How AI assistants will rewire retail.

Today: the assistant is the new Google

Right now, chat-based assistants behave almost exactly like search engines. You ask, they scrape, they summarize, they hand you links. The transaction itself still happens on someone else's website.

This is the safest possible interpretation of what is happening, and it is where most retail planning is still anchored. "Chat is the new SEO." Make sure your product pages are scrapeable. Maybe pay for some kind of placement when the time comes.

It is a reasonable place to start. It is also wrong about where this ends.

From discovery to conversion

The reason discovery does not stay separate from conversion for long is that nobody on the user's side wants it to. The whole point of asking an assistant to "find me a flight to Lisbon next weekend, under €200, morning departure" is that you would like the assistant to book the flight. Handing you ten tabs to compare is a failure mode, not a feature.

Every assistant, OpenAI, Anthropic, Google, Perplexity, plus the dozens of vertical agents being built on top of them, is racing to close that loop. Once they do, three things change at once:

1. The user no longer needs to visit retailer websites. The website's job collapses from storefront to fulfillment endpoint.
2. The assistant becomes the new point of conversion, the place where money commits to changing hands.
3. The economics of who pays whom invert. (More on this below.)

That is the moment retail stops being a search problem and becomes an agent problem.

Why direct-to-manufacturer is harder than it looks

A tempting story at this point goes: if assistants can place orders directly, will Amazon get cut out? Will every brand sell directly through the AI?

Probably not. At least not immediately.

Manufacturers are good at making things. They are mostly bad at:

• Handling one-off consumer orders at scale
• Fulfilling rapid, mixed-basket delivery
• Processing returns and refunds
• Maintaining live pricing and inventory across thousands of SKUs
• Offering bulk discounts or curation across brands

These are exactly the things aggregators spent two decades building. A shoe brand could in principle let an assistant order direct from its factory. In practice, the assistant will route most orders through whichever aggregator already has the product on a shelf, ready to ship tomorrow.

So the value chain does not collapse. It reshapes.

The full value chain takes shape

This is the picture the next few years are converging on:

A few things to notice:

• The assistant sits between the user and everything else. It is the single interface, the only thing the user actually talks to.
• There are multiple fulfillment paths, not one. Some orders go via multi-brand aggregators. Some go via curated or bulk specialists. Some go direct to the manufacturer. The assistant picks the path on the user's behalf, based on price, speed, trust, and whatever else it has been told to optimize.
• Data flows the other way. Aggregators sit on something the LLM providers desperately need: real, live, transactional data. What is actually in stock today. What prices actually clear. What returns rates look like by category. Aggregations and statistical models that no public web crawl will ever match.
• The model serves the assistant. The LLM is the engine. Without it the assistant cannot reason, compare, or transact. But the LLM itself sees almost none of the user directly. Its commercial relationship is with the assistant layer.

That last point is subtle and matters. The LLM provider and the assistant provider may be the same company today (OpenAI sells both GPT-5 and ChatGPT). They will not always be. The economics of a wholesale model business are very different from the economics of a consumer assistant business, and the industry will eventually separate them the way cloud computing separated infrastructure from application.

Aggregators get a second life, as data brokers

The most underrated part of this picture is what happens to aggregators.

The naive reading is that they get squeezed: the assistant talks to the user, the user no longer visits Amazon, Amazon's traffic collapses. End of story.

The less naive reading is that aggregators have something genuinely scarce in an AI world: ground-truth commercial data. Every order placed, every cart abandoned, every search-to-purchase journey, every return, none of that is in the LLM's training data, and most of it never will be. It lives in the aggregator's database.

Aggregators become to AI retail what Bloomberg is to finance. Invisible to the end user, indispensable to the people building tools for the end user.

That data is exactly what an LLM-driven assistant needs to make non-hallucinated recommendations. "Is this product actually in stock in Milan today? At what price? With what delivery date?" No amount of pre-training answers those questions. Only a live data feed does.

So aggregators get a new role: not the destination, but the data layer. They sell aggregations, statistical models, and real-time feeds to the LLM providers.

But for how long?

This new role is a real business. It is also not necessarily a permanent one.

The threat to aggregators is that the manufacturers eventually figure out how to feed their catalogs, inventory, and pricing directly to the LLM layer, bypassing the aggregator data brokerage entirely. The technical work to do this is not enormous. What is missing is the standard, and the trust, and the incentive.

The aggregators' best defense is the same thing that made them successful in the first place: curation, logistics, and bulk economics. A box of detergent costs less on Amazon than at the factory partly because of buying power, partly because of consolidated shipping. An assistant that cares about price will route through whoever delivers the best deal, and that is often still the aggregator.

But the moat is no longer the traffic. The moat is the operations. And operations moats erode more slowly than traffic moats, but they also erode.

Who actually pays

Here is the part that gets least discussed and probably matters most. In the new model, the user pays nothing, or close to nothing, for discovery. That is already how search works, and assistants will inherit the same expectation.

The economics:

• Users get free (or near-free) access to the assistant. Premium subscriptions exist, but the bulk of revenue does not come from them.
• Manufacturers and service providers pay the assistant for completed conversions: a referral fee, a bounty, a placement bid. This is the part most analogous to today's Google Ads, except the assistant is far more entangled in the decision than a search results page ever was.
• Aggregators pay the assistant for traffic and for keeping their inventory in the recommendation set. They also receive money, from the LLM providers, for the data feeds they supply. So they are both buyer and seller.
• LLM providers earn from serving the model into the assistant. Whether they own the assistant or just supply it, every transaction the assistant completes is, ultimately, a piece of model usage they get paid for.

The big inversion: today, retailers spend marketing budget on search engine optimization. Tomorrow they will spend it on assistant engine optimization, being the product the AI picks when a user asks an open question. Same logic, completely different surface area, almost no continuity of skills or tooling.

What this means, depending on where you sit

• If you are a brand or manufacturer Your priority is not your website. Your priority is being machine-legible: structured catalog data, real-time inventory APIs, a clear identity an assistant can resolve unambiguously. The brands that are easy for an AI to recommend will be recommended. The ones that are not, will not be.
• If you are an aggregator You have two business models to invest in simultaneously: the customer-facing one you already have, and the data-provider one you do not. The second one is less glamorous, less visible, and may end up being where your enterprise value sits in five years.
• If you are building assistants The hard part is not the model. The hard part is the supply side: the orchestration layer that turns a "find me X" request into an actual, reliable transaction. The model is becoming a commodity input. The supply network is the durable asset.
• If you are a user You probably will not notice most of this. You will just notice that fewer things go through Google, more things go through a chat box, and one day you will realize you have not visited a retailer's website in months.