AI Commerce Weekly: Week 29, 2026
AI chatbots get a fact wrong about 64% of UK high-street retailers. That, the shelf robots, and Gap's 'mindset beats mandate' study defined the week.
TL;DR
W29 read like a stack, and every layer was doing the same job for the one above it. At the bottom sat the product-data substrate, and Searchable priced the cost of getting it wrong: the AI chatbots British shoppers actually use returned a false fact about 64% of UK high-street retailers. Above it the store hardened around what is on the shelf, Tesco trialling robots while Instacart bought the computer vision to run on 600,000 phones, and SmartNovo revealed a store that is itself the AI. Newest, and most human, the people: Gap and Microsoft found that mindset beats mandate, and that the AI hangover is real. Capability was never the constraint.
What the chatbots get wrong about you
The number to carry out of this week is unflattering, and it is ours. Searchable, a London-built AI-visibility platform, put more than 72,000 questions to ChatGPT, Gemini and Perplexity about UK high-street retailers, then graded every answer against the truth. Sixty-four per cent of businesses had at least one false fact returned about them. One answer in sixteen was simply wrong. The commonest error was the wrong postcode, one in ten, and in the worst cases the address was more than twenty miles out. Around one in fifteen website answers pointed at a dead link, a lookalike, or somebody else's shop entirely.
The failure is not spread evenly across the tools, which is the part that earns a chart.
Perplexity was wrong more than twice as often as ChatGPT. If you have been optimising for one and ignoring the other, that gap is your blind spot.
I keep coming back to what this actually means for a retailer, because it inverts the usual framing. We spend the discovery budget on being found. Searchable is measuring something one step earlier and more dangerous: being found and then described wrongly, by a machine the shopper has decided to trust. Searchable cites a Rithium survey in which 58% of AI-using shoppers say they lose trust in a brand when the AI gets its product information wrong. And you do not get told when it happens. There is no bounce-rate line for "the assistant sent them to a rival because it had your closing time wrong." Sarah has argued for a while that consumer trust is a UI problem before it is a data-policy one, and this is that same argument moved upstream: the interface the customer now trusts is one you do not own and cannot see, and it is fumbling your basics one time in sixteen. Co-founder Chris Donnelly notes it hits smaller bricks-and-mortar retailers hardest, because a thin footprint gives the models little accurate material to learn from. The fix is unglamorous and cheap. Your website, your directory listings, your structured data. The same substrate everything else this week was built on.
Making the catalogue legible to the machine
If Searchable is the substrate seen from the loss column, two builds this week showed the supply side of the same problem. Both start from the premise that product data is now the thing an agent reads before it decides whether to recommend you.
Akeneo used its Summer Release to introduce Agentic Ziggy, an agentic layer inside its Product Cloud that lets a product-data team deploy and supervise fleets of AI agents to enrich, govern and check catalogue data across millions of SKUs. The bit that matters for anyone who has run a PIM in anger is the governance model: propose-and-approve, human-in-the-loop, role-based permissions, approval gates. Agents do the tedious work of schema mapping and syndication-error triage, and a person still signs off. That is the right shape. The failure mode of "let the agents fix the catalogue" is obvious to anyone who has watched an automated feed cheerfully propagate the same error across forty channels at three in the morning. A named early user is Elemis, the UK skincare brand, which is a more useful signal than the launch copy.
The consumer-facing cousin came from Meta. Meta told brands they can use its Muse Image tool to drop catalogue products into a photo of a shopper's own room inside Meta AI, compare options, then buy through the brand's own site, all off the product data Meta already holds for ads. It is US-first, and Meta is refreshingly blunt about the incentive: the brands that invest in catalogue quality now will be best placed as AI discovery scales. Different surface, same toll booth. Whether the picture is rendered in a chatbot's answer, a product-experience agent's enrichment, or a room visualiser inside Meta AI, the entry ticket is a clean, structured, machine-readable catalogue.
There was a fourth company at it this week, from the production end rather than the data end. ABOUT YOU Group launched SCAYLE STUDIOS, an AI photo studio that turns a single product image into a full set of PDP, social and campaign shots and videos, replacing fixed studio slots and weeks of lead time with a self-service tool. Inside ABOUT YOU it now runs about 90% of e-commerce productions at roughly 90% lower cost, and in the company's own A/B tests the AI-generated images beat traditional studio photography by 9.2% on GMV. It is German rather than British, and the named early users (s.Oliver, Betty Barclay, Goldner) are too, but it is available to UK brands and it makes the supply-side version of the same point: the catalogue is what the machine reads, and now increasingly what the machine helps make. If your roadmap still files product content under back-office chore, this was the week four separate companies told you it is the storefront.
The store hardens around what's on the shelf
Underneath the discovery layer, the physical store spent the week obsessing over one deeply unglamorous number: whether the thing is actually on the shelf.
Tesco is trialling Simbe's Tally robot, a five-foot autonomous unit that trundles the aisles scanning for gaps, misplaced stock and wrong shelf-edge prices, then hands the list to store colleagues. It follows Morrisons, which trialled the same robot in three stores last year, so this is now the UK's biggest grocer following its rival into shelf-scanning hardware. Simbe's global figures (ten times more out-of-stocks caught than manual audits, 607 million unavailable-product instances found worldwide) are vendor claims, so hold them at arm's length, but the direction is real enough.
Instacart went at the same problem from the other end and bought the computer vision instead of building the robot, acquiring the Israeli firm Arpalus, whose models turn a quick phone video of a shelf into real-time inventory at better than 95% accuracy. Because it runs on any camera, Instacart plans to turn its network of roughly 600,000 shoppers into a live shelf-sensing layer, then wire the result straight into e-commerce fulfilment rather than a store-ops dashboard. That last part is the tell. Availability is quietly becoming an input to the online order and the AI shopping experience, not just a shop-floor metric. For a UK grocer the live question is which capture model you can run at scale: dedicated robots down the aisle, or vision on the phones your pickers already carry.
And then there is the logical extreme. SmartNovo took the wraps off a store that is itself the AI: a 21-square-metre, staffless, container-sized unit launching in Poland under the Lewiatan brand with NVIDIA and Starlink, where, in the founder's words, the model is the manager, the cashier, the guard and the marketing team all at once. The whole box loads onto a truck and relocates in under an hour, and it is legally classified as a vending machine, which sidesteps permits and construction. I would not over-read a reveal, and the theft-detection and dynamic-pricing claims are the founder's own. But the unit economics of a relocatable, AI-run box are different enough to matter, and if you run small-format or pop-up estate it is worth understanding before a competitor pilots one.
Better products, not more products
The one customer-facing AI win of the week was British, and it was the kind of fashion-adjacent story that matters most to this desk. DCYPHER Beauty is launching exclusively inside John Lewis, starting at the Oxford Street flagship in August and rolling across the estate next spring, with patented AI that reads a customer's skin in-store and formulates a made-to-measure complexion product for them within minutes. No pre-made shades, no approximation.
What makes it interesting is not the try-on novelty, it is the inventory logic underneath. Co-founder Ruth De Leo puts it in a line I have not been able to stop turning over.
Read that as a technology strategy rather than a marketing line and it is quietly radical. Almost every AI personalisation pitch is really a recommendation engine bolted to the same warehouse: show more, cross-sell more, hold more stock to cover more variants. DCYPHER's model uses personalisation to make the inventory smaller, formulating to order rather than guessing demand across forty SKUs. For fashion and beauty specifically, where the industry is drowning in size-and-shade permutations and the returns that come with them, made-to-order is the version of AI personalisation with an actual operational payoff. It is one launch in one department store and I am not going to over-claim from it. But it is the one this week I would keep a tab open on.
The AI that ships where the shopper can't see it
Here is the pattern I actually take from the week, wearing the Head-of-Technology hat. The most credible UK AI spend right now is not the shopper-facing stuff. It is in the back office, out of sight, where the returns are easier to prove and nobody has to trust a chatbot.
The Very Group signed a three-year deal with UiPath to run agentic AI across pricing and merchandising for more than 200,000 products. The Liverpool retailer led the announcement on explainability, which is the right instinct: this is agentic AI making margin-critical decisions, and "the model said so" does not survive a buying-team review or, eventually, a regulator. It is not glamorous. It is also exactly the kind of deployment that works, because pricing is a bounded, data-rich, measurable problem, and the trust barrier that hobbles a consumer-facing agent simply does not apply when the user is your own commercial team.
Dotdigital, meanwhile, shipped an MCP server as a standard connector, alongside a loyalty product and a natural-language builder. The loyalty module is the crowded part and I would not lose sleep over it. The MCP server is the signal. It is the mirror image of the substrate story: agentic discovery is about making your product data legible to someone else's AI, and an MCP server is about making your own stack callable from inside an LLM like ChatGPT or Claude. A UK martech name shipping that as a checkbox connector, not a moonshot, tells you where the plumbing is heading. Dotdigital's own research says only 15% of consumers find the marketing they get "very relevant", which is a fair argument for making your stack easier to query and act on rather than harder.
Mindset beats mandate
The most useful thing on the wire this week looked past the storefront entirely, at the people meant to use all of this. Gap Inc. and Microsoft ran a field study of 388 Gap employees and came back with two findings that every retailer rolling AI out internally should pin up.
The first they called mindset beats mandate.
Teams forced into a rigid AI workflow did worse than teams left to figure it out for themselves: more friction, lower quality, more work left unfinished. Reframing AI as a thought partner rather than a tool produced meaningfully better work at the exact same level of access. The second finding is the one I have felt in my own building and never had a name for. There is such a thing as an AI hangover. People who struggled through a forced, joyless rollout carried the frustration into the next task, and it measurably lowered how they felt about the tools afterwards. A bad mandate does not just fail on its own terms. It sours the next attempt too.
Worth being honest about the dating: the fieldwork ran last November and the study first surfaced in April, so this is fresh coverage of not-new research. It earns its place anyway, because it is the half of AI adoption this feed almost never captures, and because of where it comes from. Old Navy, Banana Republic and Athleta are Gap brands, so this is a fashion-retail workforce lesson, not a generic one. Every retailer reading this is rolling AI out internally as well as bolting it onto the storefront, and most of us are measuring the storefront half far more carefully than the human one. The uncomfortable read across the whole week is that the technology was never the constraint. The catalogue an agent can trust, the shelf data that is actually right, the pricing you can explain, the people you brought along rather than dragooned. That is the work. The models are the easy part.
Sources
What the chatbots get wrong about you
Searchable is a commercial AI-visibility platform reporting its own 72,000-question study; the 58% trust figure is a cited 2026 Rithium consumer survey.
Making the catalogue legible to the machine
Akeneo's and Meta's capability descriptions are the vendors' own; Meta Muse Image is US-first, with UK/EU availability unconfirmed. SCAYLE STUDIOS (ABOUT YOU Group) is German in origin and its efficiency and 9.2% GMV figures are the company's own A/B tests; added post-publish on 2026-08-03 as a W29-window (20 July) development, and covered in depth in a standalone article.
- Akeneo introduces the first truly agentic product experience platform, The Retail Bulletin, 10 July 2026
- Akeneo Introduces the First Truly Agentic Product Experience Platform (primary), Akeneo newsroom, July 2026
- Meta introduces AI-powered room visualization feature (Muse Image), Retail Dive, 9 July 2026
- Introducing Muse Image (primary), Meta newsroom, July 2026
- ABOUT YOU Group launches AI-powered photo studio with SCAYLE STUDIOS (primary), ABOUT YOU newsroom, 20 July 2026
- About You launches Scayle digital AI studio to speed up photo/video production for UK brands, FashionNetwork UK, 20 July 2026
The store hardens around what's on the shelf
Simbe's 10x / 607m out-of-stock figures and Arpalus's 95% accuracy are vendor figures; SmartNovo is a reveal, and its autonomous-store capability claims rest on the founder's account. SmartNovo is CEE, not UK.
- Tesco trials shelf-scanning robots in supermarket aisles (Simbe Tally), Retail Gazette, 15 July 2026
- Instacart Acquires Arpalus to Advance Real-Time Shelf Intelligence Across Grocery Retail (primary), Instacart Newsroom, 16 July 2026
- SmartNovo takes the wraps off autonomous store as company eyes expansion across Poland and CEE, RTIH, 12 July 2026
Better products, not more products
DCYPHER's 'first beauty brand globally' and made-to-measure claims are the brand's own.
The AI that ships where the shopper can't see it
The Very Group's pricing terms were undisclosed; Dotdigital's 15% / 53% figures are its Censuswide Customer Trends Index (base 4,000, UK/US/AU/SG).
Mindset beats mandate
A readout of an internal Gap Inc./Microsoft field study, first published on the Gap Inc. newsroom in April 2026; fieldwork ran November 2025, so the mid-July RTIH coverage is fresh but the underlying research is not new.
Stay Connected
Follow LLCommerce on LinkedIn
Get the latest AI commerce insights, analysis, and industry news delivered to your feed.
Large Language CommerceAuthors
