AI Commerce Weekly: Week 34, 2026
OpenAI is taking a bidding market inside ChatGPT to 31 European countries. On its own availability table, British advertisers can already buy into it and none of those 31 can.
TL;DR
W34 was the week the assistant grew a price list. OpenAI announced ChatGPT Ads reaching 31 European countries, and its own Ads Manager availability table says British advertisers can already buy through the self-service tool while every single one of those 31 markets is listed Coming Soon. Nine countries worldwide are Available. The UK is one of them, Ireland is not, Germany is not, France is not. That is a real commercial asymmetry with a golden quarter coming, and it is worth reading carefully rather than excitedly, because the table answers a question about advertiser access and not about whether anyone is seeing an ad. The rest of the week is the counterweight. Gartner's numbers, which Mastercard leans on while quietly reshaping them, put willingness to let AI make purchase decisions at a ceiling of eleven per cent, with 54% of AI-assisted shoppers saying they had to check everything they were told and 62% saying it wasted their time. Mastercard published a genuinely detailed specification for delegated payment. Storyblok's chief executive argued the discovery problem is a decade of contradictory content rather than missing plumbing. Tilt shipped an AI product whose whole pitch is keeping a person in the shot. And in Sheffield, somebody put traffic cones behind the delivery robots.
Nine countries, and Britain is one
On 18 August, OpenAI said it was bringing ChatGPT Ads to 31 European markets. The release names nine of them, Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands and Austria, with "including" doing the work of the other twenty-two. It gives no date beyond "next week". It does not mention the United Kingdom anywhere, in any form.
Last week that silence was the story, and this feed was deliberately careful about it: an omission is not an answer, and reading one into it is how you end up publishing something you have to take down. Then OpenAI's own Ads Manager availability table answered it directly. Forty countries listed. Nine marked Available: Australia, Brazil, Canada, Japan, Korea, Mexico, New Zealand, the United States, and the United Kingdom. Thirty-one marked Coming Soon, every one of them European, including all nine of the countries OpenAI named in the expansion release. OpenAI never publishes the full list of the 31, so strictly the two sets being identical is a reading rather than a stated fact. It is not a strained one.
So on the week a paid layer arrives across Europe, a British merchant can log into the buying tool and a German one cannot. Ireland cannot. Switzerland cannot.
Read the scope precisely, because it is very easy to overstate and I would rather be dull than wrong here. The table describes availability of Ads Manager, which is the advertiser-facing self-service tool. It says nothing about whether ads are being served to users in a given country, and it gives no UK go-live date. Two different questions; only the first is answered. There is a claim circulating on aggregator sites that the UK went live on 6 June, and it has now failed verification on four separate runs of this feed. No primary source states it. It stays out.
The consumer-side rules are published too, and they are more interesting than the geography. Ads may appear for users on the Free and Go plans; Plus, Pro, Business, Enterprise and Edu stay clean. No ads to accounts identified as belonging to under-18s. Ads sit below the end of a response, labelled and visually separated, and are not eligible near personal health, mental health or politics. OpenAI says it does not currently allow political advertising in ChatGPT at all. Not in Temporary Chats. Not in the Atlas browser during the test. For this early test, advertisers get aggregated reporting, views and clicks.
And then the line that will matter most in twelve months: "Personalized ads are not initially available in the European Economic Area (EEA) or Switzerland." The UK is not named there and is not in the EEA. The page says nothing either way about UK personalisation, so nothing is concluded from it here. But the shape is visible now, which is regulatory geography producing two different products inside one assistant, and Britain sitting outside the carve-out by default rather than by any stated decision.
Eleven per cent, and what it is a ceiling of
Here is the problem with all of that, and I do think it is a problem rather than a quibble.
The most-cited number in agentic commerce right now is Mastercard's "only one in ten" willing to let an agent complete a purchase autonomously. Follow the superscript on that sentence and it lands on a Gartner release from 27 May, and Gartner's own wording is different in a way that matters: willingness to let AI make purchase decisions "topped out at 11% across lower-stakes categories, such as personal care and household supplies". Topped out. It is a ceiling, measured where the stakes are lowest, from 322 US consumers surveyed in January 2026.
Rendered as a flat "one in ten", the qualifier disappears, the base disappears, and the geography disappears, inside a press release datelined Baku. Same direction of travel as the original, but tidier and more quotable than the evidence supports.
The fuller Gartner data is more useful than the headline anyway. Thirty-one per cent would let AI narrow choices for household supplies, twenty-eight per cent for personal electronics. People want a shortlist, not a proxy.
That distinction is doing more work than it looks like it is. A shortlist is a research tool and a proxy is a delegation, and almost every product roadmap in agentic commerce this year has been built for the second one.
The number I keep coming back to is from a different Gartner survey, 846 US consumers fielded across November and December 2025, so hold the two apart rather than stacking them. Among people who used AI while shopping for a recent purchase, 54% said they had to double-check the accuracy of all the information the tools gave them, and 62% said that information ended up wasting their time. Muhl's line on it is that accuracy is now a brand issue.
That verification burden is the mechanism underneath the gap. Everything else in this section only measures it. Nobody delegates to something they are already grading, and it is the bit a bidding market does not fix: you can win the placement and still lose the sale to a shopper who opens three tabs to check what the assistant just told them about your delivery window.
All of these figures are American. We still do not have a UK-sampled equivalent, which is a gap this publication has flagged for months and will keep flagging. The closest domestic reading we have is the Vercel and World Retail Congress work from July, where 20% of UK shoppers said they start on AI search and 79% still start on the retailer's own site.
Mastercard writes the spec
Set the figures aside, because the architecture in the Mastercard Signals report is the most concrete published answer yet to a question this feed spent last week asking: what does anyone's human-in-the-loop control actually consist of, in operational terms?
Mastercard's version is a trust layer that sits between the agent's decision and the money moving. Before a payment executes, it verifies who is acting, what the user authorised and which limits apply, and keeps an auditable record of what the agent did. Agentic tokens carry task-specific authority and can be restricted by agent, merchant, category, spending limit, timeframe or usage rules. And there is a defined exception path, where an unexpected price change or a renewal is paused and brought back for approval rather than pushed through.
That is a specification. Alert-level, testable, arguable. It is the sort of thing you could actually put in a contract, which is more than can be said for most of the "the human stays in control" language sold this year. Apply the Sunday-evening test to it and it mostly holds up, which is not a sentence I write often about a vendor report.
Two honest caveats, though, and they run in opposite directions.
The first is that the answer arrived from the payments layer and not from a retailer. The question was whether any retailer would publish what its review process consists of. Still nobody. The scheme has written the spec that its customers have not.
The second is the report's own evidential hygiene. The two optimistic figures on the page, 85% open to collaborating with an agent and 74% open to letting one complete specific tasks, carry no base, no sample, no fieldwork date and no market at all. The one pessimistic figure is the only traceable one. And the $3–5 trillion by 2030 forecast sits there with no attribution whatsoever. It is a real number from McKinsey, as it happens, correctly credited over in Rain's Agentic Payments Alliance launch the same month. Just not credited here.
That Alliance is worth a paragraph on its own terms. Rain convened more than 25 organisations, Visa, Mastercard, Fiserv, Circle, Solana and Remitly among them, to work on agent authorisation, fraud prevention, and whether loyalty and rewards travel with an agent. It says it will be run collectively by its founding members rather than owned by any one company, and that members will set the charter together. Which is another way of saying there is no charter, no governance model, no funding detail and no timetable yet. Worth watching, worth not treating as an institution.
Read the roster with a UK eye and the thing that stands out is who is absent. No retailers, of any nationality. No regulators. The release gives no head-office locations, so nobody can say from it whether any UK institution is inside the room. What can be said is that the rails for agent authorisation are being specified in New York by payment companies, while the questions about who is liable when an agent buys the wrong thing are going to be answered by British regulators who are not on the list.
Delete it
The organic half of the discovery question got an unusually practical answer this week, from a slightly unexpected direction. The short version is that the binding constraint might not be missing plumbing. It might be five years or more of contradictory content that the plumbing will faithfully surface.
Retail Gazette put the machine-readability argument to a UK retail audience via Dominik Angerer, CEO and co-founder of the CMS vendor Storyblok. The headline figure is that 46% of buyer journeys already begin in AI rather than traditional search. Handle it with tongs: that is Storyblok's own research, shared with the publication, published on 3 November 2025, with no sample size, no methodology, no fieldwork dates and no market stated anywhere. Nine months old, vendor-supplied, and in a UK title with no UK breakdown. Angerer's own gloss, reported rather than quoted, is that it is likely higher now.
The argument underneath it survives the figure, which is why it is here.
Angerer's advice is not to generate more machine-readable content. He tells retailers to find out where their content actually lives, across the websites and markets and languages and help desks and FAQs a large retailer accumulates, and then to update, redirect or delete anything old or contradictory. On the deleting:
"I really mean delete it because wrong content is way more harmful than less content"
And the bit most GEO coverage misses entirely: "If you remove all your content of the websites that are publicly facing, but you still have the old PDFs online, you will still have the same issue."
Anyone who has done a platform migration knows exactly how true that is. I have personally left markup on live sites that outlived the company that commissioned it. Old PDFs, orphaned landing pages, the 2019 delivery policy nobody unpublished because it was on a subdomain nobody owned. Under the old regime that was untidy and mostly harmless, because a search engine ranked it into oblivion. A model does not rank it into oblivion. It reads it and repeats it, confidently, to someone asking whether you deliver to Orkney.
So the thing being sold as AI readiness turns out to be housekeeping. Less saleable, considerably more actionable for a mid-market retailer without a commercetools budget, and it connects straight back to the verification burden above. Every wrong fact a model surfaces about your catalogue is another tab the shopper opens to check you.
Read the source with its interest in mind, obviously. The diagnosis is a content problem and the cure is content management, sold by a content management company.
The traffic cones
Three things happened at the physical end of the week, and they rhyme.
Iceland says AI over its existing CCTV has cut losses from theft and shrink by 80%. SAI's platform runs computer vision across checkouts, aisles and entrances, spotting missed scans, concealment, walkouts and shelf sweeping, and pushes an alert to a colleague's handheld within eight seconds. False alerts down almost 50%, colleagues responding to 90% of alerts. Iceland runs more than 1,000 UK stores, around 700 of them on the high street, and its hundred worst account for roughly a third of total shrink against a theft bill previously put at around £20m a year. Andy Edwards, head of loss prevention and corporate audit, says they needed "a solution that could work alongside us, rather than behind us", which is the right instinct. Every one of those figures is retailer-and-vendor sourced with no methodology, no measurement window and no baseline, and the coverage is given only as "across its store estate", never quantified. That is the wrong habit. Iceland is now exploring facial recognition, and SAI has published safeguards in advance: no retention of innocent shoppers' images, a customer right to review the evidence, proportionate sanctions. Stated intent for a deployment nobody has announced, but stated in advance, which is more than the fortnight's other biometrics stories managed.
Tilt launched Clips, a scrollable feed of short buyable videos where the AI handles the shoot and the write-up and a listing takes under three seconds. The UK live-shopping app calls it its biggest launch since raising $26 million from the likes of TQ Ventures, Vinted Ventures and Balderton. What is interesting is that it is an AI product whose pitch is a limit on AI: the machine writes the listing, the person still has to hold the shoe up.
It shipped without a single adoption or conversion figure. Tilt's earlier listing tool, which this feed reads as the predecessor although the announcement never says so, did carry one. File the argument and wait for the evidence.
And then Sheffield. Starship has pulled its delivery robots out of the city after five months, having reached around 50,000 households since a March start with Uber Eats. Both halves of this need reporting together. There was vandalism: visibility flags snapped off, "off our street" daubed across the machines, traffic cones set behind them to stop them moving. And Starship says none of it drove the decision, calling the withdrawal commercial and the "natural conclusion" of the pilot. Separately, the walking charity Living Streets has criticised Co-op's use of the robots over their effect on older and disabled pedestrians, and is calling for clearer regulation of delivery robots on UK pavements.
Nobody has reported a technical failure in Sheffield. Somebody wrote on the robots.
My read is that the constraint on all three of these is the same one, and it is not capability. Iceland is doing well because it is publishing what happens when the system is wrong. Tilt is selling the person in the frame. Starship's pilot ended in a city where a chunk of the public treated the machine as an imposition rather than a service, whatever the commercial rationale sitting on top of that. Permission is the scarce input here, and it does not come from a better model.
Which brings the week back round to where it started, because a bidding market on a surface where 54% of people are already double-checking the answers is a market for attention rather than for conversion. Attention is worth buying. It is also a different budget line, a different owner and a different measure of success than the one the deck will imply.
I will be at eCommerce Expo at ExCeL on 23 and 24 September, and this is the question I am taking with me. Hilary Platt, a Director at OpenAI, is on the programme, alongside Deann Evans, Managing Director, EMEA at Shopify, Jordy Jordan, Director of Product Management at Asos, and Duncan Rutherford, Head of eCommerce at BT. Ten thousand people are expected across the two days. A platform director on a UK merchant stage in late September is the platform side turning up in person rather than through a partner's press release, and there is one obvious thing to ask her: when do the ads actually start showing here.
What to do this week
Sources
Nine countries, and Britain is one
The 18 August release gives no go-live date; 'next week' is its only temporal marker, so the week commencing 24 August is an inference. Its nine named countries are prefaced with 'including' and are not the full 31, which OpenAI never lists, so the identity of the 31 Coming Soon rows with the 31 expansion markets is a reading rather than a stated fact. Neither 'United Kingdom' nor 'UK' appears anywhere in the release or in the consumer-side FAQ. Both Help Centre pages carry rolling relative timestamps only and no absolute publication date, so they are cited by access date rather than by a converted one. The availability table describes advertiser access to Ads Manager, not whether ads are served to users in a country, and 'Available' is nowhere defined. The claim that the UK went live on 6 June 2026 has failed verification on four runs of this feed and has no primary source.
Eleven per cent, and what it is a ceiling of
Two separate surveys, deliberately not combined here. The 11%, 31% and 28% figures are from 322 US consumers surveyed January 2026; the 54% and 62% figures are from 846 US consumers surveyed November to December 2025. All US-sampled. The 11% is a category ceiling ('topped out at'), not a level. No UK-sampled equivalent for any of these figures exists in this corpus.
Mastercard writes the spec
The Mastercard release is datelined 'Baku, Azerbaijan — August 2026' with no day of the month. Its 'one in ten' figure carries a bare superscript hyperlink to the Gartner release with no printed footnote text, so the base, the geography and the 'topped out at' qualifier are invisible without following it. The 85% and 74% figures carry no base, sample, fieldwork date or market. The $3–5 trillion forecast carries no attribution on the Mastercard page; the McKinsey credit comes from the Rain release. Rain's release says 'Over 25 industry leaders', not a specific count, and Rain has a direct commercial interest. No charter, governance model, funding or timetable is published. No member HQ locations are given, so no claim about UK representation is made beyond the observable absence of any retailer or regulator from the roster.
Delete it
The 46% figure is Storyblok's own research, shared with Retail Gazette and published on 3 November 2025, with no sample size, methodology, fieldwork dates, respondent definition or market stated, and no UK breakdown. 'Likely to be higher now' is the publication's paraphrase of Angerer, not a quotation. His full advice is to 'update, redirect or delete', not delete only. Storyblok is a CMS vendor and the remedy described is content management.
The traffic cones
Every Iceland figure is retailer-and-vendor sourced with no published methodology, measurement window or baseline, and the share of the estate covered is never quantified. SAI's facial-recognition safeguards are stated vendor intent for a deployment that has not been announced. The Register listed Iceland among existing Facewatch retailers on 17 August, which sits unresolved against 'exploring facial recognition'; no conclusion is drawn here. Clips shipped with no adoption, sell-through or conversion figures published. Its lineage from Tilt's earlier Snap product is this feed's reading of a link in the source, not a claim the source makes. On Starship, the documented vandalism and the company's denial that it drove the decision are both on the record and are reported together. Both RTIH pages carry a byline date of 20 August against a 19 August URL slug; the slug date is used. eCommerce Expo speaker titles are given as the source renders them.
- Iceland cuts store losses by 80% with AI-powered theft detection, Retail Gazette, 19 August 2026
- Live shopping specialist Tilt lays claim to its biggest launch since securing $26 million funding round, RTIH, 19 August 2026
- Starship Technologies delivery robots pulled from Sheffield after five-month trial, Retail Gazette, 24 August 2026
- eCommerce Expo announces full line-up for 2026 show taking place at London ExCeL next month, RTIH, 19 August 2026
Stay Connected
Follow LLCommerce on LinkedIn
Get the latest AI commerce insights, analysis, and industry news delivered to your feed.
Large Language CommerceAuthors
