There's this story going around about agentic commerce, that when it is the AI agent that does the shopping, the classic marketing stops working. The agent reads the data, ranks the options, buys the best product on the merits, and is 100% completely immune to all the persuasive tactics.
A June 2026 preprint from Trine University and Texas A&M ran tens of thousands of controlled product recommendations through GPT, Claude and Gemini, holding price, rating, specs and review count identical across competing products so that only one variable moved at a time.
Four findings if you sell online
Agents are blind to urgency. Scarcity, anchoring and loss aversion moved the agent's pick 10–13% of the time, against a 4% baseline. The countdown timer does nothing.
Authority works far too well. Cited studies and expert endorsement moved the pick 50–73% of the time. The agent is naive about exactly the signals humans have learned to discount, and the authority claims in the study were fabricated. Nothing checked them.
The agent is not brand-neutral. With two products identical on every visible dimension, the recognized brand was recommended in 100% of 670 trials. The unknown brand never won.
But the moat is thin. The smallest edge the researchers tested, one tenth of a star on a rating, flipped win rates from 4% to between 64% and 80%. Bigger advantages added almost nothing on top, which tells you recognition is working as a tiebreaker of last resort rather than a standing preference, give the agent any way to discriminate and it drops the default. How much that tenth of a star buys you varies enormously by model, though: on the same test Claude switched 11% of the time where GPT switched 94%.
The agent is neither a rational buyer nor a brand loyalist. It picks whatever looks marginally better in the data it can actually reach.
Note: Keep in mind this is an unreviewed preprint, and its "unknown" brands were engineered to be names the models had never seen, which is not the same thing as being a small brand. Trust the directions here, not the decimals.
What merchants need to fix
Make specs complete. The agent weighs structured attributes far more heavily than description copy. Every variant needs machine-readable fields: dimensions, materials, compatibility, and the limits: what it doesn't fit, what it isn't for. A missing field reads as a zero.
Treat review volume as a distribution channel. A tenth of a star separates invisible from chosen. Review velocity used to be a conversion lever; it's now upstream of whether you get shown at all.
Replace slogans with evidence. "Trusted by thousands" is noise to an agent. A named customer with a measured result is not. And because fabricated authority currently works so well, verification will tighten fast, so build the claims you can actually stand behind.
Get retrievable. The incumbent advantage collapsed the moment retrieval entered the loop. If you're not the category leader, that's the opening. If you are, that's the threat. Either way it means product pages, comparisons and honest specifics an agent can pull, rather than a homepage it will never load.
Keep a channel open for the questions data can't answer. An agent that can't confirm fit, stock or a return window doesn't ask twice, it moves on. We can speak to that directly: Nuvoo now has its AI Agent handling 70% of chat sales.
The moment a buyer decides has always been the moment worth investing. That's the argument we've been making about service, and it stands. And even though your next buyer may not be a person, you should sell to it anyway.