
Every industry has a creed, and the creed of online commerce is the attribution window. You don't say it out loud. You don't print it in the pitch. It sits in a settings menu, a dropdown with a number in it, and the number decides who gets paid, who gets fired and which startup raises its next round.
I know this because I used to set the number.
In the 2000s I ran ad networks. Before that I built one of the first advertising game networks, and later I ran business development at LeadsCon, which meant years of rooms full of affiliates arguing about cookie windows the way medieval monks argued about angels on pinheads. Thirty days or seven. View-through or click. Last touch or first. Everyone in those rooms understood the dirty secret: the window was not a measurement. It was a negotiation. Make it long enough and you could take credit for the weather.
So when a PR firm sent me a pitch this fall about an AI shopping startup with "four retailers generating or approaching $10 million in chat-attributed revenue," I didn't hear a number. I heard a creed. And I wanted to know which one this company believed in.
The answer turned out to be the most interesting thing I've read in conversational commerce all year. It's also the reason I'm writing a nice story about a vendor, which, regulars will tell you, is not my usual liturgy.
The pitch and the confession
The company is Remark, out of Boston. It builds AI shopping advisors that sit on a brand's own website and talk to shoppers the way a good store associate would. Its clients run from Ariat, J.McLaughlin and Veronica Beard in fashion to PEARL iZUMi, Backcountry and Stance in gear, plus Pepper Home, MirrorMate and Seattle Coffee Gear in the house. The pitch from 5WPR led with the big numbers: the $10 million retailers, Pepper Home's $2 million, Sur La Table's $18.5 million in annual revenue through guided shopping, and "reported revenue lifts of 8-12%."
Two kinds of number in one paragraph. The industry blends them like it blends whiskey, and most buyers never taste the difference.
Remark's co-founder and CEO, Theo Satloff, answered my questions in writing. He didn't blend them. He pulled them apart himself.
"As part of the rollout process for each of our new brand partners we run a 50-50 A/B test comparing Remark against their existing user experience to evaluate the impact of Remark," he wrote. "A split test like this is the only way to truly measure incremental impact and we insist on it for each new partner. Our 8-12% lift numbers are based on the results from these A/B tests."
Then came the confession, the line no vendor is required to volunteer:
"While not the primary measure of lift, we do also report to brands the amount of chat attributed revenue. We attribute revenue from any order made by a shopper who purchases within our standard look back window of 30 days following a chat."
Read that slowly. The $10 million is the attribution number. It counts any order within a month of a conversation. The 8-12% is the causal number, the one that survives a control group. And the CEO, unprompted, ranks them in the right order.
I have sat across from a lot of executives in this business. I have watched men with better haircuts than mine defend 90-day view-through windows with a straight face. I can count on one hand the ones who told me, in writing, that their own headline number was the soft one.
The cathedral in Seattle
And then there's the cathedral. There's always a cathedral.
On its Q3 2025 earnings call, Amazon said Rufus, its shopping assistant, was on track for more than $10 billion in incremental annualized sales. By the Q4 release in February 2026, the figure was nearly $12 billion, with more than 300 million customers using Rufus in 2025. Trade coverage reported the figure was built on a seven-day attribution window: buy something within a week of talking to the bot, and the bot gets a halo.
The load-bearing word is "incremental." Amazon has never published the method behind it. No holdout design. No control population. No confidence interval. Just a number delivered from the altar of an earnings call, where nobody is allowed to ask follow-up questions.
Then the scripture changed. On May 13, 2026, Amazon merged Rufus and Alexa+ into Alexa for Shopping. On the Oct. 29 Q3 call, Andy Jassy said more than 400 million customers had used it in 12 months and that U.S. users spend 50% more per order. Trade press noted Amazon offered no methodology, no order count and no adjustment for the obvious problem: the person asking a chatbot about a $400 espresso machine was always going to spend more than the person buying paper towels.
And the incremental billions? Amazon stopped saying them in 2026. What it said instead was about ads. Sponsored prompts inside the assistant went billable in the U.S. on March 25, 2026, at existing cost-per-click bids. In July, Jassy said shoppers who click a sponsored prompt convert 48% more often and spend 21% more.
So here is the arc of the largest AI shopping assistant on the planet in one sentence: it went from claiming $12 billion in incremental sales to selling ads inside the answer, and never showed its homework on either.
Now picture a startup with roughly $27 million in total funding saying, in writing, we split-test every brand before we quote you a lift number. In that cathedral, that's not a footnote. That's Luther with a hammer.
The gospel according to Betty Halbreich
When I asked Satloff how Remark thinks about ads inside AI shopping conversations, he gave the clearest answer I've received from anyone in this space. I'm going to give it to you whole, because it deserves to be read whole:
"A recommendation is only valuable if it can't be bought. Betty Halbreich spent five decades at Bergdorf Goodman, and people waited months to see her because they trusted her to tell them the truth. No one waits months for someone they suspect is paid to push a product. That trust is why shoppers keep coming back. Our brands see 97-99% customer satisfaction with Remark. Our advisor works for the brand on the brand's own site, like a store associate, and recommends based on fit, not placement. Ads in AI assistants can work, but they need to be clearly labeled and kept separate from the answer. Once shoppers suspect recommendations are for sale, they'll tune them out, the same way they learned to scroll past sponsored search results."

I grew up in Brooklyn. Bergdorf's was the other side of the river and the other side of the universe. But everybody in New York knew who Betty Halbreich was. She talked her way onto the Bergdorf floor as a sales associate in 1976, got the store to build her a personal shopping office, and stayed until she died in 2024 at 96. Her whole value, the entire reason socialites and costume designers queued for her, was that she would tell you a dress looked terrible on you. She was a priest of the unbought opinion.
Hold Satloff's answer up against the year's product launches. Amazon's sponsored prompts live inside the assistant's conversation. Google's Direct Offers now appear inside the AI Mode answer. ChatGPT's shopping ads sit beneath the answer, labeled. Satloff is saying the third model can work and the first two are a slow poison. Search ads taught a generation to scroll past the top of the page. Conversational ads could teach a generation to stop believing the conversation.
He has an easier pulpit than Jassy, and that's worth saying fairly. Remark doesn't sell ads. Its advisor lives on the brand's site, recommending the brand's own goods. There's no auction for the second slot. But plenty of people with easy pulpits still say the wrong thing. He said the right one.
The church of the lookback window
To understand why that matters, you have to walk through the rest of the sanctuary.
Open the documentation for almost any AI shopping assistant sold to a Shopify store and you'll find the same three-part liturgy: a window, a definition of "engagement" and a revenue total. That's the whole service. There's rarely a sermon about what would have happened anyway.
Gorgias, one of the biggest helpdesk platforms in that world, publishes one of the more honest versions, and I want to give it real credit for that. Its Shopping Assistant credits revenue from orders placed within three days of a conversation. By Gorgias's own research, about one in seven assistant conversations ends in an attributed order. Influenced orders average $181 against $123 at the same stores, and the assistant touches about 1.9% of online revenue, rising to 3-4% in considered-purchase categories.
Then Gorgias does something almost nobody does. It tells you what the number isn't. Its methodology calls the result associative attribution, not proof of cause. It concedes that high-intent shoppers may simply wander into the chat on their own. It describes no holdout and no control group. And by its own breakdown, roughly 80% of influenced revenue rests on the thinnest possible thread: a conversation happened, and later an order happened. Only about 18% comes from the bot recommending something the shopper then bought.
That's a vendor reading its own fine print out loud from the pulpit. It should be the floor for the industry. It's closer to the ceiling.
Further down the long tail, the windows stretch like taffy. One Shopify chat app credits a sale within 10 days of a recommendation click. Another analytics product lets merchants toggle between 7, 14, 30 and 60 days, which is the dashboard equivalent of asking the defendant to pick the jury. Others count any order where an AI assistant shows up anywhere in the customer journey.
There are heretics, and they deserve their names in print. Quickchat only credits an order when the cart itself came from the chat, and calls that deliberately conservative. Alhena has publicly gone after vendors who count any session where the chat widget merely appeared on screen.
Remark's 30-day window sits at the generous end of that spectrum. If that were all Remark reported, this would be a different story, and you know which kind I'd write. It isn't all Remark reports. Remark puts a randomized test underneath it, and its CEO says the test is what counts.
Who these people are
Remark was founded by Theo Satloff, Ian Patterson and Carl-Philip Majgaard. Satloff was a software engineer at Amazon before this, which makes his measurement discipline either a reaction or an inheritance, and I suspect I know which. He also co-founded Outdoorly. The company raised a $10.3 million seed in May 2024 from investors including Stripe, Spero Ventures, Shine Capital and Neo, then a $16 million Series A in July 2025 led by Inspired Capital, with Stripe, Neo, Spero, Shine and Visible Ventures following.
The founding idea was almost old-fashioned. Remark started with humans. It paired shoppers with real product experts, ski techs and stylists and estheticians and new parents, then trained AI personas on what those people knew and how they talked. By the Series A, Remark said more than 60,000 experts had fed the system. It reported 4x year-over-year revenue growth, 100% customer retention and roughly 130% net dollar retention.
The customer list tells you where advice is worth money. It started in places where a wrong purchase hurts: bikes, skis, cycling kit, socks that cost more than socks should. Then it moved into made-to-order curtains and custom mirror frames, where one wrong measurement means a return nobody can resell. Now it's luxury fashion. In August 2026 Remark launched a real-time virtual try-on that renders a garment on a shopper's moving body at full HD frame rates, straight from the product page, no app. Satloff's line at launch was that shoppers ignored older try-on tools because clothes aren't stickers and people aren't mannequins.
That's the product 5WPR wanted me to write about. The thing I actually care about is how they count.
The numbers that don't drift
Remark's public case studies are what vendor case studies always are, mostly attribution totals in nice fonts:
Seattle Coffee Gear: $6.5 million-plus in chat-driven revenue, 99% CSAT
Pepper Home: $2 million-plus in chat-driven revenue, 33% chat conversion
MirrorMate: $1.5 million in chat-driven revenue, 32% chat conversion
PEARL iZUMi: 30% of chatting shoppers converted, 7x ROI, 8% increase in new revenue attributed, 24% higher average order value
Alchemy Bikes: 20x ROI, 13% higher average order value
PEARL iZUMi's study describes the design: an 8-week A/B test. Remark's homepage leads with a 9% revenue lift and a 20% conversion lift. Back in 2024, Remark told TechCrunch its partners saw about a 10% net revenue gain.
Eight. Nine. Ten. Eight to twelve. Two years, several surfaces, the same range.
I've spent a career watching vendor numbers. Invented numbers inflate. Each deck is a little braver than the last, because the last one worked and nobody checked. Numbers that come out of control groups get boring. They sit in a band and refuse to grow, because reality refuses to grow on schedule. Remark's lift numbers are boring in exactly the way honest numbers are boring.
The 32% and 33% chat conversion rates need a sentence of context, and Remark isn't hiding behind them. Somebody who opens a chat to ask about curtain fabric is already holding a credit card. One in three of them buying says the bot didn't scare them off. It doesn't say the bot conjured them. Which is, again, exactly why the CEO says the split test comes first.
The salesman in the answer box
Remark's discipline would be a nice footnote any other year. This year, every major AI shopping surface started selling placement, and the measurement question became a moral one.
OpenAI began testing ads in ChatGPT in the U.S. in February 2026. It opened self-serve to all U.S. businesses on May 5 and dropped the minimum spend entirely. That month it added product-feed shopping ads, running off essentially the same catalog file retailers send Google Shopping and shown as labeled sponsored cards under the answer. CPMs reportedly slid from $60 at launch to as low as $25 by mid-April. By August the ads had reached 31 European countries.
What OpenAI didn't keep was the cash register. Instant Checkout, which let shoppers buy inside ChatGPT, launched Sept. 29, 2025 and was wound down in March 2026. Reporting at the time put the number of Shopify merchants who ever went live at around a dozen. People were happy to confess their shopping desires to a chatbot. They just wouldn't hand it their wallet. Purchases now go back to the merchant's own site.
Google unveiled Direct Offers for AI Mode at NRF on Jan. 11, 2026, piloting with Petco, e.l.f. Cosmetics, Samsonite, Rugs USA and Shopify merchants. The pitch is an AI that decides the precise moment to slide a sponsored discount in front of a shopper, a personal salesperson approaching you with a deal. Google framed it as built to drive incremental conversions, and it has since moved from the AI Mode product viewer into the AI Mode answer itself.
The early performance claims are vendor-sourced and directional. Criteo said a February 2026 sample of 500 retailers found LLM-referred visitors converting at roughly 1.5 times other referral channels. Pacvue, StackAdapt and others now pipe ChatGPT ads into the same retail media dashboards that have run Amazon ads for a decade, measured with the same attribution plumbing.
Nobody is publishing holdouts there either. We are building the next great advertising surface on the same faith-based accounting as the last one.
Who keeps the confession
There's an older word for what happens when a shopper talks to an AI for twenty minutes about hiking boots. It's a confession. What they want, what scares them, what they rejected, the price that made them flinch. The question of 2026 is who gets to keep it.
"The brand owns the shopper & conversion data," Satloff wrote. "Remark builds the intelligence system that makes sense of the data the brand is collecting, allowing brands to package insights to use for internal goals. For example, a product recommendation from an advisor can trigger a follow-up email flow from the brand to the shopper."
Then the paragraph every brand marketer in America should tape to the monitor:
"Brand ownership is becoming increasingly important since conversations inside ChatGPT and other assistants belong to those platforms. In these cases the brands only see the referral. On their own sites, brands have the ability to thoroughly learn about their shopper behavior, which tells them more about the customer experience than referral data alone could ever. This shapes forecasting, merchandising, planning, product development, marketing, and more."
That's the whole war in four sentences. Inside ChatGPT, OpenAI hears the confession. The brand gets a click with a referrer string and, if it paid, a conversion event. On the brand's own site, with its own advisor, the brand keeps the transcript: every "does this run small," every competitor comparison, every hesitation at checkout. For a luxury house, that's what the best store associate used to carry around in her head, finally written down.
The platforms are betting shopping moves to them. Remark is betting that whoever keeps the conversation keeps the customer. I've watched the open web lose that bet before, to search and then to social. I'd like to see somebody win it for once.
Where the money goes, and where it doesn't yet
On budgets, Satloff was more modest than the market, which is to say more honest.
"Retail acquisition budgets will follow anywhere consumers are spending time, particularly places they are discovering products and doing research," he wrote. "Today brands are making investments to be 'discovered' and understood by LLMs and agents by organizing their product data and making sure it is easily consumable by AI tools. Over time, we expect to see bigger and bigger investments from brands in this area, but there is still much to be learned over the next few years about how and where consumers will prefer to do their shopping."
After Instant Checkout's six-month life, that reads less like caution and more like accuracy. Brands are spending on product feeds and AI visibility now. The great migration out of paid search and retail media into AI assistants exists mostly in pitch decks. It hasn't shown up in anyone's audited numbers.
The catechism brands should demand
So here's a short catechism, borrowed from Remark's own practice, for anyone buying an AI shopping product in 2026:
Show me the holdout. If you can't tell me what happened to the half of my traffic that never saw your bot, your revenue number tells me who used it, not what it did.
Put the window in writing. Three days, seven, ten, thirty and sixty are all in market. They aren't comparable, and your choice tells me something about you.
Split the credit. How much revenue came from a product the assistant recommended and the shopper bought, versus a conversation that merely happened before an order? Gorgias published that split. You can too.
Tell me who keeps the transcript. On my site, I do. Inside someone else's assistant, I get a referral.
Tell me whether the answer can be bought. If it can, show me the label, and tell me whether a paid placement can ever push out an honest recommendation.
Remark answers the first four on the record and holds the right line on the fifth. Its pitch led with the softer number, because that's what pitches do. Its CEO, when it mattered, put the harder one first.
I spent years in a business where the window was a negotiation and the number was a prayer. It's strange, and a little moving, to find a vendor in the newest corner of that business who'd rather show you the control group. Betty Halbreich would have told you the dress looked bad. Somebody in AI commerce is finally willing to tell you which number is real.
Method, limitations and disclosures: this piece is built on written answers from Remark CEO Theo Satloff, provided through 5WPR in November 2026, plus Remark's public case studies, funding announcements and launch materials, Amazon earnings statements, Gorgias's published methodology, and trade coverage of OpenAI's and Google's 2026 ad and commerce products. Remark's revenue, lift and satisfaction figures are company-reported and were not independently audited; ADOTAT did not see the underlying A/B test data, and Amazon's incremental figures and attribution window come from earnings statements and trade reporting, not published methodology. ADOTAT has no financial relationship with Remark, 5WPR, Amazon, Google, OpenAI or Gorgias. The author ran ad networks in the 2000s and was director of business development at LeadsCon, has known 5WPR for years, and ADOTAT has previously reported critically on 5WPR's founder. Remark did not review this story before publication. ADOTAT is reader-supported. No paid placement in editorial.

