Sephora did not open a shop inside ChatGPT

For about a week in October, a particular post went round LinkedIn. Sephora has launched inside ChatGPT. Customers can now discover products without visiting the website. Is ChatGPT the new storefront, or are we still optimising the old one?

It is a good post. It is also describing a retreat as an advance, and the detail that would have changed the conclusion was available to anyone who looked for it.

The timeline nobody put in the post

OpenAI launched Instant Checkout in ChatGPT in September 2025, built with Stripe. You could buy inside the conversation. Merchants paid a 4% fee on completed transactions, which was less than half what Amazon charges.

In March 2026 OpenAI dropped the fee and deprioritised native in-chat checkout. A spokesperson said Instant Checkout is "transitioning to apps."

Sephora's ChatGPT app was piloted in the United States that same month. Coverage at launch described beauty advice, product recommendations, and Beauty Insider loyalty benefits inside the chat. On buying, it reported that payments and checkout were planned for future updates.

Read those three paragraphs in order and the story inverts. Sephora's app is not the storefront arriving. It is the thing that exists because the storefront closed. "Transitioning to apps" is the sentence that produced the Sephora app, and the app does the half of commerce that still works in a chat window: discovery. The half that OpenAI tried at full scale, with the largest assistant in the world and the most capable payments company behind it, was pulled back within six months.

A storefront where you cannot pay is not a storefront. It is a catalogue.

Why in-chat buying stalled

The adoption numbers are unflattering and they were published. Among people familiar with Instant Checkout, 8% had used it to buy something. Across US online adults, 33% had used AI agents to find and compare products in the prior three months, but only 10% had used one to make a purchase, and 61% said they would not be comfortable doing so.

That last figure is the one worth sitting with. A clear majority of people who are perfectly happy to ask an assistant what to buy are not willing to let it take the money. Those are different acts of trust and the gap between them is not a user-education problem that will close with familiarity. People are comfortable with advice and uncomfortable with authority over their card, which is a reasonable position rather than a failure of understanding.

Now the part that cuts the other way

Here is where the sceptical read, including the one I started with, goes wrong.

Adobe Analytics, measuring more than a trillion visits to US retail sites, reports that AI-referred visits convert at a rate 60% higher than non-AI traffic. Those visitors generate 53% more revenue per visit, spend 59% longer on site, bounce 33% less, and add to cart 28% more often. July 2026 was the eleventh consecutive month AI traffic outperformed non-AI traffic on conversion.

Salesforce, on behavioural data from more than 1.5 billion shoppers across 37 countries, found traffic referred from AI chats growing between 150% and 428% year on year in every quarter it measured, while overall traffic grew in the single to low double digits.

So the dismissive version is wrong too. This is not hype with nothing underneath it. AI-referred traffic is the best-converting traffic many retailers have.

The number nobody quotes

Adobe publishes the conversion lift. It publishes the growth rate. It does not publish AI traffic as a share of total traffic.

Ahrefs did measure that, across 3,000 websites: AI chatbots accounted for 0.17% of the average site's visitors and 0.12% of views. Ahrefs itself says this is probably understated, because some AI visits land in analytics as direct traffic, and the picture varies enormously: the top sites in its sample took about 6% of traffic from AI and one reached 18%.

Put the two together and you get the honest sentence, which is less exciting than either the enthusiast version or the dismissive one. AI sends very little traffic, and the little it sends is unusually good.

A 60% conversion lift applied to a fraction of a percent of your visits is a rounding error in this quarter's revenue and a serious signal about next year's. Both of those are true at once. Almost every article quoting the lift leaves out the share, and the omission always runs the same direction.

The error underneath all of it

The lift is being reported as though the AI causes it. Nobody has shown that, and there is a simpler explanation sitting in plain view.

Think about who actually arrives through an assistant. Someone described a problem in their own words, read an answer, and then deliberately clicked through to a retailer. They have already done the comparing. They have already narrowed it. By the time they land on a product page they are further down the funnel than almost anyone arriving from a search result or an ad.

The assistant is not manufacturing intent. It is filtering for it, and handing over the people who were going to buy anyway. That is a selection effect, and it is the most ordinary explanation available. It may not be the whole story, and the trust transfer that people describe, where a recommendation reads as an answer rather than an advertisement, is probably real and probably contributes. But nobody has separated the two, and in the meantime the industry is quoting a correlation as a mechanism.

This matters practically. If the lift is selection, then "get recommended by AI" does not create demand, it intercepts demand that already existed. The strategic implications are completely different and the numbers look identical.

What this means if you sell in Lebanon or the Gulf

Three things follow, and they are cheaper than the posts suggest.

Do not budget for in-chat checkout. It was tried at the largest possible scale and pulled back inside six months. If an agency is selling you an agentic commerce package, ask them what happened to Instant Checkout in March 2026 and watch what the answer does.

The Sephora move is not replicable and that is the point. That app works because Sephora is a retailer with a loyalty database. Linking Beauty Insider is what makes the recommendations worth having. A single brand, or a store doing a few hundred orders a month in Beirut, has no equivalent to plug in. The lesson is not "build an app in ChatGPT". It is that first-party data is what buys you a seat anywhere this goes.

Be the page the assistant reaches for. This is the part that is actually available to you, costs nothing beyond work you should be doing anyway, and compounds. Specific, dated, sourced answers to the questions buyers in your market actually ask. It is the same discipline that gets you cited in a payment gateway comparison or a question about getting paid from Lebanon, and it does not depend on any platform keeping a feature alive.

What we do not know

Being straight about the holes, because the posts going round do not have this section.

• Nobody has published AI-referred conversion data for MENA. Adobe's figures are US retail. Whether they hold in markets with heavy cash on delivery and lower card penetration is unknown, and we would not assume it.
• Adobe does not explain how it defines or attributes an AI referral, which matters when a share of those visits arrive as direct traffic.
• No results have been published for the Sephora pilot. Launch coverage, stated goals, no metrics. Anyone telling you how it is performing is guessing.
• Whether Instant Checkout returns in another form is unknown. OpenAI framed the change as evolving its strategy, not abandoning commerce.

The thing worth watching

Discovery is moving into the assistant. Purchase moved back out. For the first time in two decades of e-commerce those two steps are splitting across platforms that do not share data.

Which means attribution breaks. A customer asks an assistant, gets your name, and arrives at your site some time later with no trace of where the decision was made. You will see the conversion and not the cause. Most of this will land in your analytics as direct traffic, which is where unattributable demand has always gone to die.

That is the question nobody is asking while everyone argues about storefronts: not whether to be visible in AI, but how you will ever prove it worked. We do not have a good answer yet, and neither does anyone selling you one.

Frequently asked questions

Sources

• Nuvei, OpenAI tried to checkout, it couldn't, for the Instant Checkout timeline: a September 2025 launch built with Stripe, a 4% fee on completed transactions confirmed in January 2026, and the March 2026 shutdown in which the fee was dropped and native checkout deprioritised, with the OpenAI spokesperson quote that Instant Checkout is "transitioning to apps." Also for the adoption figures: 8% of people familiar with Instant Checkout having used it to buy, and across US online adults 33% having used AI agents to find and compare products in the prior three months, 10% having purchased through one, and 61% saying they would not be comfortable doing so. Nuvei is a payments provider and therefore an interested party in how agentic commerce is judged, which is why the timeline is used here and its framing of the causes is not.
• Chain Store Age, Sephora pilots app within ChatGPT, for the Sephora pilot being US-only, the app surfacing beauty advice and product recommendations on a request such as finding a foundation for dry skin, the optional Beauty Insider account link that brings saved preferences and member benefits including free shipping promotions and samples into the chat, and the statement that payments and checkout within the app are planned for future updates rather than live. No performance metrics are reported.
• mind Retail, While OpenAI buries native payment, Sephora and Carrefour bet on discovery in ChatGPT, for the Sephora pilot date of 24 March 2026, Carrefour's listing in the OpenAI app store in France, and the framing that retailers are entering ChatGPT for discovery rather than purchase while OpenAI steps back from native payment. Partially paywalled, so only the visible portion is relied on here.
• Digital Commerce 360, Adobe: AI-referral traffic spending, converting more than others, for the Adobe Analytics figures on US retail sites in July 2026: AI-referred visits converting 60% higher than non-AI traffic for an eleventh consecutive month, 53% more revenue per visit, 59% longer time on site, 33% lower bounce rate, 28% higher add to cart, and traffic up 62% year on year and 1,219% against October 2024. Based on more than a trillion visits. Adobe does not publish AI traffic as a share of total traffic, and the article notes Adobe does not explain how it defines or attributes an AI referral.
• Salesforce, Shopping's new first step: agentic search grows 200%, for traffic referred from AI chats growing 150% to 428% year on year in every quarter measured between Q1 2024 and Q1 2026 while overall traffic grew in the single to low double digits, on behavioural data from more than 1.5 billion shoppers across 37 countries. Salesforce publishes no conversion-rate comparison in this piece, which is why the conversion claims above are sourced to Adobe instead.
• Ahrefs, AI traffic study, for the share nobody else quotes: AI chatbots accounting for 0.17% of the average website's visitors and 0.12% of views across 3,000 websites, tracking seven assistants, counting only visits whose referral source was visible in analytics. Ahrefs states the figure is likely understated because some AI visits appear as direct traffic, and reports wide variation, with top sites in the sample at about 6% and one at 18%. Smaller sites took a larger share, 0.56% for those under 999 monthly visitors against 0.13% for those over 10,000.

Every source above was opened and read on 24 October 2026. The selection-effect argument in this article is our own reading of the Adobe data and is not a finding Adobe published: Adobe reports the correlation, and no study we could find separates selection from causation. The practical recommendations are our own. This story is moving quickly and the platform behaviour described here, in-chat checkout in particular, could change again without notice, so confirm the current position before acting on any of it.