June 17, the Agentic Commerce Collective convened its third roundtable to work through that exact problem: as AI agents increasingly decide which products get discovered and selected, how do you make sure you’re surfaced, trusted, and chosen by a machine?
The session, led by Eric Shea (Strategy and Digital Transformation, PwC), Scot Wingo (Co-Founder, ReFiBuy), Jameelah Calhoun (VP Market Strategy, Forter), and Jason Grunberg (CMO, Forter), skipped the hype and focused on what has genuinely changed: machine-readable data, trust signals, recommendation logic, and how brands show up when the shopper is a machine.
One line captured the shift better than any statistic. As a participant put it, and the panel agreed: “SEO used to be the CMO’s problem. GEO is the CEO’s problem.“
Agentic shopping went from interesting to real
“Last year there was a lot of excitement. This year it’s real,” said Eric, who now spends “120 percent” of his time on agentic commerce for PwC’s clients. The data backs him up. In a recent PwC consumer survey, more than 63% of shoppers said they now start product discovery using AI tools. Among high-income millennials, roughly 58% use AI as their primary shopping method, and about half of all consumers reach for it beyond discovery, for final consideration and post-purchase tasks.
The behavior is real, even if the checkout isn’t. Most AI surfaces have stepped back from completing transactions and now push shoppers back to the brand or retailer site to buy. That referral traffic is worth paying attention to: Eric noted it converts about nine times better than social commerce.
“More than 20% of all prompts across AI surfaces now carry some shopping intent, roughly a third of it pure discovery, a fifth comparison, and only 13% for actual purchase.” — Eric Shea, PwC
The categories leading the way are the considered ones, where shoppers want help getting the decision right: electronics, home goods, apparel, beauty, and hardware.
Showing up is not the same as being recommended
Here is where most brands get a false sense of security. They appear in AI results when someone names them directly, so they assume they’re covered. They aren’t.
“It’s not just about if you’re showing up. It’s about being recommended — and that’s where there’s a huge discrepancy.” — Eric Shea, PwC
Eric shared a PwC case study to make it concrete: a top-three global CPG brand was losing share among younger shoppers, so his team ran 5,000 prompts across that company’s target shopper profiles. The brand surfaced just 16% of the time, and it was never actually recommended. The root causes were telling: the company had over-optimized its site for conversion and stripped out the brand storytelling and product claims that AI engines look for. The team projected a 2% to 3% share recovery over twelve months, roughly $100 million in upside, from closing that gap.
That economic weight is why this conversation has moved from the marketing team to the boardroom. Eric pointed to one top CPG now monitoring GEO at the board level, brand by brand. As Jameelah added, the models increasingly reward authoritative content and verifiable claims, not just volume of user-generated reviews. Knowing why an agent recommends you, or doesn’t, has become a leadership question.
Prepare your site ready for AI-driven commerce.
The gap closes in your product catalog
Scot took the discussion from visibility to mechanics, and named the core problem: data asymmetry. AI engines know an enormous amount about the shopper, often including memory of past chats and access to email receipts across retailers. Brands, meanwhile, still feed those engines about five attributes per product: brand, size, color, price, and inventory.
“The old Google shopping was kind of like this creaky dirt road… and now we have four-lane super highways.” — Scot Wingo, ReFiBuy
The product detail page is the delivery vehicle for that data, because crawlers keep returning to it even when feeds exist, so blocking bots or burying data in images and carousels costs you placement before the race starts. New protocol “lanes” (ACP, used by ChatGPT and Stripe, and UCP, used by Gemini and Shopify) open four ways to feed engines richer information:
- Expanded basic attributes, beyond the old five-field standard
- Conversational attributes, like scent profile, use occasions, and skin type
- Reviews you’re comfortable sharing
- Product-level FAQs
Because FAQs are the highest-leverage lever of the four, Scot recommends making yours robust: closer to 200 questions than 10. That content shouldn’t all go on the product page because it’s not meant for human consumption. Instead, you should put that in your metadata and data feed so an agent can find it — and share it back with the consumer.
Scot also recommends going “secret shopping”: run specific prompts across every answer engine and watch how your products appear on the product card and the offer card. The prize positions are the left product card and the number-one offer card. The platforms themselves are moving fast. Gemini leads, with a cross-merchant Universal Cart reportedly days from launch, Claude has just added product cards, ChatGPT has pulled back to referral-only, and Perplexity is unwinding its checkout experiments. With only about a quarter of search traffic now leaving Google for external sites, the brands that win will be the ones whose catalogs are ready to be read.
What’s next
The agents are already shopping. The brands that treat search and discovery as a leadership priority, with a measurement strategy, a catalog built to be read, and the claims their customers are actually searching for, will be the ones that get recommended.
Want to participate in more agentic conversations? Attend IMPACT
