The conversation about agent commerce within marketing teams has boiled down to a single question: How do we get AI to recommend us? It’s a reasonable question, but it’s the wrong one to start with.
A brand purchased by an agent is the one selected by the algorithm. A brand that a customer asks for by name is something a brand has to earn. AS strategist Jess Graham he says, the first case is procurement, and no one has ever fallen in love with sourcing.
This distinction is important as agents take on a larger role in product discovery. A trademark can be perfectly machine-readable and completely absent from the person. The agent evaluates, the customer receives a product, and the act of choosing silently disappears from the process.
For marketers who have martech and customer data, this is where the strategy can be tested. Algorithmic readability is important, but so is staying present to the human behind the agent.
10x your SEO with Semrush for business.
The most powerful SEO platform in the world, created specifically for businesses.
Request a demonstration
AI agents are already reshaping product discovery
The trading agent is already on the market, although not in the form promised by the first headlines. OpenAI instant checkout went live in September 2025 and has been retired in March 2026five months later.
Only a dozen merchants integrated it; usage remained low and shoppers who searched on ChatGPT still preferred to complete their purchase on the retailer’s site. OpenAI has not withdrawn from business. It has moved to a discovery-focused model that directs shoppers to merchant apps and storefronts, where the retailer handles payment.
Google went in the opposite direction with its own protocol. THE Universal Trade Protocolannounced at NRF in January 2026 with Shopify, Etsy, Wayfair, Target, Walmart and Visa, is now live and continues to expand: Cart support, catalog access, and identity link are already in use. Read the two stories together and one thing becomes clear. The checkout layer is still taking shape, but agents are already in the discovery layer.
Consumer research shows that AI is already influencing product discovery. Salesforce found 39% of consumers, including 54% of Gen Z, have used generative AI to discover and evaluate products. Adobe has too tracked traffic from generative AI tools to retail sites.
What research doesn’t yet demonstrate on the same scale is that agents buy on behalf of consumers. Salesforce found that 63% of Gen Zers are interested in having AI agents make purchases for them, but this reflects a stated interest rather than measured behavior.
Consider the momentum this might create. A customer asks an AI agent to purchase a moisturizer. The agent evaluates a thousand options with respect to price, reviews and delivery speed and buys one. The customer has never seen the packaging, never read the brand story, and never compared it to the product they may have already used and loved. The agent made the choice and the customer received the result.
That scene comes from Graham, who has mapped out what happens to brands when agents start making purchasing decisions for people.
Being readable to AI is not enough
Graham draws a clear distinction between algorithmic readability and brand preference.
Algorithmic readability involves structuring product data so that the agent can read it, classify it, and position a brand in the consideration set, or, in his words, be on top of the algorithm. This is where budgets go and the first returns are real. Brands that emerge within the responses generated by artificial intelligence gain more trafficand that traffic tends to spend more time on site. Work matters and should continue.
The other challenge is staying present to the human behind the agent. Graham’s name for the failed state is agent invisibility. The business implication follows directly. A brand purchased without being deliberately chosen by a human may have given up its pricing power.
To an agent optimizing price, ratings and delivery, an undifferentiated brand reads like a commodity and commodities compete on price until someone loses. Graham gives this a name that belongs on an income statement: the discovery taxthe cumulative cost of how customers find a brand and then decide it’s worth choosing again.
The cost of giving up control of discovery
Other industries have done it already paid the price for relinquishing control of the discovery.
- Hotels have turned discovery over to online travel agencies and now pay 15% to 30% per booking without owning almost any of the guest data.
- Musicians have turned discovery over to playlist algorithms and now earn fractions of a cent per stream, with the platform deciding who gets featured.
- Third-party sellers created demand in a marketplace, then watched the marketplace study what sold and launch their own competing products.
The sequence repeats in a loop. A new intermediary (in this case, AI agents) offers convenience. Brands accept worse economic conditions to maintain access. The intermediary captures the customer relationship and data, and brand differentiation comes down to a comparison of features.
Agentic trading is the most complete version of this model so far because the customer is not even present when the evaluation takes place.
Brand preference requires data you control
Graham’s prescription for brands it’s about building discovery experiences that agents can’t capture and giving people a reason to choose your brand from the start. The martech stack plays a critical role because earning preferences requires more than just making a brand readable for an agent.
Most brands that chase readability may provide agents with incorrect data. Static product feeds and last-touch attribution models describe what a customer did and don’t explain why they chose. It’s enough to rank. Being preferred requires more.
An agent that optimizes over a few structured fields treats each well-structured competitor as interchangeable. This instability manifests itself in the results of artificial intelligence.
SparkToro found less than 1% chance that the same brand appears in two identical AI queries. A brand flickers, present in one response and gone in the next, with no explanation offered to the marketer or customer.
Escaping agent invisibility then becomes a data decision before it becomes a campaign decision. It requires first-party, zero-data that captures preferences and relationship signals – those that explain why a customer chose a brand – held somewhere within a brand’s martech stack.
In practice, this means owning the moment of discovery rather than borrowing it (via third-party data or walled gardens, for example). It means capturing signals that come from community and direct conversation, treating delivery and unboxing as relationship data rather than logistics, and building a consensual identity that persists whether an agent is in the middle or not.
A brand that has the discovery, direct relationship, and underlying data becomes the brand that an agent can be instructed to search by name. This is the durable version of the top of algorithm and works on data controlled by the brand.
What marketers should do now
From a data perspective, the harder truth is that this is a problem of architecture rather than strategy. A team that fails to connect the tools it already has won’t suddenly capture customer preferences and signals. The acting moment does not create that gap but exposes and amplifies it, while increasing the cost of leaving it open.
Start with an audit rather than a tool. Map out your actual zero- and first-party customer data and be honest about it. A data set that records what was purchased and almost nothing about why it was chosen is built to be readable and stops there.
Then decide, deliberately, where relationship data is built in places that are unique to your brand: property discovery, community, the delivery moment that Graham is aiming for, and a direct channel that the platform can’t freeze. Treat algorithmic readability as the minimum due to the machine, then invest beyond that.
Put a real number on discovery tax this quarter. A board that hears “we win the deal and lose the relationship” can fund the solution faster than another dashboard can.
Place ownership of agent-facing data where the customer relationship already exists, within marketing, rather than letting it slide into a purely technical decision made downstream.
Agentic trading will deliver the transaction to whoever is most machine readable. Whether it will also transfer the relationship is a martech and data decision being made right now. Marketers need to guide that decision rather than leave it to default.
