Persuasion backfires when the buyer is an agent

Add a countdown timer to the checkout page. Put “Only 3 left in stock” under the buy button. Cross out the old price so the new one looks like a break. These moves are the bread and butter of selling online, and they work, because they push on how people actually decide.

Now imagine the buyer isn’t a person. It’s an AI assistant doing the shopping on someone’s behalf, told to find the best option and come back with one. The countdown means nothing to it. The scarcity line means nothing. And the strike-through price might make it trust you less.

Here’s what’s changed in the world, in plain terms. AI assistants are moving from recommending things to buying them. Ask one to pick a project-management tool, a hotel, or a supplier, and increasingly it doesn’t hand you ten links. It narrows the field itself, or makes the call outright. When it does, the human who used to weigh your pitch may never see it.

That shift breaks an assumption most of us have been carrying: that winning in the AI era means getting surfaced. Get quoted by the model, show up in the answer, be the recommended option. The whole craft has been about visibility.

Getting surfaced and getting chosen run on opposite incentives. The first rewards persuasion. The second can punish it. And when an AI agent is the buyer, you’re playing the second game.

Look at the evidence. In Harvard Business Review, in a piece dated July 6, 2026, researchers Graham Kenny and Ganna Pogrebna describe what happens when AI agents act as what they call autonomous purchasing intermediaries, assistants that don’t just surface options but select or narrow them. Their finding: agents weight objective attributes. Performance specs. Structured pricing. Factual differences you can check. The psychological triggers that a decade of e-commerce built its funnels on, things like urgency, scarcity, and anchoring, don’t reliably move an agent, and in some setups they lower your odds of being picked. That last part is what they report, from early cases rather than a controlled trial, so hold it as a signal, not a law.

Sit with that second half anyway. The tactics don’t just go quiet on a machine. They can score against you. A strike-through price is a persuasion signal, and to a system built to weigh verifiable facts, a persuasion signal can read as noise, or as a reason to discount you.

Now the discovery game, the one the field has already mapped. The software company PostHog documented 41x growth in traffic from LLM recommendations over two years, and their guide to answer engine optimization, the craft of shaping content so an AI assistant quotes you, lands on one discipline: citability. Clear subheadings. Conversational language. A direct answer in the opening sentence, so you’re the fragment the model pulls. Natalia Amorim, who wrote that guide for PostHog, is right about it. That advice is about being mentioned.

Being mentioned and being selected are not the same job, and they pull your content in opposite directions. Citability rewards the rhetorical: the quotable line, the warm phrasing, the confident claim. Machine-legibility rewards the literal: the spec sheet, the number, the structured field with the rhetoric stripped out. A team that polishes its copy to be maximally quotable can sand off exactly the terse, checkable detail an agent needs to compare it fairly. You can get better at one while getting worse at the other.

Watch what companies actually did about it. Kenny and Pogrebna report three cases. A boutique hotel stopped optimizing for human attention and started correcting the AI’s description of itself, fixing pricing and amenity details so the model represented the place accurately. A B2B software firm dropped its periodic review-management cycle and began monitoring AI-generated perception of its brand in real time, treating the model’s account of the company as a signal to watch rather than a channel to win. A manufacturer used AI-generated inquiries as a filter, screening leads before spending engineering time on quotes. The funnel ran backward: the agent screened them.

None of those companies was trying to get mentioned more. Each was fixing what the machine believed about it, which is a correctness problem. The old instinct, get surfaced, would have walked right past it.

The honest counter is that the citability advice was never wrong, just bounded. In the discovery regime, where a human still makes the final call, persuasion works exactly as before. Scarcity and anchoring push on the person downstream, right where they always did. If your buyers are people who use AI to build a shortlist and then decide for themselves, keep your funnel. Nothing here retires it.

And the weaker point I have to be straight about: this rests on one HBR article, and every claim in it is reported, not independently verified. One source describing a new behavior is a lead, not a settled fact. The three cases are examples chosen to illustrate, not a controlled result. The pipeline that surfaced this filed it as something to watch, not a conclusion, and that’s the right call.

So the question I can’t yet close is the one that decides how much this matters. Are the persuasion signals ignored, or penalized? Those are different worlds. If an agent simply skips your countdown timer, you’ve lost nothing. Your other attributes still get weighed, and you adapt whenever you like. If the timer measurably lowers your selection odds against a competitor who just states plain specs, the sign has flipped, and the marketing reflex to add one more nudge is now subtracting.

The test is specific enough to run. Put a product dressed in human-tuned persuasion, the urgency and strike-through and scarcity, against the same product described in neutral structured specs, and let an agent choose between them across many trials. If the persuasion version loses, a new discipline is born, with its own number to move: selection rate, sitting where citation rate used to. And probably its own owner, because the person who writes quotable copy is rarely the person who maintains a clean, comparable spec.

So there’s a clean experiment sitting in front of any team that wants it. Point an agent at your own checkout page, many times over, and watch which version it buys. If the plain one wins, you’ve found the edge of the old playbook and the first inch of a new one. The behavior is already live. The scoreboard is the only thing we’re still waiting on.