How Much Shopping Really Happens Inside ChatGPT?

ByRetailgeek

Americans are having about 566 million conversations a month about shopping with ChatGPT. Add Gemini, Copilot and the rest and the real figure is even higher.

I got to that number the hard way, because nobody publishes it. I wanted to answer what sounds like a simple question. How many Americans use AI chatbots to find and buy products?

The published answers disagree by a factor of ten.

The estimates run from 2% to over 20%

Here is the spread.

  • 2.1% of messages are about a product someone can buy, according to the paper OpenAI’s own researchers published with academic economists in September 2025.
  • 5.2% of conversations are about buying things, according to Google’s ATLAS report from August 2026, which covered 15 million conversations with Gemini.
  • 21.6% of sessions show some commercial interest, according to a research panel called Measure Protocol. Only 7.1% showed a real signal that someone was close to buying.
  • 19.2% of conversations are commercial, according to Profound, a company that sells AI visibility software.
  • “Roughly 20% of queries in ChatGPT have a direct commercial intent,” according to OpenAI’s head of global ad sales, speaking at Cannes Lions in June 2026.

Four of those five are about ChatGPT. Same rough time period. Numbers ten times apart.

Two definitions are doing all the work

The first problem is what counts as shopping.

“What laptop should I buy for under a thousand dollars” is clearly shopping. What about “help me plan a beach trip”? What about “how do I fix a leaky faucet”? The beach trip ends in a hotel booking. The faucet ends in a hardware store run. OpenAI’s researchers count only the laptop, and that gets you 2.1%. Count anything that touches a buying decision and you get 20%.

Neither one is wrong. They are just answering different questions.

The second definition is what you are counting. A message is the actual thing you type. A conversation is the whole back and forth, usually about five messages. Profound counts conversations. OpenAI’s researchers count messages. Mix those up and your answer moves by five times, in whichever direction happens to flatter your argument.

Watch who is doing the counting

Two of these numbers come from scientists. The OpenAI paper and Google’s ATLAS report both publish their sample sizes, their methods, and the exact instructions they gave their classifiers. You can check their work.

Most of the others come from companies that do better when the number is big. That does not make them wrong. But it does mean you should want to see the method, and in most cases it has not been published.

The 20% figure needs the most care. It was said out loud at an advertising festival by the person who sells ChatGPT’s ads. There is no paper behind it. No sample, no time period, no definition of what counts. It also happens to be the highest defensible number available.

Here is the part that bothers me most. OpenAI has not said how many messages ChatGPT handles since July 2025. That figure is fourteen months old now, which is about twenty years in AI time. There has never been a US-only version of it. The one number you most need to size any of this does not exist.

My best estimate, and how I got there

Start with 2.5 billion ChatGPT messages a day worldwide, which is OpenAI’s last published figure. About 19% of ChatGPT’s traffic is American. Group those messages into conversations. Then apply the commercial share.

You land at about 566 million commerce conversations a month in the US, with ChatGPT alone.

I checked that a second way. Profound publishes its own worldwide count of commercial conversations. Take the US share of that and you get 448 million. So the range I use is 448 million to 566 million.

Remember that this is one chatbot. Gemini, Copilot, Claude and Perplexity are not in the number at all. Comscore has ChatGPT at roughly 69% of American AI assistant users, so counting everybody would push this up by a third or more.

Some of my assumptions can’t be validated (yet):

The first is that a conversation runs about five messages. That ratio is implied by the OpenAI research paper rather than stated in it, and if the real number is four or six, my estimate moves by 25% in either direction. The second is Profound’s 19.2%. Profound has never published the rules it uses to decide what counts as a commercial conversation, so I am trusting a classification I cannot inspect, from a company that sells software to brands who want to show up in AI answers. I used their number because a second, independent route landed close to it. But that is reassurance, not proof.

This should get better. Similarweb and NIQ say they will start measuring shopping inside ChatGPT and Gemini by the end of this year. OpenAI could publish a current message count tomorrow and settle half of this. Until somebody does, everyone working in this space is estimating, and the honest move is to show the arithmetic so we can all argue with it.

What this looks like compared to e-commerce sites

Monthly US visits, according to Similarweb, with my ChatGPT estimate slotted in at the middle of its range:

• amazon.com 2,300 million • ebay.com 512 million • ChatGPT shopping conversations 448 to 566 million • walmart.com 492 million • temu.com 357 million • etsy.com 250 million • target.com 226 million

ChatGPT sits right alongside eBay and Walmart, and at about a quarter of Amazon. That is a real number. It is also not the number some people want it to be.

It is fair to ask whether any of that turns into a sale. ChatGPT can take your money now. Instant checkout is live, and Best Buy checkout showed up in the wild this week. But OpenAI has never published a transaction count, only about 30 Shopify merchants were actually live in February against a launch promise of more than a million. For now this is a discovery channel with a cash register in the corner.

So this is not ChatGPT eating retail. This is ChatGPT showing up as a large commerce property, about the size of the second tier.

Where I could be wrong

A conversation is not a visit. Someone on walmart.com has already decided to shop. Someone asking ChatGPT about running shoes may be a long way from buying anything. The comparison is fair on units. It is not fair on intent.

The traffic AI actually sends to retail sites is also still tiny. Amazon got about 0.16% of its visits from AI referrals in June 2025, according to Similarweb. Adobe reports that AI referred traffic to US retail sites more than doubled in the year to May 2026, but has never published what share of total traffic that represents. Fast growth on a base nobody will state.

One thing cuts the other way. This is ChatGPT alone, so every other assistant is upside on the number.

What to do about it Monday

Stop asking whether customers are shopping inside AI. Enough of them are that it is worth your time now.

Start asking whether your products show up when they do. Your product data, your reviews and your specs have to be readable by a machine that will boil them down to two sentences. Most product pages are still written for a shopper with a browser and time to scroll.

Anyone quoting one of these numbers at you should be able to tell you which one they mean, and where it came from.

Which number have you been using?

Similar Posts

  • First post!

    ByRetailgeek

    I’ve been meaning to start this blog for years… it’s amazing what you can accomplish when you take a half day off from work.

    Anyway, I spend my days thinking about how to create great customer experiences that help shoppers make purchase decisions, and while I get called upon to work in a lot of interesting product categories, I’m especially fortunately that a good portion of my work is in the consumer electronics industry, allowing me to feed my inner geek. So it’s only natural that I named this blog, retailgeek (I’m just not that clever).

  • CES 2008

    ByRetailgeek

    January 7-10, 2008, in Las Vegas, Nevada. I didn’t realize this until after the fact, but this year marked the 25th anniversary of my first visit to CES. CES has changed a lot in 25 years and sadly I don’t think it’s what it once was. Back in the 80’s, tens of thousands of independent retailers represented a large portion of the consumer electronics distribution channel. Manufacturers needed a show like CES to introduce new products to dealers, and decide what to carry for the coming year. You could actually see exhibitors taking orders from retailers in their booths. Today, a handful of national retailers represent the bulk of the market, and they expect private visits from the manufacturers to sell their lines. The cost of exhibiting at CES (or any show) has continued to rise, and with the consolidation of the market and the emergence of the internet, the value of exhibiting has shrunk. Today, the main job of CES is to generate industry buzz and serve as a press opportunity for the Consumer Electronics industry. Small, innovative companies have very little chance of elevating their message above the noise. As often as not, Mac World, successfully steals the entire news cycle from CES anyway.

    Here are some of my favorite consumer gadgets and retail technology from the show: (more)

  • Retail’s Big Show (National Retail Federation)

    ByRetailgeek

    January 13-16, 2008, New York City. The Big Show is the trade show put on by the National Retail Federation (NRF). It’s a great trade organization that does a lot of useful work for different functional areas in the retail industry (IT, Marketing, Supply Chain, etc…).

    Because NRF’s show is so close to CES, I rarely get the chance to attend NRF, but I did make the trek to New York for this years. Frankly, I was disappointed to see that the overwhelming majority of exhibitors were focused on IT. I like a good POS system as much as the next guy, but I expected to see more variety of exhibitors than I found. (more)

  • Digital Signage Expo

    ByRetailgeek

    Las Vegas, Feb 25-26. Digital Signage Expo (DSE). This was the fourth year of the show, which used to be called the Digital Retail Expo. This is one of several shows trying to become the main destination for vendors and customers of Digital Signage and Kiosks Digital Merchandising. The show moved from Chicago to Las Vegas this year, which prompted some discussion in the blogosphere (Bill Gerba and Dave Haynes). The show is making nice progress with more exhibitors and more attendees each year. Most exciting the attendees this year weren’t just integrators looking at Digital Signage as a new industry to enter, but also end users (like retailers and public venue owners) looking to learn more about best practices. If I had to bet on one stand alone Digital Signage trade show to survive, I’d bet on this one right now. (more)

  • EuroShop

    ByRetailgeek

    Dusseldorf, Germany February 23-27. Euroshop is held every three years in Germany.

    I always get a kick out of the fact that Europe has the largest retail show in the world with over 100,000 attendees and over 1900 booths and they call it EuroShop, while in the US we have a retail show with 15,000 attendees and over 900 booths and we call it GlobalShop. (more)

  • Next up…. Global Shop

    ByRetailgeek

    The next stop on the tour is Chicago for Global Shop. My company MTI has an exhibit here, and the team has been working long hours to get the booth constructed while complying with the all the quirky Chicago exhibitor rules. The show opens tomorrow, and I’ll look forward to posting a recap. If you happen to be attending the show, feel free to stop by booth 3416 and say hello. (more)

Leave a Reply

Your email address will not be published. Required fields are marked *