Why Commerce Is Becoming Invisible — and What That Means for AI Defensibility

Commerce is quietly splitting into two categories, and most businesses reliant on e-commerce are prepared for only one. 

On one side, buying is becoming invisible. Low-consideration, mundane purchases — replacing a set of tyres, reordering the same cleaning product, topping up a subscription — are starting to get triggered automatically, from device data and behavioural signals, without a human ever opening a storefront. A car’s health diagnostics flag worn tyre tread, the pantry item is low, and the transaction just happens. No interface, no browsing, no decision to make.

On the other side, buying is becoming more personal, not less. Categories where social proof matters, or where the shopping itself is part of the pleasure — colour cosmetics is the clean example — are moving in the opposite direction. Nobody wants an agent quietly reordering their foundation shade without them. What they want is advertising and commerce that gets more enjoyable and precisely tailored to them: model imagery matched to their actual skin tone, skin type, aesthetic, and lifestyle, so they can feel inspired and judge fit accurately before they buy, not after a return.

Both trends are being driven by the same underlying shift: an operating-system or agent layer sitting above commerce platforms, acting as the new signal source. It's the layer that knows your discontent towards existing products, and it's the layer that knows your skin tone and your aesthetic preferences well enough to feed a commerce platform tailored copy, imagery, pricing, and cross-sell that actually fits what you’re looking for next. Commerce platforms don't lose relevance in this shift — but their job changes. They stop being where decisions get made and start being where decisions get executed, fed by signals from somewhere else.

Between January and April 2026, 68% of US Google searches ended without a single click through to a website, up from under 60% two years earlier. 

The searches that used to send someone to your storefront are increasingly resolving inside the answer itself. Meanwhile, the traffic that does arrive from AI platforms converts noticeably better than ordinary traffic: over a third more revenue per visit, and close to double the conversion rate. Less traffic. Much better traffic. Arriving through a layer you don’t own and, until recently, wasn’t being measured on at all.

That’s not hypothetical, and the clearest evidence is a very public experiment that failed fast. When OpenAI launched in-chat purchase completion in late 2025, retailers who joined it saw conversion rates land at roughly a third of what they got from ordinary click-through traffic. Within five months, the feature was quietly scaled back. Shoppers were happy to research inside the chat window. They finished the purchase somewhere else, on their own app, their loyalty programme, their own checkout. What replaced native in-chat buying wasn’t retreat, it was reinvention: several of the retailers who’d tested it moved to branded assistants that handle discovery and link straight back to their own site to close.

Call it the law this year has actually written: discover in AI, buy on site. The site isn’t dying. It’s being demoted, from destination to something closer to a settlement layer, the place trust gets confirmed, and money changes hands, fed by discovery that’s now happening somewhere the brand doesn’t control.

You might think this sounds futuristic, but we're already living the reality. 

🎸 Spotify queues up your next album before you ask.

🪒 Subscribe & Save reorders your razor blades on a schedule you never consciously set.

📺 Netflix decides what's worth your evening before you've opened the app. 

None of that is a full agent transaction yet, but the soft purchasing power behind it, an algorithm quietly making the small decisions on your behalf, is already normalised. The invisible layer isn't arriving. It's expanding.

That's the uncomfortable part for anyone whose commerce value has been built on interface work — the storefront, the checkout flow, the on-site personalisation widget. If the OS or agent layer is where the decision-making signal actually lives, then interface-level work is exactly the layer that's commoditising fastest. Anyone can build a competent storefront. Far fewer people can build the data and signal-integration layer that connects a personal agent's understanding of a customer to what a commerce platform actually serves them.

The AI defensibility test for commerce infrastructure

Here’s the part worth an afternoon of a strategy team’s attention rather than a shrug. The same clean product data that makes a brand legible to an AI answer engine is also what a retail-media placement needs, what a marketplace listing needs to syndicate correctly, and what a growing wall of product-disclosure regulation is going to require regardless of whether any AI ever gets involved. It’s one substrate, and most businesses are currently being sold four separate, badly-built versions of it by four different suppliers who never talk to each other. Build it once, properly, and it pays for itself every time you use it. Build it four times badly, and none of the four quite works.

This is where it gets messy, and it’s worth sitting in the mess rather than skating past it. The team that owns the website is rarely the team that owns product data, and neither of them owns the brand mentions and third-party coverage that increasingly decide whether an answer engine cites you at all.

🛍️ Marketing owns the storefront brief. 

📊 Ecommerce operations owns the feed. 

🎙️ PR owns earned coverage and has no idea that coverage now functions as a ranking signal for a system nobody in the building has ever logged into. 

Ask three people in most commerce organisations who’s responsible for whether an AI answer engine knows the business exists and trusts it, and expect three different answers, or a long pause. 

Where this tends to break: teams respond to falling organic numbers by pouring more budget into the interface, a slicker storefront, a flashier on-site personalisation tool, while the underlying product data stays inconsistent across channels. It feels like progress because there’s something to point at in a review meeting. It isn’t, because the layer that’s actually eroding visibility was never the interface. 

The second common failure is treating this as a purely technical fix, handing it to engineering to clean the feeds and ticking the box, without anyone asking the harder strategic question: what does the business actually want an agent or answer engine to say about it when nobody’s watching? Clean data with no point of view behind it is just clean data.

This is the same infrastructure argument that underpins AI defensibility more broadly, applied specifically to commerce: the durable, hard-to-commoditise work isn't the visible feature layer; it's the infrastructure beneath it — content, data, and signal management that keeps getting more valuable exactly as the interface layer keeps getting easier to spin up. 

For a commerce business, that means the strategic question isn't "how do we build a better storefront experience?" It's "how do we own the signal-integration work between the agent layer and the commerce platform" — the unglamorous, defensible layer that determines whether the personalisation on that colour-cosmetics site is actually good, or just competent.

Businesses that keep investing purely in interface and feature work are building on ground that's eroding under them in real time. The ones that move now into the data and signal layer — figuring out how a commerce platform should consume and act on signal from an increasingly invisible buying layer — are the ones still standing when routine purchases have gone fully silent, and the only visible commerce left is the kind that has to work hard to earn attention.

This is the thinking behind the AI Defensibility Audit — a practical framework for any business working out which parts of their offer are infrastructure-level and defensible, and which are interface-level and about to be commoditised.

---

👉 Question worth sitting with:

Of everything your team currently sells as commerce work, how much of it is actually agent-proof?