AI Agents and Online Shopping

How autonomous purchasing agents will reshape discovery, trust, and the entire merchant-customer relationship.

The first thing that surprised me about AI shopping agents was not how good they are. It was which merchants they choose. In informal tests across categories I know well, the agent frequently did not pick the brand with the best photography, the strongest social presence, or the biggest ad budget. It picked the listing where the information was complete.

That should worry a lot of marketing teams, and it should excite a smaller number of product teams.

Discovery stops being a funnel

Every discovery model I have built has been shaped like a funnel. Impressions at the top, some fraction clicks, some fraction of those converts. At AliExpress US the AI recommendation system I worked on served more than 50 million monthly users and lifted GMV around 40 percent, and even that system was fundamentally a funnel optimizer. It got better at surfacing the right item in a finite slate of options a human would scan.

An agent does not scan a slate. It runs a search, evaluates candidates against a specification, and returns one or two answers. There is no impression-to-click ratio, because there are no impressions in any meaningful sense. There is a query and there is a verdict.

This collapses the funnel into something closer to a qualification test. You either satisfy the constraints or you do not appear. And unlike search rankings, where position ten still gets some traffic, position three in an agent's internal evaluation gets nothing, because the agent does not show its shortlist.

The strategic consequence is that marginal improvements in the middle of the funnel stop mattering. A two percent lift in click-through on a category page is irrelevant to a buyer that never sees the category page. What matters is whether you pass the filter.

Trust becomes a computed variable

Human trust is built through repetition, aesthetics, and social proof. It is slow to build and slow to lose. A customer who likes your brand will forgive a bad experience once, maybe twice.

Agent trust works nothing like this. It is a running estimate updated on evidence, and it has no loyalty and no sentimentality. If your listed price does not match your checkout price, if your stated stock is wrong, if your shipping estimate is optimistic, an agent will register that as unreliability and weight you down on the next relevant query. It will not tell you. It will not complain. Your conversion rate from agent traffic will simply drift downward and you will have no clear attribution for why.

This is a genuinely new failure mode, and I think it is underrated. Merchants are used to trust failures being loud, in the form of reviews, chargebacks, support tickets. Agent trust failures are silent.

None of these are marketing problems. They are operations problems that were previously buffered by the fact that humans do not systematically track them. Agents do.

The relationship gets mediated, and that is the real shift

Here is the change I think most people are underweighting. For twenty years, the direction of ecommerce has been toward disintermediation. Brands built DTC channels specifically to own the customer relationship, the data, and the email list. That was the entire strategic thesis of the DTC era, and it is the thesis I operated under running consumer electronics growth at Next2Market.

Agentic shopping re-intermediates. If a customer's assistant is doing the browsing, the comparing, and increasingly the buying, then the merchant's relationship is with the agent, not the person. Your beautifully crafted post-purchase email sequence is being read by a machine that summarizes it or discards it. Your retargeting is pointed at a human who was not the one shopping.

I do not think this means brand stops mattering. It means brand operates on a longer loop. The human still forms a preference, still tells the agent I like this brand or avoid that one, still notices when a product delights or disappoints. But the day-to-day transactional relationship moves to a layer you do not control.

Owning the customer relationship was the DTC bet of the last decade. Agentic commerce quietly reprices that bet, because the party placing the order may no longer be the party you built the relationship with.

What happens to affiliate and creator channels

I spent years building affiliate infrastructure. At AliExpress the program ran to more than 10,000 partners across 40-plus countries and drove roughly a 30 percent revenue lift, and the mechanic was straightforward: partners create content, content reaches people, attributed links close the loop.

Agents break the link in the middle. If an agent reads a comparison article and then buys directly, the affiliate link was never clicked, and the publisher who did the work that informed the purchase gets nothing. That is an attribution problem, but more importantly it is an incentive problem. If the content that trains and informs agents cannot be monetized, less of it gets made, and agents get worse.

I expect this to be resolved, but messily and slowly. The plausible outcomes are some form of agent-readable attribution, direct licensing between commerce platforms and content producers, or a shift toward content whose value is in the human audience rather than the purchase event. Probably all three in different niches.

For affiliate marketers specifically, my advice is unsentimental: the content that survives is content an agent cannot replicate. Original testing. Long-term ownership reports. Category expertise that reflects actually using fifteen products rather than reading fifteen spec sheets. The volume play in affiliate marketing has been dying for a while. Agents are going to finish it.

The practical read

I am wary of predictions with timelines attached, so I will frame this as what I would bet on rather than what I think happens by a specific quarter.

I would bet that agent-driven purchase volume stays small for longer than the enthusiasts expect, and then grows faster than the skeptics expect, because the bottleneck is checkout infrastructure and payment trust rather than model capability. I would bet that the first categories to tip are replenishment and commodity replacement, where the specification is precise and the emotional stakes are low. I would bet that brands built on genuine product differentiation do better in an agentic market than brands built on distribution advantages.

And I would bet that most merchants find out they are losing agent-mediated sales about a year after it starts happening, because nobody instrumented for it. That is the cheapest thing to fix, and almost nobody has fixed it.