What I Change Before an Agent Shops for Us

I would earn third-party proof, a machine-readable identity, and one consistent story across sources before I spent another quarter on promotional tactics aimed at people. Harvard Business Review's May 2026 research tested eight common promotional mechanisms across four models and found that most of those tactics had no effect on AI shopping agents or backfired. The agent is comparing evidence. The buyer may never see the landing page the tactic was built for.

The operating method for getting into that consideration set is How to Get Your Brand Recommended by AI Shopping Agents. That page is the deeper resource, the method. This essay is how I would use it as a founder: which of the three signals I would fund in which order, which human tactic I would stop, and what result would put persuasion back in the plan.

The buyer shift I would budget against

Accenture's Agentic Commerce report found that 90% of frequent AI users in North America would switch from a previously preferred brand if their assistant recommended a better alternative. Up to 45% of shoppers are expected to shift at least half of their commerce activity into agent-mediated systems within two years. A quarter of executives in that research say AI agents will be their top audience for content within three years, ahead of search engines.

NielsenIQ's research describes the same move as a change in the path: discovery, ranking, and recommendation happening inside the agent, rather than as a search the person conducts and then judges.

BCG reports that AI traffic to retail sites rose nearly 700% year over year in the 2025 holiday season. Shoppers who start through an agent spend 32% more time on site, browse 10% more pages, and bounce 27% less. I would treat those sessions as already filtered. The filtering happened before the human arrived, which means the work that matters is whatever the agent used to build the short list.

Evertune's series puts a clock on ChatGPT specifically. Shopping experiences were triggered in 87% of relevant responses by February 2026, up from 8.3% in October 2025. That is a tenfold increase in five months. I would not wait for a rounder anniversary to decide whether our proof is inside the set those experiences draw from. A tenfold change in the trigger rate is a reason to audit the consideration set this quarter.

Signal one: proof the agent can verify

The HBR test is the clearest account I have of what the agent responds to. The mechanisms included scarcity cues, countdown timers, strike-through pricing, vouchers, and bundles, run on GPT-4.1-mini, GPT-5, Gemini 2.5 Pro, and Gemini 2.5 Flash Lite. Most of the classic persuasion tactics failed or reversed. The more capable reasoning models looked skeptical of overt persuasion.

Two signals worked in the direction a merchant would hope. Star ratings raised selection. Competitive pricing cut selection when the price was inflated: the agent noticed the inflation instead of accepting a framed "value." I would keep ratings and a defensible price on the list of things to maintain. I would take countdown timers and strike-through theater off the list of things we build for this channel.

The way in is third-party credibility the agent can check: reviews, editorial coverage, and validation that does not live only on our domain. Evertune's companion finding is that more than 60% of ChatGPT responses in this line of research originate from base-model knowledge rather than a live search. The agent is not required to discover us at the moment of the question. A large share of the opinion was formed from what the training record already contained, and that record is heavy with third-party editorial sources. I would fund the sentence that could have been in that record, and I would stop funding the timer.

Signal two: an identity the agent can parse

Uberall's work with Jes Scholz found that a new user entering a category gets the market leader by default, and that the agent's "market leader" is a function of reviews, citations, and structured presence, not of offline share. A smaller brand that is cleaner in the structured layer can be the leader inside the agent while a larger brand is harder to parse. I would take that as a reason to make our attributes boringly consistent: name, category, and the few facts a feed can state without a slogan.

Commercetools reports that 65% of consumers trust AI to compare prices and 14% trust it to place the order. The gap is the current window. Agents are informing and shortlisting more than they are checking out. The brand with the cleaner machine-readable identity wins the short list during the years when the human still confirms. I would use the window. I would not assume the 14% stays 14%.

Structured product data, the same entity signals on every platform that carries us, verified reviews, and a stable brand-to-category association are the moat I would invest in. I would not describe that work as a search-ranking project. It is identity, at the level an agent reads.

Signal three: the same story everywhere the agent looks

Agents synthesize training data, live retrieval, structured feeds, editorial coverage, reviews, and product data. The arXiv paper on commercial persuasion in AI-mediated conversations found that behavioral shifts in how agents process promotional information account for 90% of the performance gap between overt persuasion and authentic product information. Consistent signals across independent sources strengthen the category association. Conflicting or manipulative signals get the brand downgraded.

A perfect product feed can still lose if the editorial record, the reviews, and the third-party mentions tell another story. I would pick one description and make the independent sources match it, rather than letting each channel optimize its own adjective.

Signal	What I would put in market	What I would stop building for this channel
Third-party authority	Reviews, editorial coverage, independent validation	Scarcity cues, countdown timers, self-claims
Machine-readable identity	Structured attributes, one entity, stable category	Visual-only branding, unstructured slogans
Cross-source consistency	The same facts in editorial, reviews, feeds, and coverage	Channel-specific claims, framed inflation

I would translate the three signals into owners, because a signal nobody owns becomes a slide. Editorial and reviews own signal one, and the test is a sentence a stranger can check, not a logo. The team that ships the catalog owns signal two, and the test is whether the attributes match the sentence. Whoever runs campaigns owns signal three, and the test is a ban on channel-specific claims that contradict the feed. Accenture's 90% willingness to switch, the 45% shift of at least half of commerce, and the 25% of executives who expect agents to be the top audience are the reasons those owners exist. They are not three different strategies. BCG's nearly 700% holiday traffic jump, with 32% more time, 10% more pages, and 27% less bounce, is what the short list is worth after the agent has already filtered. Evertune's move from 8.3% to 87% in five months is the clock. I would review the three owners against that clock monthly, and I would not add a fourth owner called "promotions" unless a replication of the HBR test says the timers started working.

The decision

I would audit the three signals in that order and fund the first one that is thin. If we lack third-party sentences, I would buy coverage and real reviews before any feed project. If the sentences exist and the attributes disagree, I would fix the feed and the name. If both are sound and the channels contradict each other, I would cut the contradictory campaign. Paid placements and urgency cues stay in the plan only where a human still sees them. I would not count them as agent strategy.

The share of agent recommendations follows the share of sources the agent will cite. The place I keep that connection, from a placement to a recommendation a buyer never had to search for, is Machine Relations. The name I use for the identity work in the second signal, so a feed project does not get mistaken for a branding project, is the entity chain.

Scope: brands whose buyers can hand comparison or purchase to an assistant, using the HBR models named above, Accenture's North American frequent-AI-user base, the NielsenIQ and BCG retail figures, Evertune's ChatGPT shopping-trigger series, and the Commercetools trust split. It does not tell a founder to abandon human marketing on surfaces an agent never reads. It says those surfaces are a different budget, with a different success condition.

What would change my mind: a replication of the HBR design, across models of the same class, in which scarcity, countdowns, and strike-through pricing moved selection in the merchant's favor often enough to matter. I would also put promotional mechanics back into the agent plan if Evertune's "more than 60% from base-model knowledge" finding reversed and live, on-site persuasion started dominating responses. I would not put them back because a human team preferred the countdown, or because the 8.3% to 87% trigger jump made the quarter feel urgent. Urgency argues for fixing the three signals. It does not argue for rebuilding the tactic the test rejected.

FAQ

How do I expect a shopping agent to choose?

From the three signals in the HBR and companion studies: third-party authority, a parseable identity, and agreement across sources. Star ratings moved selection up. Inflated prices moved it down. Scarcity cues, countdowns, strike-throughs, vouchers, and bundles, tested on GPT-4.1-mini, GPT-5, Gemini 2.5 Pro, and Gemini 2.5 Flash Lite, failed or backfired. More capable models were the more skeptical ones.

Will paid advertising show up inside the agent's choice?

Evertune found more than 60% of the ChatGPT responses in this research coming from base-model knowledge, which is shaped by editorial and other third-party sources in training, not by the ad the human would have seen. I would keep paid media for human attention. I would not forecast agent recommendations from it.

What is changing in the path to purchase, in the figures I would actually cite?

Accenture: 90% of frequent AI users in North America would switch on an assistant's recommendation, up to 45% of shoppers may move at least half their commerce into agent-mediated systems within two years, and 25% of executives expect agents to be the top content audience within three years. BCG: AI retail traffic up nearly 700% year over year in the 2025 holidays, with agent-started sessions at 32% more time, 10% more pages, and 27% lower bounce. Evertune: ChatGPT shopping triggers from 8.3% of relevant responses in October 2025 to 87% by February 2026. Commercetools: 65% trust a comparison, 14% trust the agent to order. I would cite each figure with its publisher, and I would not average them.

What would I fix first if all three signals are weak?

Third-party proof. The HBR result says the agent does not accept our persuasion, and the 60% base-knowledge figure says a large share of the opinion is already formed from other people's pages. Identity and consistency matter, and they can be built in parallel once a third party has said something specific. A feed that states a claim nobody else corroborates is a well-structured claim. It is still ours alone.