Assistants Are Marketplaces. The Link Economy Collapses

Breaking this week: assistants close the loop

Two announcements in late September created a clean before and after. On September 29, 2025, OpenAI moved ChatGPT from guide to cashier by introducing a native buying flow that turns a product recommendation into a completed purchase without sending you to a merchant site. Etsy is live at launch in the United States, with Shopify next, and payments ride on Stripe. See the Reuters report on Instant Checkout.

A day later, on September 30, 2025, short video began shifting from human shot to model made. Meta rolled out a feed of AI-generated clips in its Meta AI app, and OpenAI positioned Sora not just as a creator but as a feed of synthetic video. Coverage framed it as a flood of machine-made content. Read the Washington Post coverage of Sora and Meta.

Put the two together and the pattern is clear. Assistants now own both sides of the funnel. They can generate the content that captures attention and they can complete the transaction that captures value. Links used to be the connective tissue of the web. As of this week, they look more like an optional courtesy.

The assistant ceases to be a detour. It becomes the point of sale.

The end of the click-through habit

For thirty years we trained users to click out. The economy around the web formed to monetize that habit. Search results and social feeds acted as paid gateways that sent people to brand domains where pixels, analytics, and checkout stacks lived. The model rewarded traffic brokers that delivered qualified clicks at scale.

When an assistant both recommends and transacts, the click becomes an escape hatch, not the default. The assistant sits on the demand, interprets intent in natural language, selects offers using its own ranking logic, and completes the order inside its interface. The chain from query to cart compresses into a single conversation. That compression is not a small user experience tweak. It is a power transfer.

What changes first:

Measurement loses its trail. Referrer data and last-click attribution lose resolution because there is no click path to trace.
Merchandising shifts to answer rank. Shelf position becomes answer position. If you are not in the assistant’s first list, you effectively do not exist.
Loyalty migrates to the assistant. The assistant holds the payment credentials, the preferences, and the habit. Brand loyalty must be earned through the assistant’s trust layer, not through your own cookie jar and email list.

It is tempting to call this just another channel. It is closer to moving from malls to concierges. A mall gave you frontage and foot traffic. A concierge interprets the request, decides what to present, and can complete the purchase on your behalf. The store sign matters less than the concierge’s defaults.

For a deeper dive on how policy and incentives shape this layer, see our take on the invisible policy stack.

Choice architecture moves from pages to defaults

Search results pages make choice visible. You see multiple blue links, labelled ads, and shopping modules that list sellers side by side. In assistants, the choice architecture hides in prompts and defaults.

The first list wins. If the assistant proposes three winter boot options under 150 dollars, most users pick one of those three. The fourth-best boot never gets a chance.
The flow nudges the outcome. A Buy button inside the chat suggests closure. An Open in browser link suggests extra work.
Defaults silently shape margins. If the assistant defaults to a particular payment method, shipping partner, or return policy, that path shifts both conversion and unit economics.

This is not theoretical. The Instant Checkout announcement notes that seller attributes such as availability, price, quality, and support for native checkout influence ranking and eligibility. Reasonable signals, but they create a loop. Features that make the assistant’s job easier become the features that win distribution inside the assistant. Participation becomes table stakes.

Identity dissolves into a trust layer

On the link-first web, brand identity lived in your domain name, your design, and your checkout. In an assistant-first world, identity dissolves into a trust layer that the assistant surfaces.

Think about picking a driver in a ride-hailing app. You do not remember the driver’s personal brand. You remember the platform’s ratings, guarantees, and refund flows if something goes wrong. That is the future of product shopping inside assistants.

Guarantees and returns become platform primitives. If an assistant offers instant refunds for certain categories, that promise is more salient than an individual retailer’s fine print.
Safety and authenticity are standardized. Provenance claims, recall notices, and sustainability scores appear as badges in the chat, not as paragraphs on a product page.
Merchants compete to be legible. Your first job is to be compatible with the assistant’s schema so the assistant can vouch for you.

You should still tell your story through video and community channels. But when the purchase happens in a dialog, the levers are fulfillment reliability, policy transparency, and product data quality. Your logo will be present. The assistant’s trust layer will be the billboard.

If you are tracking how synthetic media will shape brand storytelling, see our analysis of compiled video.

Regulation must adapt to conversational shelves

Law and policy assume we can see the shelf and the ad. With assistants, the shelf is a sentence and the ad can be a suggestion. Regulators will need updated rules and updated test methods.

Four changes that fit both U.S. and EU traditions:

Disclosure inside dialog. Define what counts as clear and conspicuous when a recommendation is a generated sentence. Labels should be readable by humans and machines and must travel with the recommendation when shared.
Ranking audits that include protocol incentives. If eligibility or ranking depends on using a native checkout, disclose the dependency and test its reasonableness. Sample outputs across intents, demographics, and regions to detect systematic steering.
Separation of roles. If one platform builds the assistant, runs the payment rails, and controls the marketplace, governance should prevent unfair self preferencing. Remedies can be simple. If the assistant recommends its own house brand, disclose it clearly. If it blocks rivals’ promos or coupons, make that measurable.
Redress in one tap. Require fast refunds, clear dispute flows, and human escalation when automation fails. The assistant can initiate the fix because the assistant initiated the purchase.

The goal is not to freeze the future. It is to keep incentives honest when recommendation and sale occur in the same breath.

Build the rails: provenance, competition, reputation

Acceleration is the safe assumption, which means the rails must exist before abuses do. Here is a blueprint for the next four quarters.

Provenance rails

Signed product records. Require products presented in assistants to include signed metadata: origin, authorized seller status, warranty, recall status, and a unique item hash where practical. The assistant verifies signatures before presenting options.
Cryptographic receipts. After checkout, the assistant issues a receipt that includes the model version used, the ranking inputs, and a signed record of the seller’s commitments. That trail supports disputes and audits.
Watermarked media. For model-made video and images, watermark at render time and propagate labels in headers and transcripts. Downrank media that loses provenance during editing.

Competitive safeguards

Protocol neutrality tests. Publish the protocol specification and a conformance suite that exposes all fields needed for fair comparison across sellers. Let merchants validate how their data will render.
Choice moments. On the first purchase in a new category, show a choice screen to set default payment provider, shipping profile, and a preference for marketplace versus direct brand fulfillment. Make settings portable across assistants.
Auction transparency. If paid placement exists, attach a small expandable panel that explains why a suggestion appears, who paid, and what other candidates were considered. Keep the explanation in plain language.

Reputational markets

Assistant-level ratings. Let users rate the helpfulness and fairness of a recommendation, not just the product. Publish anonymized category scores so third parties can study assistant behavior.
Verified merchant identity. Require government-backed checks and beneficial ownership disclosure for merchants that use native checkout. Tie penalties to identity, not just accounts.
Post-purchase performance. Feed shipping times, return rates, and dispute outcomes back into ranking models in a way merchants can monitor. Publish standardized dashboards that reveal how to improve standing.

Agent to agent shopping arrives next

The phrase artificial intelligence can sound abstract. The near future of everyday purchases will be concrete: agents negotiating with agents.

The user agent knows preferences and constraints, is allowed to spend from a budget, and asks clarifying questions. It remembers sizes, allergies, and delivery rules.
The merchant agent knows inventory, prices, shipping promises, and bundle offers. It confirms delivery windows and accepts or rejects counteroffers.

Over the next 12 to 24 months, we will see protocols that let these agents negotiate directly.

Intent negotiation. The user agent expresses a structured intent: winter boots, size 8, budget 150 dollars, must arrive by Friday, prefer recycled materials. The merchant agent responds with options and tradeoffs.
Price and policy bargaining. Merchant agents make limited, logged concessions to close the sale, such as free expedited shipping in exchange for final sale on clearance items or a discount for flexible delivery windows.
Fulfillment confirmations. Once a deal closes, the merchant agent issues a signed promise with handling time, carrier, and a delivery confidence score. The user agent stores this next to the cryptographic receipt.

At maturity, the assistant orchestrates many merchant agents in parallel and chooses whom to ask first based on relevance and reliability. That is why today’s data work matters. If your catalog is messy, your return policies are opaque, or your shipping estimates are unreliable, your merchant agent will lose the negotiation before it begins.

For a platform view on diversity among models and why it matters, read our take on why model pluralism wins.

What builders and retailers should do now

You do not need to boil the ocean. You need to be first in line when assistants become the default doorway.

Join the checkout protocol early. If the assistant publishes an integration guide, implement it. Do not wait for multi item carts. Start with your top 50 products and learn how ranking reacts to your data.
Clean your product data. Use consistent attributes, high quality images, and unambiguous variants. Mark authorized reseller status and warranty terms in machine readable fields.
Model your policies. Convert returns, exchanges, warranties, and service levels into rules your merchant agent can enforce. Ambiguity wastes negotiation cycles and will quietly downrank you.
Instrument refunds and disputes. Make instant refunds possible for low risk items. Log and expose dispute resolution times. Expect those metrics to shape your assistant ranking.
Build brand outside, trust inside. Keep storytelling and community on your own channels. Inside the assistant, focus on the proof points the trust layer surfaces: reliability, clarity, provenance.
Design for explainability. Add fields that make it easy for an assistant to explain why your product is a good match. Think feature summaries, policy shortcuts, and clear tradeoffs.
Price for negotiation. Expect limited, logged concessions. Build margin room where policy changes can substitute for price cuts, such as flexible delivery windows or credit for future purchases.

The risks and how to blunt them

The same dynamics that simplify buying can amplify misaligned incentives.

Conversion over everything. An assistant could quietly steer to higher fee offers. Countermeasure: auction transparency and third party audits that correlate recommendations with take rates.
Dark patterns in sentences. Nudges move from pop ups to phrasing. Countermeasure: dialog disclosure rules, user level controls, and a report abuse flow inside the chat.
Vertical integration without checks. One company could own recommendation, payment, and marketplace distribution. Countermeasure: separation of roles and simple, testable conduct rules.
Synthetic sameness. Machine made feeds could drown out human creativity. Countermeasure: provenance signals, diversity targets in feeds, and user controls to steer curation.

None of these outcomes are inevitable if we lay the rails and keep measurement honest.

The bottom line

This week felt like a small step, but it is a crossing. Once assistants can both shape demand and capture it, links stop being the spine of the web and start becoming a side door. The habit of clicking out will survive for a while. Then it will feel like writing a check at a grocery store. You can still do it. You will not want to.

The collapse of the click-through economy into model mediated demand capture is not a catastrophe. It is a redesign of the interface between people and markets. Build the provenance rails so we know what we are seeing. Install competitive safeguards so ranking rewards relevance, not rent. Grow reputational markets so good actors win and bad ones are caught.

Assistants are becoming marketplaces. That is not a headline for next year. It is the operating assumption for the next release cycle. Design for it now, while the defaults are still being set.