(Credit Emily Forlini)
OpenAI has unveiled a new shopping experience for ChatGPT, and we were among the first to demo it at a press event in New York City.
"Since people are already using ChatGPT for shopping, we might as well double down and help them use it more," an OpenAI spokesperson said at the event.
Well, it's more like a shopping "discovery" experience, as OpenAI employees put it, since you can't purchase products within the chat window. The company hopes to eventually integrate it with the Instant Checkout tech that allows you to do one-click purchases from Walmart and Target inside ChatGPT. However, for this launch, OpenAI didn't have enough time to enroll merchants and release it before the holiday shopping season.
The tech is quite ambitious, with multiple new interfaces. OpenAI built it on a fine-tuned version of the GPT-5 Thinking mini model. It's available now for all logged-in users to try.
In my tests, it incorporated the correct price and flagged deals. However, I wouldn't hand over all your shopping needs to ChatGPT just yet. Like most AI-powered experiences I've tested, it's not ready to revolutionize online shopping. In OpenAI's internal testing, when given complex queries with tight constraints, the model's product accuracy was 64%. That's up from 37% when using ChatGPT Search.
Still, it's "not perfect," OpenAI admits. "Shopping research might make mistakes about product details like price and availability, and we encourage you to visit the merchant site for the most accurate details."
ChatGPT: Help Me Find a Specific Couch
To get started, either select the shopping option or type /shopping in the chat window. My initial prompt asked for an off-white sectional priced below $6,000. I don't necessarily have that much to spend on a couch, but I wanted to see how expansive the tech could search, rather than just finding the cheapest option on the web.
OpenAI researcher Manuka Stratta recommended I specify delivery before Christmas to see if it could take into account shipping timelines.
The first thing that popped up was a sort of mini algorithm trainer, shown below. It surfaced images of about 10 sample couches in different styles, and I gave them all a quick yes or no. (Apologies for the picture quality. It's a photo of the demo computers set up at the event.)
Then, it crunched the data and surfaced its Top Pick, complete with bullet points, a table, and links to websites. The links mostly go to the manufacturer's website, although there were also reviews from Architectural Digest. You can also see all the other products ChatGPT initially surfaced, just in case you want to go back. This is one of the extra-mile features OpenAI built, and part of the reason the product comes off as ambitious and unique.
The Top Pick was close, but I wasn't ready to buy without further research.
(Credit: Emily Forlini)Result: A Flawed, Yet Interesting Departure From Core ChatGPT
One issue is the tech can't distinguish between real and fake customer reviews, and it doesn't know if a product appears popular simply because it's being advertised by the company. "It's impossible to know which is real; it's a hard task," an OpenAI employee explained at the event.
To get around this, the model leans heavily on Reddit reviews since the company has deemed them more "organic." (OpenAI also has a content licensing partnership with Reddit.)
ChatGPT didn't cite any Reddit posts in my couch search, but there were other problems. For one, most of its suggestions failed my price test and were far below the $6,000 range. When I gave that budget, I expected them to be close to it, since it suggested I was willing to spend more. Instead, the recommendations were mostly around $1,300-$1,600.
I mentioned this to Stratta, who said the team would look at tweaking the price programming. They were still refining the product, even just a few days before launch. If you have this problem, she recommended giving it a more specific price range, such as $4,000 to $6,000.
Another problem was how long it took to search. I stared at the monitor for what felt like a lifetime, although it was only about 3 minutes. If I was surfing Google or another website, there would be almost no wait time as I browsed the selection, changed the filters, etc. Other searches we did later took over five minutes, and one never finished before the event ended.
Stratta said the Wi-Fi might be struggling, but long load times is a common problem with ChatGPT's deep thinking products. The idea is the longer you wait, the better the answer. However, I could see this losing shoppers' attention. The best solution might be to leave ChatGPT up in a separate tab and do something else in the meantime.
I refined my search, adding that the couch should be under 110 inches on the longest side and pet-friendly. It shopped for 2 minutes and surfaced a $1,759.96 option from a brand I'd never heard of (and wasn't really my taste). I realized I had never heard of almost all the brands it had produced so far, despite my relative familiarity with this market after hours of searching on my own.
That could be a positive thing since it exposed me to new sites. But it felt off to not include the most popular furniture websites, like Wayfair, Ashley Furniture, Raymour & Flanigan, West Elm, Crate & Barrel, and Room & Board.
I refined the search and asked it to look for Pottery Barn and West Elm couches. Happily, it was able to take this into account in the new suggestions it surfaced. But the link for the West Elm couch I liked took me to a strange version of the company's website, with no price listed, and no place to click to buy. Perhaps ChatGPT had crawled a cached page. I'm not sure. In the end, it would be faster to browse West Elm myself and simply set filters for what I'm looking for.
Prediction: Chatbots Are More Likely to Be Personal Shoppers
(Credit: OpenAI)In the end, I commend OpenAI for designing a unique experience I had never seen in ChatGPT. It helped find information in a customized fashion, but given the half-baked feel, I'm not rushing to use it over a typical Google search or looking on the actual websites of the couch brands.
We can expect more shopping experiences from ChatGPT as OpenAI seeks new revenue streams. This is likely a test to see how consumers react, and the company will refine it over time. OpenAI needs to recoup its billions in funding, and an established way for a website to make money is by referring its users to other sites and taking a cut of whatever they buy.
Theoretically, with all of ChatGPT's user data, OpenAI should be able to give you expert recommendations, and the strength of its own brand should make people comfortable adding to cart once Instant Checkout is integrated. If it can this right, OpenAI may be seeing dollar signs.
Disclosure: Ziff Davis, PCMag's parent company, filed a lawsuit against OpenAI in April 2025, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.