AI Agent Commerce: Where the Money Moves

AI agents now complete commerce transactions on behalf of users. This is no longer a demo. It is a production behavior happening millions of times a day across consumer and business use cases. The financial plumbing behind it is forming fast. So is the advertising layer. This post covers where the money flows, what part of that flow Surfacedd fits into, and what we do not do.

The Growth Numbers Are Real

Adobe's 2025 Holiday Digital Insights report put year-over-year referral growth from AI sources at 805%. That is referral traffic from chatbots, answer engines, and agents into retailer sites. The number is large partly because the base was small — AI referrals were near zero two years ago — but the shape of the curve is what matters. AI-originated commerce traffic has gone from experimental to measurable in three holiday seasons.

Salesforce's 2025 Connected Shoppers report found that 17% of US online shoppers had at some point completed a purchase either inside or through an AI assistant. The figure is higher among under-35 users and higher still in travel and software categories.

Shopify reported in its Q4 2025 earnings that AI-assisted checkouts, including those initiated by agents operating on behalf of users, grew from roughly 2% of sessions in Q4 2024 to 11% in Q4 2025. The company has invested in tooling to make storefronts more legible to agent crawlers.

These numbers are noisy. Different sources define "AI commerce" differently. Some count any session that started with an AI referral. Others count only transactions completed without a human clicking through a traditional interface. But the direction is the same across all of them. AI agents are handling a rising share of commerce.

Agent Checkout Patterns

Three patterns dominate in 2026.

Pattern 1: The research-to-redirect agent. The user asks an agent for a recommendation — a camera, a flight, a B2B tool. The agent generates a ranked list. The user clicks through to a retailer or vendor site and checks out there. This is the pattern closest to traditional search-plus-click, and it is the most common. Adobe's 805% figure is dominated by this pattern.

Pattern 2: The assisted-checkout agent. The user stays inside the agent and the agent completes the checkout by calling payment and fulfillment APIs. The user approves. Shopify's agent checkout and several travel booking agents operate this way. The retailer's brand appears, but the transaction surface is the agent.

Pattern 3: The autonomous-purchase agent. The user delegates a recurring or routine purchase to an agent that executes without per-transaction approval. Reorder pet food. Renew a software subscription. Rebalance a portfolio. The agent runs on a schedule or rule set. This is the smallest category today but growing fastest among developers building vertical agents.

Each pattern puts the brand in a different position relative to the user. In pattern 1, the brand still has a storefront moment. In pattern 2, the brand is a line item inside the agent's output. In pattern 3, the brand is invisible unless the agent surfaces it.

Advertising works differently in each. The advertising question in pattern 1 is how to earn a citation and a click. The question in pattern 2 is how to appear as a disclosed recommendation. The question in pattern 3 is harder: how to stay in consideration for an agent that may never expose the brand to the user at all.

Where Advertising Fits

Surfacedd's scope inside this picture is narrow on purpose.

We do disclosed recommendations. When an agent returns a list of options or makes a suggestion, a Surface can appear alongside the organic output as a clearly labeled sponsored recommendation. The Surface carries the disclosure. The organic output is not altered. The user can distinguish the two.

This is the advertising equivalent of a sponsored listing that is visibly separate from organic search results. It works in pattern 1 — where users are reading ranked recommendations — and it works in pattern 2 — where a disclosed option appears alongside the agent's preferred choice. It does not work well in pattern 3, where the user is not looking at the agent's reasoning at all.

We do not do paid organic mentions. If a brand pays for its name to appear inside the agent's answer text, with no disclosure, we treat that as manipulation of the organic layer. It is the digital equivalent of product placement passed off as editorial. We do not build it. We do not sell it. When we see it advertised by other networks, we flag it.

The distinction matters commercially. Paid organic mentions are what some early agent-advertising vendors are selling under different names. The margin is higher. The click-through rates look good. The long-term effect on user trust is corrosive. When users discover that the agent's answers have been commercially influenced without disclosure, they stop trusting the agent. The publisher loses. The advertiser loses. Only the vendor who sold the shortcut benefits, briefly.

For the principled version of this argument see honest AI advertising. For the developer-side economics see AI agent monetization. For the general category frame, see advertising for AI agents.

Transaction Fees Are Outside Our Scope

A large share of agent-commerce revenue in 2026 comes not from advertising but from transaction fees. An agent that completes a booking, a purchase, or a subscription can earn a commission from the merchant, a take rate from the payment processor, or a referral fee from an affiliate program. Several prominent agent platforms, including some chatbot-native ones, have made transaction fees their primary monetization model.

Surfacedd does not operate in this layer. We do not handle payments. We do not take transaction cuts. We do not broker affiliate relationships. Our scope is the advertising layer — disclosed Surfaces rendered inside agent output — and stopping there is deliberate.

Three reasons.

First, focus. Transaction-fee monetization requires a different set of relationships — direct merchant contracts, payment compliance, chargeback handling, fraud operations. Mixing it with an advertising product dilutes both.

Second, conflict of interest. If the same platform both takes transaction fees from a merchant and sells advertising inventory to that merchant, the incentives to favor the paying merchant in the agent's output multiply. Keeping advertising separate from transactions keeps the conflict surface small.

Third, user trust. The disclosure story is simpler when a platform only does one thing. "This is a sponsored recommendation" is clean. "This is a sponsored recommendation and also we take a cut of the purchase" is confusing, even when both are disclosed.

Agent developers who want transaction-fee revenue should build it. It is a valid model. Surfacedd is not a fit for it, and that is fine.

What to Watch Through 2026

Four things are worth tracking over the rest of the year.

Agent checkout standardization. Several industry groups are working on standards for how agents communicate purchase intent to merchants. Shopify, Stripe, and a handful of agent platforms have published drafts. The shape of the final standard will determine how smoothly transactions move through agents and how much friction remains.

Disclosure norms. Regulators in the US, EU, and UK are all looking at agent-generated commercial content. Expect clearer rules on labeling by end of 2026. Honest operators will be ahead of the rules. Dishonest ones will be caught by them.

Attribution models. The measurement problem in agent commerce is severe. Click attribution mostly fails. New models — brand-lift studies, incrementality tests, post-purchase surveys — will have to take over. The vendors who solve this credibly will win trust.

Inventory explosion. The number of agents capable of hosting commerce is rising fast. Every major chat product, a growing number of vertical agents, and a long tail of developer-built agents will all be candidates for advertising inventory. Matching brands to this inventory at scale is the problem the network layer exists to solve.

The Short Version

Agent commerce is real, the growth rates are large, and the advertising layer inside it is being built right now. Surfacedd's piece of that build is narrow: disclosed recommendations rendered as Surfaces, alongside agent output, with the organic answer left alone. We do not do paid organic mentions. We do not take transaction fees. We think the discipline of doing fewer things cleanly is the advantage, not a limitation.

The money is moving. Where it lands depends on which norms the ecosystem picks. The right norms are the ones where users can tell, at a glance, what is sponsored and what is not.