- AI for resale is strongest on the boring half of the work, which is turning a photo into a complete, consistent listing.
- Condition grading is a description problem, not a judgement problem. Write what is visible and let the buyer grade it.
- Authentication from photographs produces a signal, never a verdict, and treating it as a verdict is how a seller ends up making a claim they cannot support.
- Clothing, footwear and leather goods dominate counterfeit seizures, which is exactly the category most small resellers trade in.
- A liability rule worth adopting: never state that an item is authentic on the strength of a tool. State what you checked and what you know about its history.
- The Digital Product Passport for textiles is coming and will eventually make resale provenance far easier, with the delegated act planned for the fourth quarter of 2027.
Somebody buys a coat at a house clearance for eleven pounds. It has a label they half recognise, a small mark on the left cuff and a lining that looks newer than the shell. They want to list it tonight. The two questions standing between them and a sale are whether it is genuine and how to describe its condition without getting a return in a fortnight.
Resale is one of the few kinds of selling where every single item is a fresh judgement. There is no product page to reuse, no supplier spec, no second unit to learn from. That repetition of one off decisions is what makes it exhausting at volume, and it is also what makes it an obvious target for automation. Which parts genuinely automate is the question worth answering carefully, because getting it wrong here is a different order of problem from getting a product description wrong.
Is resale actually big enough to build habits around?
Yes, and the growth is in the direction of more sellers rather than fewer. ThredUp's fourteenth annual resale report, published on 2 April 2026, puts the global secondhand market on a path to 393 billion dollars by 2030, growing roughly twice as fast as apparel retail overall, with the United States market projected at 78.8 billion dollars and secondhand already accounting for around 10 percent of total apparel spend.
The same report makes a claim about technology that is worth reading closely rather than accepting: that AI is the engine helping scale resale by improving search and discovery, automating everything from pricing to verifying authenticity. The first half of that sentence is well supported. The second half, verifying authenticity, is doing a great deal of work in four words, and the rest of this piece is largely about why.
One other finding is operationally useful. Nearly half of secondhand shoppers now discover items through social feeds and creators rather than through search. That changes what a listing is for: it has to survive being seen as a single image in a feed, which puts more weight on the first photograph and less on the keyword stuffing that used to drive marketplace search.
What can a model genuinely do with a photo?
Quite a lot, provided you ask it to describe rather than to conclude.
Given four decent photographs it will reliably tell you the garment type, the apparent material, the colour in words a buyer would search, the visible construction details, the presence of a mark and roughly where it sits. It will read a care label. It will notice that a zip pull is a different metal from the studs. Every one of those is an observation, checkable by looking at the same photo, and none of them requires the model to know anything about the world beyond the image.
That is the useful frame: it is doing description and extraction, which is the same class of task as any other vision work, and it fails in the ordinary ways such systems fail. Our explainer on what the different computer vision tasks actually are is the background reading, and the relevant line from it is that detecting something and identifying it are different problems with different reliability.
Where it stops being reliable is any question whose answer is not in the pixels. Does this smell of damp. Is the seam about to fail. Was this worn twice or fifty times and carefully repaired. Those are the questions returns are made of, and no amount of photograph improves the model's access to them.
Condition grading, and why the shortcut backfires
The temptation is to ask for a grade: excellent, good, fair. It is the wrong question and it produces the wrong kind of error.
A grade is a promise. A description is a statement of fact. If you write that a coat is in excellent condition and the buyer thinks the cuff mark disqualifies that, you have a dispute about a word. If you write that there is a two centimetre grey mark on the left cuff, visible in photo four, that did not lift with cold water, there is nothing to dispute. The buyer grades it themselves and they do so before paying rather than after.
This is where a model earns its keep in resale, and it is unglamorous. Ask it to list every visible flaw with location and approximate size, from the photographs, in plain language. Then check the list against the item in your hands, add the things it cannot see, and that becomes the condition section. The consistency matters more than the eloquence: a seller whose descriptions always cover the same eight points in the same order gets fewer questions and fewer returns, because buyers learn to trust the format.
Add one line the model can never write: what you did to it. Washed on a cool cycle, steamed, not altered, stored flat. That sentence does more for buyer confidence than any adjective, and it is the part of the listing that is unambiguously yours.
Can it tell a fake?
It can raise a flag. It cannot give you a verdict, and the distinction is not pedantry, it is the whole of your legal exposure.
Counterfeiting concentrates precisely where small resellers operate. An analysis of the EUIPO and OECD findings records footwear at 33.7 percent of seizures, clothing at 17.3 percent, perfume and cosmetics at 9.6 percent and leather products at 8.7 percent, with more than half of goods seized at European borders coming from online trade and 91 percent of seizures involving small packages sent through postal services. Those figures come from a study of 2017 to 2019 data published in 2021, so treat the proportions as the shape of the problem rather than as this year's number.
The picture has not improved since. A later OECD and EUIPO study on dangerous fakes found online sales accounting for 60 percent of seizures of dangerous products destined for the European Union, with China and Hong Kong together the source of around 75 percent of those seized.
What this means for a one person resale business is that you will handle counterfeits, not as an unlikely accident but as a normal event, and your process has to assume it rather than hope otherwise.
| Signal | What a model reads from photos | How much it is worth | What you do with it |
|---|---|---|---|
| Font and logo detail | Shape, spacing, alignment | Moderate, and easily faked now | Flag for closer look, never conclude |
| Stitching pattern | Density, regularity, thread colour | Moderate on structured goods | Photograph it and let the buyer judge |
| Hardware and engraving | Finish, weight cues, wear pattern | Low from a photo alone | Handle it, describe what you felt |
| Serial or date code | The characters themselves | High as extraction, low as proof | Record it, state it, do not interpret it |
| Label and care tag | Wording, fibre content, country | Useful for inconsistency | Compare against the era of the item |
| Price and source | Nothing, this is your knowledge | Highest single signal | Trust it above all of the above |
The bottom row is the one experienced resellers already know and newcomers underrate. Where the item came from, what was paid for it and what else was in the same lot tells you more than any visual check. A house clearance with forty years of one household's clothes is a different provenance story from a pallet bought unseen, and no vision model has access to that difference.
How should the listing be worded?
Conservatively, and in a way that separates fact from belief. Three formulations do most of the work.
Say what you observed: the item carries a serial number reading as follows, photographed in image five. Say what you did: purchased from a private seller in a house clearance, not professionally authenticated. Say what you offer: full refund on return within thirty days if you believe it is not genuine, no questions.
What you avoid is the flat assertion of authenticity, because that assertion is a statement about the item that you cannot personally support, and a model's opinion does not become your knowledge by passing through your keyboard. It is also the statement that turns a refund into a misrepresentation claim if it turns out to be wrong.
The related risk is the platform rather than the buyer. Listings suspected of infringement get removed, and repeat removals get accounts suspended, often with an opaque notice and a slow appeal. That process has its own rules about what you are entitled to be told, which we set out in our piece on what a marketplace owes you when it suspends a seller account. The defensive posture is the same one that helps with buyers: a record of where each item came from, dated.
Photography, and the rule that catches resellers
Generated and heavily edited imagery is a specific hazard in resale that does not exist in new goods. If you sell a new mug, an idealised studio render is a representation of the product line. If you sell a used coat, the photograph is evidence about that specific object, and improving it changes the evidence.
Removing a background is generally fine. Removing a stain is misrepresentation. Colour correcting to daylight is fine. Shifting a colour to a more saleable shade is how you get a return and a bad review at once. The boundary is whether the edit changes a fact the buyer is relying on, and it is worth writing your own rule down because the tools make the wrong edit as easy as the right one. Marketplaces have their own positions on this too, which we covered in the rules platforms apply to AI generated product photos.
There is a practical upside to keeping images honest beyond avoiding disputes. Buyers of secondhand goods look for flaws. A listing that shows the mark clearly converts better than one that hides it, because the shopper who was going to find it anyway now trusts everything else you said.
Does AI pricing work better here than elsewhere?
Better than it does for new goods, and for a reason worth understanding. Pricing a new product is a strategy decision about your position in a market. Pricing a used one is closer to an estimation problem with a lot of comparable evidence, because thousands of similar items have sold recently at observable prices.
The caution is that comparables are only comparable if the condition matches, and condition is the variable the comparison tool understands least. A listed sold price does not tell you whether that jacket had a mark on the cuff. So a pricing suggestion drawn from sold listings is a reasonable ceiling for an excellent example and a poor guide for anything flawed, which is most of what passes through a small resale business.
The practical method is to take the suggestion as an anchor, then discount explicitly for each documented flaw and write the reasoning in your own notes. Over a few hundred items that private record becomes the most valuable asset in the business, because it tells you what your buyers actually pay for a described flaw rather than what a general model assumes.
Speed matters more than precision here in any case. Resale inventory has a carrying cost in space and attention rather than in cash, since the money was already spent, and an item sitting unsold for five months at the perfect price has cost more than one sold in a week at ten percent less.
What is coming that will change this
Provenance is about to get easier, slowly. The European Union's Digital Product Passport will attach structured information to products through a data carrier such as a QR code. The Commission's page on textiles and apparel states that the passport is expected to cover product identification, fibre composition, origin information, identification of relevant economic operators and data relevant to reuse, resale, disassembly, refurbishment and recycling, with the delegated act for textiles planned for the fourth quarter of 2027 and exact requirements to be defined then.
Resale is named in that list, which is the part worth noticing. A garment that carries a verifiable record of what it is and who made it removes most of the authentication problem for anything manufactured after the rules bite. It does nothing for the forty year old coat in the house clearance, which is to say it does nothing for a large share of what small resellers actually sell, and it will not take effect for years.
The obligation also lands on economic operators placing products on the market, with distributors required to ensure a passport is available for what they handle. How that applies to a reseller of an already sold item is exactly the sort of detail the delegated acts will settle, and it is worth watching rather than worrying about now.
Running resale on your own site
One structural note. Resale inventory is almost always single quantity, which breaks a lot of assumptions built for catalogues. Every item is its own product, sells once and then must disappear cleanly, and a listing that stays live after the sale produces the worst customer experience in the business.
Most sellers start on marketplaces for the traffic and move some of the volume to their own site as they build a following, particularly given how much discovery now comes through social feeds rather than marketplace search. If that is the direction, the thing to check is whether the platform handles one of one inventory naturally rather than as a workaround, and our comparison of what to look for when moving off a hosted platform covers the trade offs involved in that move.
The other piece worth automating early is the cross listing record: which item is live where, and what happens when one sells. That is dull integration work rather than intelligence, and it prevents the single most common resale failure, which is selling the same jacket twice.
In resale the model is a very fast pair of eyes and a very poor witness. Use it for the first and never call it as the second.
A working process
Photograph the item properly, including every flaw, in daylight. Let the model draft the description and the flaw list from those photographs. Add what your hands and nose know that the camera did not. Record where it came from and what you paid, in a private column nobody sees. Write the authenticity language as observation rather than assertion. Then list.
That sequence takes a couple of minutes per item once it is habitual, and the minutes it saves are real. What it protects is the part of a resale business that takes years to build and an afternoon to lose, which is a buyer's belief that your descriptions are accurate.