- Best Buy told shoppers on 14 September 2026 that they can find gifts and finish the purchase inside ChatGPT, two days before four other retailers described what that traffic is really worth to them.
- AI referrals sat under 2% of traffic for leading retailers in early 2026 and reached roughly 5% by late 2026, with some shops approaching 10%.
- Four named shops measured the same channel and disagreed completely: one saw conversion rise from 3% to over 13%, another saw it land below its own site average.
- The largest independent study, covering 973 shops and 20 billion dollars of orders, found ChatGPT referrals underperformed organic search, affiliate and paid search.
- Nothing in this week's news requires a small shop to change its storefront. It does change what you should be measuring by November.
On 14 September 2026, Best Buy told its customers they could look for gift ideas and finish the purchase without leaving ChatGPT. Two days later, Digital Commerce 360 published interviews with four other online retailers about how much traffic those assistants actually send them. The two stories landed in the same week and they pull in opposite directions, which is exactly why they are worth reading together.
One says the biggest retailers are wiring themselves into the assistant before the holidays. The other says the merchants already receiving that traffic cannot agree on whether it is any good. If you sell things and you have been wondering whether to do something about AI shoppers this season, the honest answer sits in the gap between those two reports.
What actually happened in the middle of September?
Best Buy opened gift discovery and checkout inside ChatGPT, and a separate report put hard numbers on what AI referrals are doing at four smaller retailers. Neither event is a platform change that forces anything on you, and that distinction matters for how you spend the next six weeks.
The Best Buy move is the more visible one. Digital Commerce 360 reported that the retailer announced on 14 September that shoppers can find gift ideas and shop its products through ChatGPT, finalising the transaction in the chat window. It follows Ask Blue, the conversational assistant Best Buy put on its own site in August, which compares products, checks compatibility and answers support questions using its product data, customer reviews and stock information.
The context around that announcement is a retailer with room to experiment. Its marketplace now carries nearly a million products from more than 5,000 brands, produced 300 million dollars of gross merchandise volume in the second quarter, and is projected to reach 1.3 billion dollars across the full fiscal year. Next day delivery reaches nearly every major metropolitan area, against fewer than half in the first quarter. A company operating at that scale can afford to put a bet on a channel that might not pay off.
The second report is the one a small shop should read more carefully, because it is made of merchants rather than press releases.
How much traffic are AI assistants really sending?
Roughly 5% of website traffic for leading retailers by late 2026, up from under 2% at the start of the year, with some shops approaching 10%. That is a real channel and a small one, and both halves of that sentence are load bearing.
In July 2026 alone, AI associated referral traffic to ecommerce sites rose 62% against the previous year, and shoppers arriving that way generated 53% more revenue per visit than everyone else, according to the same Digital Commerce 360 survey of four retailers. Growth like that is genuinely fast. It is also growth measured from a base small enough that a single good month distorts it.
What makes the report unusually useful is that it names the shops and quotes the people who read the dashboards. Their accounts do not line up.
| Retailer | AI share of traffic | What they say it does |
|---|---|---|
| Bero, non alcoholic beer | Under 5% | Hyojin Park sees ChatGPT in the referrers only occasionally and calls it nothing significant |
| Povison, furniture | About 5%, heading for 10% by early 2027 | Ayden Lin describes buyers who are already informed and move through the decision quickly |
| La Joya Jewelry | 5% to 6%, up from 2% to 3% | Nishit Mehta says a normal 3% conversion rate rises above 13% for AI arrivals |
| Edible Brands | Under 10%, of which ChatGPT is about 95% | Erica Randerson reports conversion above 5% but below the site average of more than 10% |
Read that table twice. La Joya says AI traffic converts more than four times better than its normal rate. Edible Brands says AI traffic converts at roughly half its site average. Both are describing 2026, both are talking about the same handful of assistants, and both are almost certainly telling the truth about their own data.
Why do four shops get four different answers?
Because conversion from any channel is mostly a fact about the shop, not about the channel. A gift retailer with a 10% site average and a jewellery shop with a 3% one will report opposite results from identical visitors.
Edible Brands is the clearest illustration. Its conversion from AI referrals is above 5%, which would be a strong number almost anywhere. It looks weak only because the rest of the site converts above 10%, which Erica Randerson attributes to an audience arriving with a gift already in mind. A visitor who is researching is worth less to that business than a visitor who has decided. At La Joya, where the baseline is 3%, a researched visitor is worth a great deal more.
This is the part that gets lost when a statistic travels. A headline saying AI shoppers convert better is not wrong, it is simply an average across shops whose baselines have nothing to do with each other. Your own number is the only one that tells you anything, which is the same logic behind our piece on which assistants actually send referral traffic.
Does AI traffic convert better or worse than search?
The published evidence says both, and the disagreement about AI shopper conversion is not a rounding error. Two serious datasets reach opposite conclusions, and knowing why protects you from acting on whichever one you happened to read.
On the optimistic side, Adobe Digital Insights reported that AI referred visitors in March 2026 were 42% more likely to purchase and generated 37% more revenue per visit, spending 48% longer on site, visiting 13% more pages and bouncing 32% less. Similarweb put ChatGPT referral conversion at roughly 11.4% against 5.3% for organic search.
On the other side sits the largest independent measurement anyone has published. Practical Ecommerce assembled the comparison, and the underlying study by Maximilian Kaiser and Christian Schulze analysed twelve months of first party data from August 2024 through July 2025, covering 973 ecommerce websites, 20 billion dollars of orders, nearly 50,000 transactions attributed to ChatGPT and 164 million from traditional channels. It found ChatGPT accounted for less than 0.2% of ecommerce traffic. It found ChatGPT referrals converted about twice as well as paid social, and worse than most other channels: organic search converted about 13% better, paid search 45% better, affiliate 86% better.
Both results can be correct. Adobe measured post click behaviour across many tools in the United States in 2026. The academic study used last click attribution across 49 countries in a window ending July 2025, which undercounts an assistant that shaped a decision days before the sale. Different question, different answer.
Timing alone explains a lot. Adobe's own series shows AI traffic converting 38% worse than other channels in March 2025 and 42% better in March 2026. A study whose data ends in July 2025 is measuring a different product than one reading September 2026, and neither is stale research so much as a snapshot of a channel that keeps moving.
Is it worth chasing checkout inside an assistant?
For most small shops, no, not as a project for this season. The integrations are real but narrow, and what you give up is the customer relationship you spent years building.
ChatGPT has signed checkout partnerships with Target, Etsy, Walmart and Instacart, and Perplexity extended checkout to all United States users after launching it for subscribers, as Modern Retail documented. Shopify enabled purchases inside chat for merchants including Skims and Glossier. That is genuine infrastructure and it is arriving quickly.
The constraints are equally genuine. Adam Brotman, a former Starbucks executive, lists them plainly: purchases limited to a single product, no loyalty programme integration, limited first party customer data, anonymised transactions. Rachel Levy, chief operating officer at Brooklinen, reduced the problem to one sentence about wanting to know who is new and who is a repeat customer. Jamie Norwood of Winx Health called AI visibility a black box that is very hard to track and constantly changing.
There is also a demand question nobody has answered. Omnisend research cited in the same report found only 34% of consumers would authorise an AI agent to make purchases on their own. Mikey Vu, who leads Bain's retail AI practice, compared the arrangement to the food delivery and travel platforms, where retailers effectively handed over the top of the funnel and never got it back. We covered the mechanics of that trade in more detail in our look at what agentic commerce changes at checkout.
What should a one person shop actually do before November?
Three things, none of which require new software. They are all measurement and copy, and they pay off whether or not AI traffic ever becomes a large channel for you.
Tag the referrer before the season starts. Hyojin Park's comment is the most quietly important line in the whole report: session reporting does not yet identify AI traffic cleanly, so she has to dig through referral sources by hand. If you cannot separate assistant traffic from everything else in your analytics, you will finish the holidays with no idea whether any of this mattered. Set that up in September and the January post mortem becomes possible.
Read your own product page as a stranger. Ayden Lin and Nishit Mehta describe the same visitor from different businesses: someone who has already researched, expects detailed product information, and has little patience for a page that hides the specification. Assistants also lean on the same material. AI engines draw on product descriptions, customer reviews, articles and community posts when deciding what to recommend. A page that answers the obvious question in plain words serves both audiences, which is the argument we made about product pages that AI shoppers can actually read.
Decide deliberately whether checkout elsewhere is a trade you want. Not this week, and not under deadline pressure. The relevant question is not whether the technology works, it is whether an anonymised single item sale with no loyalty data is worth more to you than the customer record you would otherwise keep. For a shop whose repeat rate carries the business, that answer is often no.
How do you measure any of this without a data team?
Start with the referral source field you already have and one segment. You do not need attribution modelling to learn whether assistant traffic is arriving and what it does when it lands.
The practical detail worth knowing is that this traffic does not look like your mobile traffic. During the 2025 holidays, 73.4% of AI referrals came from desktop devices even though mobile drove more than half of all online sales, Practical Ecommerce noted in its review of the season. People research on a large screen and buy on a small one, which means an assistant may have started a purchase that your analytics eventually credits to direct mobile traffic.
The same review is useful for a second reason: it is openly sceptical about its own headline number. Adobe reported a 693% increase in AI driven holiday traffic in 2025 without publishing the baseline volume or AI's share of total referrals, and Adobe had reported a roughly 1,300% jump the year before. Growth from a small base produces enormous percentages. The number is true and it is not evidence that AI drove the season.
One more caution belongs in any honest measurement plan. AI discovery does not always end in a visit at all. In September 2025 Penske Media sued Google over AI Overviews, saying its affiliate revenue had fallen by more than a third from peak and traffic to its sites had halved. Publishers met the zero click problem first, and there is no reason product research is immune to it. An assistant that tells a shopper your price and your return policy without sending them to you has influenced the sale and left no trace in your referral report.
Which products do AI shoppers research before they buy?
The ones people already researched before assistants existed. Electronics, toys, appliances, video games and personal care drew the most AI referrals during the 2025 holidays, which are the categories where a buyer wants a specification explained rather than a photograph admired.
That pattern is worth sitting with if you sell something else. A shop selling a 30 dollar candle is not in the same position as one selling a 900 dollar mattress, because nobody opens an assistant to compare candle specifications. The businesses in the Digital Commerce 360 survey that reported the most engaged AI visitors, furniture and jewellery, both sell considered purchases with real differences between options. Bero, which sells non alcoholic beer and saw nothing significant in its referrers, sells something people buy on taste and habit.
So the honest version of the advice is conditional. If your catalogue rewards comparison, assistants are a channel worth preparing for this season. If it does not, the same effort spent on your email list will almost certainly return more, and you can revisit the question next year without having missed anything. That judgement is yours to make about your own products, and it is a better use of an afternoon than reading another forecast.
There is a second pattern underneath the category list. Every one of those categories has substantial written material about it: reviews, comparison articles, forum threads, manufacturer specifications. Assistants recommend what they can read. A category with thin published information is one where the assistant has little to work with, which suppresses referrals regardless of how good the products are. Our guide to writing product descriptions with AI covers the structural side of making a listing legible to both readers and machines.
Can you stop assistants from using your catalogue?
Technically yes, practically only if you are very large. Amazon blocked ChatGPT's crawlers from its product pages, prices and reviews, and Bain's Mikey Vu observed that smaller retailers cannot sustain that kind of resistance.
The asymmetry is straightforward. Amazon can refuse an assistant because shoppers will go to Amazon anyway. A shop nobody has heard of is in the opposite position: being absent from the assistant's answer costs it the consideration it was never guaranteed in the first place. Blocking is a lever available to companies that do not need it.
What smaller merchants are doing instead is closer to old fashioned publishing. Jamie Norwood of Winx Health added citations and accolades to her company's Wikipedia page specifically because AI systems read it. That is not a trick so much as an acknowledgement that assistants assemble answers from whatever authoritative text exists, and most small brands have very little of it. If you have never thought about what a machine reading about your business would find, our piece on whether to block or allow AI crawlers on a shop works through the decision properly.
Ali Furman of PwC framed the whole shift in a way worth keeping: ecommerce is not going away, but there is a new channel to consider. A new channel at 5% is not an emergency. It is a thing to have an opinion about before it is 15%.
What this week does and does not change
It does not change what your shop should look like. Nothing in the Best Buy announcement creates an obligation for a small merchant, and nothing in the four retailer survey suggests a shop is losing sales today by ignoring assistants. The channel is around 5% of traffic for businesses larger than most readers of this blog.
What it changes is the cost of not knowing. Two years ago you could reasonably treat AI referrals as noise. Now there are named merchants reporting five to ten percent of arrivals from a channel that behaves differently from search, and a large retailer building a storefront inside the assistant before the busiest weeks of the year. The shops that will be able to make a sensible decision in January are the ones that tagged the traffic in September.
That is a smaller task than the headlines imply, and a more useful one. If you are building or rebuilding a storefront and want the underlying product data structured so assistants and search engines can both read it, our AI ecommerce store builder generates catalogue pages with that structure by default. If you are not rebuilding anything, the referrer tag and a clear specification on the product page will carry you through the season perfectly well.
The four shops in this report will have better answers by February. So will you, but only if you start counting now.