For most of the last decade, I built and scaled marketplaces, including at some of North America's largest food delivery platforms. The playbook was clear: attract customers, bring them into an experience, help them discover products, optimize conversion and bring them back.
A year ago, I asked myself what commerce would look like five years out. I expected AI to make that playbook more personal. I was half right. Personalization is happening, but on the shopper's side. Their agent handles discovery, comparison and decision-making, and the merchant's storefront may never be part of the experience at all.
That shifts what matters. The product data underneath the storefront now matters more than the storefront itself. I'd built with AI before, but always as a feature inside a traditional product. In agentic commerce, AI is the market itself, and that has changed how I think about both product and growth.
Lately, we've built against MCP, OpenAI and Stripe's ACP, and Google and Shopify's UCP, while new models have shipped every few weeks. The landscape is moving quickly enough that product decisions made at the beginning of a quarter can look very different by the end of it.
Keeping up with announcements is the easy part. Knowing which ones should change your roadmap is harder. I learned early in building with AI that anything you build purely to work around a model's weakness has a shelf life. The next release might make it unnecessary. In agentic commerce, that lesson became existential because the model, the protocols and the shopper's agent all move underneath you at once.
The model is an input that gets better over time. What you own is the layer it reads, so we built it around the catalog.
Regardless of which model, protocol or agent experience wins, every agent trying to make a commerce decision needs reliable product information. Availability, attributes, variants, pricing, descriptions and the relationships between products don't become less important as models improve. They become more important.
Building on that foundation gave us a head start. When Anthropic open-sourced its Claude for Commerce blueprint, our team went from the first walkthrough to building a shopper agent and a merchant agent in 11 days.
AI has dramatically changed the cost and speed of building software. That moves the bottleneck from engineering to evaluation.
Evals weren't new to me. While building AI features, I'd spent years testing our own models inside experiences we designed. Agentic commerce added a harder version of the problem: you also have to evaluate how someone else's agent reads your customers' products, in a system you don't control.
On an AI team, the hard skill now is knowing whether the output is good enough to ship. So I write evals before specs, and they now cover both sides.
Before building an experience, I want a golden dataset of real products and real shopper questions, along with a clear definition of what a correct answer looks like:
- Does the agent identify the right product?
- Does it understand the attributes?
- Does it make an accurate recommendation?
- Does it hallucinate availability or pricing?
If we can't define what good looks like, we're probably not ready to build. The same thinking influenced how we approached our Shopify app. Instead of optimizing for the complexity of the integration, we worked backwards from the merchandiser's experience. What should they never have to deal with? CSV exports. Engineering tickets. Theme rebuilds. The result is an app that syncs a full catalog, enriches it and publishes in under 20 seconds.
In an AI-first company, product craft and taste matter even more. When the cost of producing something goes down, judgment becomes a bigger differentiator.
This has probably been the biggest adjustment for me coming from a growth background.
For most of my career, my mental model assumed I owned the surface. The landing page. The ranking. The product detail page. The checkout flow. The A/B test.
Agentic commerce changes that. The merchant doesn't necessarily own the interface anymore. The shopper might discover a product through an agent, compare it against alternatives and make a purchase without ever visiting the merchant's website.
You can't A/B test a page you don't render.
Chat is the obvious starting point. It's intuitive, direct and accessible. But we haven't figured out the final form factor.
Voice is a strong contender. I still can't confidently tell an agent to plan and book a trip to watch Arsenal without going through a dozen prompts and confirmation screens.
But the shift for growth is already happening. For years, we optimized the experience customers saw. Increasingly, we need to optimize the information agents consume:
- Is your product discoverable?
- Can an agent understand what makes it different?
- Is the information accurate enough to recommend it?
- Does it have the context needed to match it to the right shopper?
The fundamentals of growth still apply. We're just optimizing a different surface.
AI-referred traffic is producing some impressive numbers, but we still don't agree on what they tell us. Adobe reported that shoppers arriving from AI assistants generated significantly more revenue per visit than other visitors in September 2026. Earlier, research from Frankfurt School of Finance & Management found that ChatGPT referrals made up less than 0.2% of ecommerce traffic across the sites studied.
The numbers look contradictory, but they measure different parts of the funnel: one the value of visitors who arrive, the other how many arrive at all. Neither captures the shopper who asks an agent for a recommendation, gets an answer and never clicks through. The first half of the funnel is increasingly happening somewhere merchants can't see.
The teams that won early in SEO didn't wait for perfect attribution. They found leading indicators they could influence and built from there. One we're interested in is agent share of voice. Take a fixed set of shopper prompts and measure:
- How often does your product appear?
- How accurately does the agent describe it?
- How often is it recommended over alternatives?
We've seen why this matters in our own testing. We ran the same trail running shoe through our preview agent twice, once on its original product page and once after enrichment. On the original page, the agent found four missing specs, dismissed the tagline as marketing and moved on without recommending the shoe. On the enriched page, it answered every question the shopper asked and still flagged the one detail that was missing: a rock plate.
Nothing about the shoe changed, only the information the agent had to work with. You'd never see that difference in a traffic report, but you would see it in share of voice.
The pace reminds me of my early days in food delivery in the mid-2010s, when we were building the plane while flying it. It's the same here, except the airspace keeps changing too. Models, protocols and shopper behavior shift while you're mid-flight. When everyone is figuring this out at once, it helps to share what we're seeing, including what we get wrong.
Now the question I keep asking is different: what does a merchant still control when an agent does the shopping? Product discovery, customer relationships, merchandising, distribution and measurement are all being reconsidered. We don't have the answers yet, and that's the problem I want to work on next.
