has paid for itself several times over. If they go into nothing in particular while the listing quality quietly drops, you have traded a differentiator for an afternoon. The sellers who come out ahead are the ones who take the generated draft and spend the recovered time on the four fields above rather than on the description that was already fine.
The same logic applies to written copy generally, and we set out the editing question in more detail in a piece on how much editing AI written copy needs before it ships. On a marketplace the answer is unusual: less editing of the prose than you would expect, and far more attention to the fields that are not prose at all.
Photographs are the new input, and that changes the risk
The first generation of these tools worked from text you typed. The current one works from a picture, which is a different proposition entirely.
When you typed a few features, the generator was expanding your claims. You had already asserted the facts, and the tool was making them read better. When it works from a photograph, it is asserting facts on your behalf, derived from what a model believes it sees. Those are not the same kind of output and they do not carry the same liability.
The practical consequence shows up in categories where visually similar items differ in ways a camera cannot capture. Two cables that look identical and terminate differently. Two garments in the same cut and different fibre. A component whose revision is printed on a face you did not photograph. In each case the draft will be fluent, specific and wrong, and it will be wrong in the confident register that makes a buyer trust it.
There is a cheap defence and almost nobody uses it. Photograph the label. Whatever carries the model number, the fibre composition, the size, the batch, put it in the image set you feed the generator and keep it in the listing. It improves what the tool produces, it gives a buyer the evidence directly, and if a dispute ever arrives it is the single most useful thing in your file.
The strategic point underneath all of this
Every one of these tools makes it easier to list on somebody else's platform, using their categories, their fields, their generated words and their audience. That is a real benefit and it comes with a familiar cost. The listing is not yours, the customer relationship is not yours, and the moment the platform changes a fee, a category tree or a policy, your catalogue changes with it and you were not consulted. The sharper version of that cost is an automated suspension you have to appeal.
None of that argues for abandoning marketplaces, which are where the buyers are. It argues for not being only there. A channel you control is where you can write what you want, keep the customer record, and hold the canonical version of your product data that everything else copies from. That is the case for a storefront whose content and customer data stay yours, and the generated listing era strengthens it rather than weakening it: when the words on every marketplace page converge, the place where you can still sound like yourself becomes more valuable, not less.
Practically, this means keeping your own site as the master record and treating every marketplace as an export target. Fix the product there, generate from it, override the four fields, move on. It is a small discipline and it stops the two versions of your catalogue drifting apart, which is the failure that costs the most and announces itself the least.
One closing observation about where this is heading. Every marketplace building these tools is also building the buyer side: assistants that answer a shopping question by reading the same structured fields the generator just filled in. That closes a loop worth noticing. The fields you skipped because a blank looked harmless are the fields a recommendation engine reads to decide whether your item matches what somebody asked for. A listing with a generated description and empty specifics is optimised for a browsing buyer who no longer exists in the numbers you care about, and invisible to the one who does.