Retailers built their tech stacks for a web that no longer exists
Carrefour put its entire grocery catalogue inside ChatGPT in March. Shoppers in France can now build a basket, check stock and pick a delivery slot without ever opening the retailer's website. One integration, 26 million ChatGPT users in reach. And it exposed a problem most retailers haven't had to face yet: their commerce infrastructure assumes a storefront first world that's disappearing under them.
The next sale probably won't happen on your homepage.
It'll happen wherever the customer already is. A chat window, a social feed, a marketplace, an AI assistant haggling on their behalf. Winning there has less to do with the shop window and a lot more to do with the plumbing, meaning back-end systems that can push accurate product data to any surface, instantly.
The front door is moving
For twenty years, e-commerce strategy meant driving traffic to a website and converting it there. That playbook is wearing thin. McKinsey projects that agentic commerce could orchestrate up to $1 trillion in US B2C retail revenue by 2030, with AI agents handling discovery, comparison and checkout on the shopper's behalf. Whether it lands on that number or half of it, the direction is set, and you can already watch it in browsing behaviour.
We covered the rise of AI guided product discovery earlier this year: shoppers increasingly arrive at a product because an algorithm surfaced it, not because they typed a query. Vinted users report buying mostly from their personalised homepage feeds. Debenhams Group now lets customers finish a purchase inside an AI chat, with PayPal handling payment in the background.
None of this works if your product data lives in a rigid monolith. An agent can't recommend an item it can't read. It won't recommend one with stale pricing twice.
Composable wins, mostly
The architecture most mid-market retailers are landing on is composable commerce: storefront, catalogue, feed engine and checkout as separate pieces stitched together with APIs, so you can swap one out without rebuilding the lot. SportShoes.com, a UK running specialist founded back in 1982 and now turning over £100 million-plus, rebuilt on exactly this logic. Stable commerce foundations first. The native app only came once the stack could carry it.
Worth saying plainly: composable isn't automatically better. A small retailer with one channel and a simple catalogue will get more from a decent all-in-one platform than from wrangling six vendors and their six invoices. Composability pays off when you sell across many channels and your product data changes fast. Below that line it's mostly self-inflicted complexity.
Above it, the vendor market has consolidated around connected ecosystems rather than point tools.
One example of the direction is Commerce, the company behind BigCommerce, Feedonomics and Makeswift, which now packages its platform, feed management and visual storefront builder as flexible commerce solutions built to keep product data AI ready across hundreds of channels from one connected stack.
The "AI-ready" bit matters more than it sounds. Feed quality was a nice-to-have when Google Shopping was the only machine reading your catalogue. Once autonomous agents read it directly, structured and current product data is the difference between getting recommended and being invisible.
Checkout is still where the money leaks
All the discovery cleverness in the world can't rescue a broken final step. Baymard Institute's meta-analysis of 50 studies puts average cart abandonment at 70.22%, a figure that has barely budged in a decade. Seven in ten shoppers who add an item never finish. On mobile it's nearer 80%. Surprise costs at checkout remain the biggest fixable reason people bail, cited by 39% of abandoners in Baymard's research, with forced account creation close behind.
The number I'd pin to the wall, though, is 35.26%. That's the conversion lift Baymard reckons the average large e-commerce site could gain purely through better checkout design. Not new traffic. Not new channels. Fixing forms, surfacing costs earlier, trimming the field count from the typical 23 form elements down to the ideal 12 to 14 that Baymard's checkout testing keeps arriving at. Retailers on modular stacks can test express payments, one-page flows and wallets without a replatforming project; retailers on ageing monoliths usually know exactly what's wrong with their checkout and still can't ship the fix.
And then someone has to post the thing
Commerce technology gets treated as a purely digital problem. It isn't. The order still arrives in a box, and the box is now part of the stack. Our recent look at how retail technology is transforming e-commerce packaging found retailers using data analysis to optimise box dimensions, cut transit damage and tie packaging specs straight into warehouse automation.
Unglamorous next to shopping agents, sure. But a customer opening a crushed parcel doesn't care how clever the recommendation engine was, and damage-driven returns land straight on the margin. Standardised box sizes also speed up packing lines, which matters as same-day delivery windows keep shrinking.
The next 12 months
Resist the urge to chase every headline. Most of the ground is covered by three moves, and they aren't equally weighted.
The big one: audit your product data as if an AI agent were the customer. Structured? Current? Reaching every surface a shopper might meet it on? If not, feed management is your highest-return project, well ahead of anything cosmetic.
Then run your checkout against Baymard's benchmarks. A 70% abandonment rate is normal. Accepting it is optional.
The last one takes an afternoon. Ask your team how long it would take to plug into a major AI platform's shopping integration if it opened tomorrow. Quarters rather than weeks? That's your architecture telling you something.
Retailers who treated the web as permanent got caught out by mobile. The ones treating mobile as permanent are about to get caught out by agents. Infrastructure that can follow the customer to whatever surface comes next is the whole game now.
