The Nudge Era Is Over. Now What?
When an AI agent does the first pass of a purchase, the cues built for a hurried human reader may not land where marketers expect.

Scarcity badges, anchors and social proof were built for a person scrolling quickly, and as more shoppers ask AI to help them choose, the first visitor to a product page may more often be an agent working for that person, which turns next year’s marketing plan into a choice of lens.
Picture a product page that any marketer would recognise, with a small red badge saying “Only 3 left”, a crossed-out price above the real one, a line telling you that 2,000 people bought this item last week and a timer counting down to midnight, all designed for someone on a phone at ten in the evening with half an eye on the television, who will take each cue at face value rather than stop to check it. Now imagine that the first visitor to the page is an AI agent that has been asked to find a quiet dishwasher under £500 and report back with three options, and ask yourself what the countdown timer is doing for you.
In February, the behavioural scientist Lea Karam, founder and chief executive of the behavioural science lab Mindscope, told Performance Marketing World that “The ‘Nudge era’ is long gone.” Her argument was that behavioural science had moved beyond prompts towards “infrastructures and IPs, across platforms, algorithms, and culture”, and that “Many marketers are still using old playbooks for a very different behavioural environment.” She was not talking specifically about shopping agents, so carrying her point over to them is my reading. The STAR Framework starts from motivation and from the way people process what is in front of them, and from there the nudge playbook looks like a set of tools built for one way of processing. If a growing share of first passes is done by something that processes differently, a 2027 plan becomes partly a choice of lens, and STAR’s Realist and Adventurer lenses pull in different directions. None of the research discussed here tested STAR, and everything I say about STAR is interpretation.
Built for a hurried reader
STAR’s Dual Process pillar draws on research describing how fast, intuitive processing interacts with slower, more deliberate processing, both of which everyone uses. Jonathan Evans and Keith Stanovich describe rapid autonomous processes that yield default responses unless higher-order reasoning intervenes, and they prefer to speak of Type 1 and Type 2 processing rather than two separate systems in the head, and the classic nudges sit comfortably on the fast side of that distinction. Amos Tversky and Daniel Kahneman described three heuristics, among them adjustment from an anchor, as “highly economical and usually effective” while leading to “systematic and predictable errors”, which is roughly what a crossed-out price is hoping for. In a much-cited experiment, Worchel, Lee and Adewole found that participants rated cookies as more desirable when they were in scarce supply, and more valuable when a plentiful supply had dwindled, particularly when the drop was put down to high demand, which is the scarcity badge in its original, edible form.
On the usual account, these cues tend to work, when they work, because a person does not stop to interrogate them, which is the quiet assumption under the whole playbook and precisely the one an agent may not share.
The agent arrives
McKinsey estimated in October 2025 that by 2030 agentic commerce could orchestrate between $900 billion and $1 trillion of revenue in US business-to-consumer retail, a figure that covers goods only and rests on “moderate assumptions about merchant readiness”. In August, Boston Consulting Group published Global Consumers Have Moved On. Has Your Growth Strategy Caught Up?, based on a survey of more than 13,000 consumers in 12 markets, and reported that 31% use AI at least occasionally in their purchase journeys, roughly triple the level of 18 months earlier. It found that 19% are what it calls AI loyalists, who rely on AI regularly, and that 70% of those loyalists ultimately buy what the AI recommends, which BCG’s own text puts at 13% of everyone surveyed. BCG’s summary of the change is blunt: “increasingly, AI assembles that shortlist on the consumer’s behalf.”
That is self-reported behaviour, and on BCG’s figures most of those surveyed did not report using AI in a purchase at all, and using AI somewhere in a purchase journey is not the same as an agent reading the product page, but it is enough to suggest that the first reader may change in a meaningful share of journeys while next year’s plans are being drawn up.
What the agent noticed
The most useful evidence I have found on how an agent actually responds to the nudge toolkit comes from Amine Allouah, Omar Besbes, Josué Figueroa, Yash Kanoria and Akshit Kumar, of MyCustomAI, Columbia Business School and Yale University. Their paper, What Is Your AI Agent Buying?, presents ACES, an Agentic e-CommercE Simulator that pairs an AI shopping agent with a mock marketplace in which product positions, prices, ratings, reviews and badges are randomised, so that their effect on an agent’s choice can be measured causally. In the main experiments with Claude Sonnet 4, GPT-4.1 and Gemini 2.5 Flash, an “Only X Remaining” scarcity tag had an effect that was “weakly negative or statistically indistinguishable from zero”. A platform’s “Overall Pick” endorsement, by contrast, lifted a product with a baseline selection probability of 10% to 24.3%, 19.9% and 42.6% respectively, while a “Sponsored” tag pushed the same baseline down to between 7.9% and 8.9%. All three models favoured cheaper, better-rated and more widely reviewed products, which I would call social proof of a more checkable kind, although the strength of those preferences varied sharply between models, and every model showed position biases that differed from one provider to another and persisted even when the agent was given a plain list of products with no images.
The limits matter as much as the findings, because this was a simulated marketplace, testing particular models at particular moments, and the authors report that a model update could “drastically reshuffle market shares”, so none of these numbers should be read as a permanent property of agents in real shops. The study did not test human shoppers, and the authors note that the agents’ response to endorsements echoes the behaviour of human buyers. What it does suggest, in my reading, is that an agent need not respond to a product page in the way the classic heuristics assumed of a hurried human, while being far from neutral itself. It appears to weigh some cues as information and treat others as noise or a reason for suspicion, and it has sensitivities of its own, such as position, which can change overnight.
Two lenses on the same plan
Regulatory Focus Theory, another of STAR’s pillars, gives the planning question its shape. E. Tory Higgins distinguished self-regulation with a promotion focus, concerned with accomplishments and aspirations, from self-regulation with a prevention focus, concerned with safety and responsibilities. STAR’s own type-level framing, which is a framework mapping and was not tested by any of this research, links the Realist lens with prevention and the Adventurer lens with promotion. The Realist lens foregrounds security, stability and reliability, along with simplicity and ease, and it looks at the nudge playbook and sees something that has worked for years, is understood by the team and still reaches every customer who never touches an agent. The Adventurer lens foregrounds autonomy, exploration and freedom to act, and it looks at the same playbook and sees cues designed for a reader who is gradually handing over the first pass.
These are lenses for a planning meeting, with nobody in the room sorted into either, and each has a characteristic failure. Planning only through the Realist lens risks polishing a page whose first reader no longer scans it the way the page expects, while planning only through the Adventurer lens risks rebuilding everything around a position bias that the next model update quietly reverses. My reading is that the useful move is deliberate: run the plan through one lens, then the other, and notice where the answers disagree.
A toolkit for the first reader
- Map the funnel by reader. For each step from discovery to checkout, ask whether a person or an agent is likely to see it first, because a cue written for the wrong reader is wasted at best.
- Sort your cues into claims and pressures. Ratings, review counts, specifications, delivery terms and honest prices are claims an agent can weigh; countdowns and vague urgency are pressures that rely on someone not checking, and the one such cue ACES tested, a scarcity tag, made little or no difference, or slightly hurt, for the three main models, although among the newer models the same paper tested, one was put off by it and another was drawn slightly towards it.
- Audit with agents, and keep auditing. Test how the assistants your customers use describe and rank your products, then repeat the test after model updates, since the authors found preferences could reshuffle when a model changed.
- Ask the prevention question and the promotion question separately. What must this plan protect for the customers who still shop by hand, and what would it gain from being chosen by an agent? Write both answers down before choosing.
- Keep the human moment honest. The final decision often returns to a person, and the nudge at that point remains a nudge, so it deserves the same ethical scrutiny it always did.
The test question for any page in next year’s plan is a simple one: who is reading this first?
The red badge on the product page will go on saying “Only 3 left” for as long as someone leaves it there, and for the person on the sofa it may still do its job. In one simulated marketplace, though, the three main agents tested treated that badge as roughly irrelevant, or as a small reason to look elsewhere, and responded far more to the rating and the endorsement. The nudge era may or may not be over, but its first reader is changing, and the timer keeps counting down for an audience that is not always there.
If you would like to see which Mindsets tend to lead for you, the STAR Cognitive Profile Assessment at starassess.com takes 33 questions, and the assessment and its full report are both free, which may be a useful starting point for noticing whether you plan more readily through the lens of protecting what works or the lens of chasing what might.
David Chadderton is a behavioural strategist, Chief Marketing Officer and STAR Dynamics. He writes The Unoptimised Human.
References
Allouah, A., Besbes, O., Figueroa, J. D., Kanoria, Y., & Kumar, A. (2025). What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, & Emerging Implications for Agentic E-Commerce. arXiv preprint 2508.02630, version 3, 17 December 2025. A short version appears in Proceedings of the ACM Web Conference 2026, pp. 8697–8700.
Bharadwaj, A., Taylor, L., Barrios, G., Sanghi, K., Malby, A., Barton, C., Ariav, Y., Gell, J., & Portera, T. (2026, 26 August). Global Consumers Have Moved On. Has Your Growth Strategy Caught Up? Boston Consulting Group.
Evans, J. St. B. T., & Stanovich, K. E. (2013). Dual-process theories of higher cognition: Advancing the debate. Perspectives on Psychological Science, 8(3), 223–241.
Higgins, E. T. (1997). Beyond pleasure and pain. American Psychologist, 52(12), 1280–1300.
Langford, R. (2026, 23 February). ‘The nudge era is long gone’: Mindscope’s Lea Karam on merging behavioural science with performance marketing. Performance Marketing World.
Schumacher, K., Roberts, R., & Giebel, K. (2025, 17 October). The agentic commerce opportunity: How AI agents are ushering in a new era for consumers and merchants. McKinsey & Company.
Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, 185(4157), 1124–1131.
Worchel, S., Lee, J., & Adewole, A. (1975). Effects of supply and demand on ratings of object value. Journal of Personality and Social Psychology, 32(5), 906–914.
Eight minutes. Thirty-three questions.
A precise picture of how you operate: motivations, decision style, and blind spots. Free to start.
- S
- T
- A
- R
The canonical source.
Essays, when there is something worth sending. Choose the subjects you want. I will ask you to confirm the address before I write.