8 min to read
Hi clients, readers, fans, haters, I am back to writing to all of you after short holidays and so much has happened. I don't even know where to start.
Well considering the topic I am writing about here today, two things: happened in the last six months that should be sitting on the same page in every commerce director's notebook. They almost never are.
The first. OpenAI killed Instant Checkout in March, roughly six months after launching it with Stripe and a lot of fanfare. Their own explanation was that the first version "did not offer the level of flexibility we aspire to provide," so merchants went back to owning their own checkout while OpenAI refocused on product discovery. A lot of people read that as agentic commerce failing.
The second. Google shipped Universal Cart at I/O in May and started rolling it out across Search and Gemini in the US over the summer. Nike, Sephora, Target, Walmart, Wayfair, plus Shopify merchants like Fenty and Steve Madden. The Universal Commerce Protocol, which Google launched at NRF in January with Shopify and a stack of payment partners, now has simplified onboarding directly through Merchant Center.
So agentic commerce didn't fail. It changed hands.
And while everyone was writing think pieces about whether OpenAI's retreat meant the whole thing was overhyped, the actual infrastructure got built somewhere else, on a protocol that reads from a file most brands treat as an IT chore.
Your product feed.
The part nobody wants to hear
I've been running paid media for eighteen years. I've watched a lot of channels arrive with a lot of noise. This one is different in a specific and uncomfortable way, and it took me a while to articulate why.
Every previous shift was a shift in where the customer was. Search to social. Desktop to mobile. Display to video. Annoying, expensive, survivable. You moved budget, you learned a new auction, you carried on.
This is a shift in who the customer is.
When a shopper asks Gemini for waterproof hiking boots under a hundred euros that will fit a wide foot and arrive before Friday, the thing evaluating your product is not a person. It's a model reading structured data. It will never load your homepage. It will never see the hero image your creative team fought about for three weeks. It will never encounter your brand film, your founder story, your carefully art-directed lookbook.
It reads your feed. That's it.
Estimates put consumer adoption of agentic shopping going from around 19% to 46% by the end of this year. Even if that's optimistic by half, it's the fastest behavioural shift I've seen since mobile. And 71% of marketing leaders already say agents have weakened their ability to connect directly with customers, which tells you the disintermediation is being felt before it's being measured.
Why this is a 60-day problem, not a 2027 problem
I want to be precise here, because "the future of commerce" articles are cheap and I've written a few I'm not proud of.
Three things make this urgent right now rather than next planning cycle.
Universal Cart went live in Search and Gemini in the US this summer. Not announced. Live. If you sell into the US market and your feed isn't UCP-ready through Merchant Center, you're not in the consideration set for a growing share of high-intent queries, and there is no bid you can place to fix that.
Second, Google introduced Conversational Attributes in Merchant Center at Marketing Live. Six new optional fields, including document links to manuals and spec sheets and ingredient lists, related products for accessories and required parts and substitutes, item group titles that separate the family name from the SKU-level title, structured variant options, and popularity rank within your own catalogue. Optional is the key word. They don't affect product approval and they don't change your standard Shopping ads. Which means most merchants are ignoring them completely. That gap is the entire opportunity, and it will close.
Third, sub-hourly price and inventory updates are now the baseline for AI Max for Shopping, Universal Cart, and Direct Offers. If your feed refreshes overnight, an agent trying to assemble a cart across four retailers will route around you rather than risk a broken transaction. Speed isn't a nice-to-have in this layer. It's an eligibility requirement.
None of that requires a strategy offsite. It requires a feed engineer and about six weeks.
What happened with a client in Germany
We work with a Portuguese footwear manufacturer that exports into the German market. Family business, genuinely excellent product, terrible at telling anyone about it. Classic Iberian export story.
Their German paid search performance had been degrading for about four quarters. CPCs creeping, conversion rate flat, the usual diagnosis of "the market is more competitive now." We ran the standard playbook and got the standard incremental improvement. Better creative, tighter feeds, smarter bidding. Roughly 9% better on ROAS. Fine. Not interesting.
Then we ran their catalogue through our own LLM visibility tracking to see what the models were actually saying when a German shopper asked for durable leather boots made in Europe. The answer was that our client didn't exist. Not badly positioned. Absent. The models were citing three competitors, one of which is materially worse on product and considerably better on structured data.
We looked at why. Their feed had titles, prices, images, GTINs. Everything Google's approval process wanted. Nothing a model could reason with. No material composition in a machine-readable field. No width fitting. No care documentation. No relationship between the boot and the replacement laces and the leather conditioner. No indication of which SKUs were actually popular. A German shopper asking a nuanced question got a nuanced answer that had nothing to do with our client, because there was nothing in the data to build a nuanced answer from.
We spent seven weeks rebuilding the feed. Material composition, fitting data, care documents attached as document links, related-product relationships mapped across the whole catalogue, popularity rank populated from their own sales data. Boring work. No creative awards for this.
Citation share in German-language product queries went from effectively zero to appearing in roughly a third of the prompt set we track. Traffic from AI surfaces is still small in absolute terms. But it converts at a rate that made the commercial director ask me to check the tracking twice, which matches what the wider data shows. LLM referral traffic converts at around 15.9% for ChatGPT against 1.76% for Google organic, and AI-referred visitors view roughly twice as many pages.
Small numbers, extraordinary quality. That's the shape of this channel right now. It's what a channel looks like eighteen months before everyone is complaining about how expensive it is.
The thing I keep coming back to
I reread Byron Sharp's How Brands Grow over the summer, mostly because I wanted to argue with it. Sharp's central claim is that brands grow through mental availability and physical availability. Be easy to think of, be easy to buy.
I think we now need a third one. Machine availability. Be easy for a model to reason about.
And the awkward implication is that machine availability is not downstream of the other two. A brand can have enormous mental availability, the kind built over decades with beautiful advertising, and still be invisible to an agent because its structured data is thin. Meanwhile a competitor nobody has heard of, with a well-built feed and honest specifications, gets recommended by name.
That should bother anyone who has spent a career building brands. It bothers me.
On the Voice of Experts episode I recorded with Marcus Lancaster from Quad back in June, he said something I've quoted in at least a dozen client meetings since.
"Customers don't always want more options. They want confidence."
He was talking about the crisis of sameness, about how generative AI is flooding commerce with content that all sounds identical, and about how the winning move is radical simplification rather than more. He also made the point that traditional SEO was about being found, and AI recommendation is about being understood and trusted. The audit question isn't where you rank. It's whether an AI agent would know when to recommend you, why, and what evidence supports the claim.
That framing is what pushed us to build the LLM Brand Intelligence Tracker we launched alongside that episode, and it's the same instinct behind LLM Search Console. You can't optimise what you can't see, and until recently nobody could see any of this.
Marcus also cited Harris Poll data that I think is underrated. 73% of consumers feel uneasy about how AI uses their shopping data, and 75% trust AI shopping tools less when the results are sponsored. Read that twice if you're currently building your agentic strategy around buying your way in. The consumer has already priced in their suspicion of paid placement in this environment. The organic recommendation is worth more here than it ever was in blue links.
I've also been working through the TechCrunch coverage of Universal Cart from May and Stripe's original ACP documentation, which is genuinely well written and worth an hour even if you never touch the implementation. Reading the protocol spec tells you more about what the models want than any marketing blog will, including this one.
The three things I would fix first
I'd start with feed completeness, and I'd measure it against what a shopper actually asks rather than what Merchant Center requires for approval. Approval is a floor. Write down the twenty questions your best customer asks a salesperson before buying, then check whether the answer to each one exists somewhere in your feed as a structured field. Most brands find that fewer than half do. Fill in the conversational attributes while adoption is still low enough that it's an advantage.
Then I'd fix refresh frequency, because it's unglamorous and decisive. If price and stock update once a day, you're gambling that no agent tries to build a cart in the other twenty-three hours. Get to sub-hourly. Your platform probably supports it and nobody has asked.
Third, I'd instrument the thing. Not with a spreadsheet of manual ChatGPT queries run by an intern once a month. Actual tracking of share of voice, citation provenance, and how the models describe your products across the markets and languages you sell in. Around 93% of AI search sessions end without anyone clicking through to a website, which means your analytics platform is structurally blind to most of what's happening. If your only evidence is sessions, you are measuring the small fraction of the iceberg that happens to be above water.
What this actually costs you
Here's the part that makes CFOs pay attention.
The agentic layer compresses your ability to differentiate on anything other than product truth and data quality. No hero image, no brand film, no clever landing page copy. Just specifications, price, availability, reviews, and whether the model trusts the source.
For brands with a genuinely better product and weak marketing, this is the best news in twenty years. The playing field just tilted toward them.
For brands whose advantage was distribution and advertising spend rather than product, it's the opposite. And a lot of category leaders are quietly in that second group, which is why I think the resistance to this shift inside large organisations is going to be emotional rather than analytical.
There's also a margin question nobody has answered yet. When an agent assembles a cart across four retailers and optimises for total delivered cost, you're in a price comparison you didn't consent to, at a moment you can't influence. The brands thinking about that now are going to handle it better than the ones discovering it during peak season.
One question
I'll leave you with the same question I've been asking every commerce director I meet since May, because the answers have been more revealing than anything in the industry data.
Pull up your product feed. Pick your three best-selling SKUs. Now imagine you are an AI agent with no access to your website, your ads, or your brand, and someone has asked you to recommend the best option for a specific, nuanced need in your category.
Reading only that feed, would you recommend your own product? And could you explain why?
If the answer is no, or if it takes you more than ten seconds to be sure, that's your next quarter's roadmap and it costs a fraction of what you're about to spend on paid social.
If you want to see what the models are actually saying about your brand right now across ChatGPT, Claude, Gemini, and Perplexity, that's exactly what we built LLM Search Console for. Reply to this email or find me on LinkedIn and I'll run your category. No pitch, no deck. Just the data, and you can decide what to do with it.
I'm curious about one thing in particular. If you've already started rebuilding your feed for agentic surfaces, what broke first? I'd like to know, because the failure patterns are still being discovered and most of us are learning them the expensive way.
Bruno Gavino is the CEO and founder of Codedesign, a digital marketing agency headquartered in Lisbon with offices in Boston, Singapore, and Manchester. He hosts Voice of Experts, a podcast and newsletter on AI, business, and marketing.
Track your LLM visibility at llmsearchconsole.com. Agency and insights at codedesign.org. Find me on LinkedIn.
