For two years, the line on AI shopping has been the same: one day, agents will do the buying for us. It has always been one day. Never this week.
Then OpenAI launched ChatGPT Wallet, and agents can now actually pay. In the same week, SparkToro found that 93.1% of ChatGPT conversations showed no intention to purchase, and HUMAN Security found that AI agents mostly browse product pages and rarely reach the checkout.
So which is it? Both. Agents are arriving on your website in growing numbers, they now have a way to pay, and most of them are still window shopping.
That makes this the cheapest moment you will ever have to get ready. The brands that win the agent era won’t be the ones that shout loudest about AI. They will be the ones machines can read, trust and get through to checkout.
Here is what mattered, why it mattered, and what to do with it on Monday.
Prefer this in your inbox? Board Analysis, our newsletter for UK marketing leaders, goes out every Friday with the week’s most useful marketing updates and what to do about them. Subscribe here.
OpenAI has launched ChatGPT Wallet, which it describes as “the first wallet that lets agents pay without access to users’ actual card details and that requires a security check for every purchase.” It was built with American Express, Chase, Mastercard and Visa, and will initially be available to ChatGPT Pro subscribers in the US. OpenAI says “1.2B people use ChatGPT each week, and many of them turn to it to research products, compare options, and decide what to buy.” (OpenAI)
The US-only launch is easy to dismiss from the UK. The playbook that came with it is not. OpenAI says plainly that “some sites may unintentionally block purchases through ChatGPT Wallet or other agentic payment products”, and its Commerce Partner Playbook gives businesses a to-do list: claim your business in ChatGPT, upload a product feed, let verified agents through your firewall and work with your payment and risk providers. None of that depends on when Wallet reaches the UK, and all of it takes time.
Read more: Introducing ChatGPT Wallet, OpenAI
SparkToro analysed 41,961 ChatGPT conversations from a US opt-in panel and found purchase intent in just 6.9% of them. “In a total of 4.9% of conversations, the person planned to purchase and was continuing to look at the options; another 2.1% were asking questions related to purchasing a specific brand, product or service; while a total of 93.1% showed no intention to purchase.” And “of ChatGPT’s responses in this sub-sample, 87.9% did not name a single brand, product or business.” (Search Engine Journal)
This is the reality check to sit beside the Wallet launch. Agents may soon buy, but today most people use ChatGPT for images, information and writing, not shopping. SparkToro’s Rand Fishkin concluded that “the primary reason AI visibility matters is Google’s implementation of AI Overviews and AI Mode, rather than ChatGPT itself.” Getting machine-readable is the right move. Pouring budget into ChatGPT on the assumption that everyone is shopping there is not.
Read more: 6.9% Of ChatGPT Conversations Showed Purchase Intent In U.S. Panel, Search Engine Journal
LinkedIn has brought its CTV (connected TV, meaning streaming on smart TVs) ads to the UK, letting advertisers reach decision-makers “by job role, industry and seniority while they watch ad-supported streaming content.” It has partnered with LG and Titan OS, and the product is “available to all advertisers in the market.” LinkedIn claims “people exposed to a brand’s CTV ad cost 30% less to convert into sales leads.” (Digiday)
For B2B marketers, this is one of the few ways to buy TV-style reach with professional targeting attached. It fits this week’s theme too: the value is in the audience data being accurate and verified. But the figures come from LinkedIn and one Cisco case study, and the questions about cost and accuracy are fair. Treat it as a test with measured incrementality (what the ads added that wouldn’t have happened anyway), not a channel to scale on day one.
Read more: LinkedIn brings its B2B ad targeting to UK connected TVs, Digiday
“Websites running Optimizely Web Experimentation served malicious JavaScript to visitors on Nov. 16 after an attacker added it to their projects inside Optimizely.” Report URI found the code in snippets belonging to 178 organisations, including Zoom, the BBC, CNN, Salesforce and GitLab. It swapped crypto wallet addresses and looked for card details typed into sites’ own forms. “Out of the total of 159 organizations that were reviewed, 68 had no Content Security Policy at all and 50 did have a Content Security Policy that permitted both the malicious code and transmission of sensitive user information.” (Search Engine Journal)
This is the other side of being machine-ready. The more tools you plug into your site, from A/B testing to tag managers to chat widgets, the more people can change what your visitors see. A Content Security Policy (a browser instruction listing which scripts your site is allowed to run) only helped where it was strict. Trusting a vendor’s domain did nothing, because the malicious code came from that same trusted domain.
Read more: Optimizely Says Attacker Used A Signup Flaw To Inject Malware, Search Engine Journal
“Unilever reported that it now relies on 500,000 content creators, up from the previously reported 300,000 and significantly more than the 20,000 utilized before CEO Fernando Fernandez joined the company.” It says it is “producing twelve times more content than in 2024 at a 90% lower cost”, and that creator content yields ROI that is “1.5-times higher than brand-led content.” (Marketing Dive)
The detail that should worry every brand team is the shelf life. Unilever says AI is one of the factors “reducing the lifespan of video content to four days”, after which returns turn negative. Fernandez also set out what AI recommendation engines reward: “geographical proximity, independent recommendation, clinical proof, patents published, ratings and review.” In other words, the same trust signals that machines can read. Volume keeps you in the feed. Proof gets you recommended.
Read more: How Unilever’s social-first marketing is evolving amid reorganization, Marketing Dive
“AI agents from Meta and OpenAI made nearly 90% of the agent requests reaching sites and apps protected by HUMAN Security in October”, and “Meta’s Muse accounted for 59% on its own.” Total agent requests in October “were approximately 8 times greater than in August”. Yet “85% of all agent requests in the report were directed toward product and search pages”, while requests to checkout and payment pages “decreased from 2.4% of the total to 1.9%.” (Search Engine Journal)
Put this beside ChatGPT Wallet and the picture is clear. Agents are already on your product pages in growing numbers, reading your prices, specs and stock levels. Payment tools like Wallet are designed to turn that browsing into buying. Whether it turns into your revenue depends on whether your product data is accurate and your site lets them through.
Read more: Meta & OpenAI Agents Made Nearly 90% Of Agent Requests In Bot Data, Search Engine Journal
OpenAI’s playbook asks for accurate product feeds. Search engines and AI assistants also read your structured data, the code on your product pages that spells out what you sell, what it costs and whether it’s in stock. In our guide to structured data for eCommerce, Helen Halfpenny explains that “Crawlers for engines like Perplexity, Gemini, and ChatGPT read JSON-LD to confirm real-time availability, product specs, and prices before recommending products in AI answers.” She also warns that “Search engines will cross check the info you provide, and if it doesn’t match, you might find your product listings getting disapproved or suspended.” If you do one thing before agents start paying, make sure your feed and your schema are saying the same thing.
TwentyCi provides UK residential property data and analytics. Their brief was refreshingly blunt: no vanity metrics, only conversions that support business goals. We built a content calendar and SEO-optimised copy backed by keyword research, used sharp PPC targeting and remarketing to secure quality leads, and used email marketing to share industry knowledge and keep relationships warm.
“We have definitely seen an upturn in the quality and volume of incoming enquiries. The PPC, SEO & content strategy is achieving cut through with the right audiences.” Nick McConnell, TwentyCi
With agents now doing more of the browsing, chasing the quality of what converts rather than the volume of what visits is exactly the right instinct. See how we did it.
Every story this week came back to one question: can machines read you, trust you and get through to you? ChatGPT Wallet gave agents a way to pay, but only on sites that let them in. HUMAN Security showed agents are already reading your product pages. SparkToro reminded us that most ChatGPT use still isn’t shopping, and that Google’s AI features matter more for now. Unilever showed that AI recommendation engines reward proof: reviews, ratings and independent recommendation. And Optimizely showed what happens when the tools you trust become the weak point.
None of this needs a big AI budget. It needs accurate product data, structured data that matches it, a website that lets the right bots in and keeps the wrong code out, and a brand with enough proof behind it to be recommended. Those are the foundations, and they pay back whether the agents arrive next month or next year.
Whether it’s getting your product data and structured data ready for AI agents, deciding how much ChatGPT deserves in your media plan, or locking down the scripts running on your website, our team can help you turn this week’s changes into a plan. We combine SEO, PPC and web development so your marketing works as one joined-up approach, with every decision tied back to revenue.
Sign up to The Digital Maze Newsletter