ThredUp Uses AI to Price Secondhand Clothing, Achieving a Gross Margin of 79.5%

How do you deal with your old clothes?

Of course, I recycle them on Zhuanzhuan. Zhuanzhuan is an official second-hand platform, more formal than ordinary second-hand ones...

Every year, about 17 million old clothes set out from strangers' wardrobes, pass through it, and enter another stranger's wardrobe.

Behind every new piece of clothing is a definite stock-keeping unit. The same shirt, the same color, the same size—thousands of them look exactly alike. Warehousing, pricing, and listing are not complicated for the system to handle.

But ThredUp's processing center receives millions of garments sent in each year, and every single one is a unique item. This plaid shirt has been washed twenty times and the collar is a bit curled; those jeans are labeled size 32 but actually run large at the waist; another sweater still feels new, yet its tag has fallen off.

So how does the platform estimate value? Clerks at consignment stores generally take in goods based on experience—they feel the fabric, glance at the brand, and make a mental estimate. If an item doesn't sell for a few months, they cut the price or return it to the consignor.

Many charity second-hand shops set uniform prices by category: a shirt for a few dollars, a pair of jeans for a dozen or so dollars. Good and bad items sit under the same price tag; bargain hunters get a good deal, but the shop has a hard time knowing how much a good item is really worth. This works at small scale, but once volume rises, manpower is hard pressed to keep up.

ThredUp is the one in this industry that has thoroughly industrialized the process. According to Modern Materials Handling's 2023 report on its distribution center, incoming clothes must first pass 12 quality checks; only those that pass are hung on hangers with QR codes. Staff enter the category, brand, color, size, fabric, and washing method, and pricing is left to proprietary machine-learning software. High-priced clothes are photographed on mannequins, low-priced ones are shot flat, and photos are checked before listing.

Later, AI took over more steps, able to automatically identify color, category, and size, and then factor condition and market demand into algorithmic pricing. According to a PYMNTS report this April, the platform consistently has more than 4 million items for sale, with tens of thousands added daily; a reinforcement-learning model watches real-time demand and automatically cuts prices on slow-selling clothes.

The company has also invested more than $400 million in supply-chain automation, covering product recognition, size measurement, and photography.

The result of these investments is that ThredUp can sell about 17 million garments a year. In the second quarter of 2025, its gross margin reached 79.5%, while traditional apparel retailers rarely exceed 30%.

On the consumer side, in August 2024 ThredUp launched three AI features: upgraded search, image search, and Style Chat.

Style Chat is like a conversational stylist. Enter "what to wear to an autumn wedding," and it gives you a complete outfit; enter further requirements, and it keeps adjusting.

Image search lets users upload a photo or share an Instagram post to find similar styles in inventory. Company-disclosed data shows that consumers who use image search are 85% more likely to eventually place an order than ordinary browsers.

People who use this feature come with specific needs, and AI helps them lock onto the target. The clearer the intent, the smoother the conversion. Because the goods on second-hand platforms are simply too scattered, being able to help people find that "right one" among millions of old clothes is where the barrier lies.

The latest Adobe data shows that, compared with ordinary shoppers from other traffic sources, shoppers who reach retailer websites through AI channels such as ChatGPT are more likely to browse longer, add items to their carts, and, most importantly, complete a conversion.

In July this year, shopping users from AI channels had a conversion rate 60% higher than those from other channels, generated 53% more revenue per visit, had a 34% lower bounce rate, spent 59% more time on site, and viewed 22% more pages.

This conversion-rate advantage has held for 11 consecutive months. A year ago, the situation was the opposite: at that time, customers who did not come from AI recommendations spent more per visit. This means AI assistants are getting better at showing relevant products to customers with higher purchase intent.

Back to the second-hand clothing scenario: the inventory on second-hand platforms is enormous—ThredUp alone has more than 4 million items on sale, and every one is different. There is no hit item that people can blindly follow. Browsing such shelves easily tires people out.

There is a famous jam experiment in psychology. In 2000, Iyengar and Lepper set up a stall in a supermarket. When 24 kinds of jam were displayed, 60% of passersby stopped to taste, but only about 3% ended up buying. When 6 kinds were displayed, only 40% stopped, but about 30% bought.

Too many options make decisions harder—that is how the idea spread.

Later, others raised different views. A 2010 meta-analysis summarizing dozens of experiments found that choice overload is more likely to occur when options are complex and people have no clear preference of their own.

According to a ThredUp survey, 48% of second-hand shoppers have used AI tools while shopping. People whose choices have been narrowed arrive at the product page having already made half the decision; all that remains is to confirm size and price, so it is no surprise that conversion rates are high.

Of course, the users brought in by AI are not a monolithic bloc. They also come with high demands born of high expectations. If recommendations are inaccurate, that instead reinforces disappointment—after all, "AI recommended it to me" implies an endorsement, and once accuracy fails, the loss of trust is even greater than with random browsing.

This explains why every platform is making its recommendation models more and more granular. Knowing the brands you like is only the beginning; it also needs to know that your last return was because the size ran large, that you prefer matte over glossy, and that you usually scroll on your phone on Friday nights.

ThredUp disclosed that in all of 2024, the number of new customers brought in by AI grew 32% year over year. The second-hand clothing category inherently has a barrier to entry—uncertainty about texture and condition makes many people hesitate—and AI's involvement is dissolving that hesitation.

From helping you choose to buying for you

Recognition, pricing, and recommendation—"helping you choose" has matured, and there is already a proven example like ThredUp. The next step is "buying for you," letting AI take your authorization and go place orders and make payments itself.

In September 2025, OpenAI and Stripe teamed up to launch instant checkout within ChatGPT, with initial partners including Etsy and some Shopify merchants. When users spot something they like in the conversation, they can complete payment directly within the interface without jumping elsewhere. This was the first time a mainstream AI conversation platform tried to bridge the last gap between "discovery" and "purchase."

But by March 2026, OpenAI adjusted direction, shifting its focus back to "product discovery" and allowing merchants to take orders using their own existing checkout processes rather than uniformly going through the closed-loop payment within ChatGPT.

Media citing analysts' views pointed out that Instant Checkout did not gain many users, for reasons including the hassle of merchant integration, the difficulty of keeping product data accurate, and the challenges of handling multi-item shopping carts and loyalty points. OpenAI itself believed the initial version "did not provide the flexibility we expected."

A Morgan Stanley report in December 2025 estimated that by 2030, AI shopping assistants could account for 10% to 20% of U.S. e-commerce, corresponding to $190 billion to $385 billion in sales. The report also noted that about 23% of people in the U.S. had bought something through AI in the past month, with groceries and everyday consumer goods being the first categories to scale up.

What users typically give an Agent is a mandate with a scope, a budget, and a time limit, rather than nodding approval for every transaction. A spending cap can block expensive mistaken purchases, but not cheap ones. Currently there are no public rules for handling "the Agent exceeded your intent," and merchants often can only absorb the losses themselves.

Before buying clothes, people look at images, size charts, and return policies; an Agent mostly just reads data, which pushes return rates up. ecommerceguide points out that clothing, shoes, and furniture are the highest-risk categories.

For second-hand clothes, each old garment is one of a kind—once sold, it's gone—and returns require re-inspection, re-photographing, and re-pricing. The per-item processing costs that ThredUp has driven down through scale may well be pushed up here.

Agents look at price and specifications, not layout and styling recommendations, so the displays merchants put effort into become ineffective, and second-hand goods whose condition descriptions are not detailed enough get ranked lower in the Agent's eyes.

ThredUp long ago turned non-standard goods into structured data, and this work originally prepared for human operations may become its advantage in the era of Personal Agents.

This article comes from the WeChat public account "APPSO," author: APPSO, which discovers tomorrow's products, published by 36Kr with authorization.