Tech research
When the customer
is a bot.
AI agents now search, compare and choose for people. The newest research says they are not cold, rational buyers. They are more impressionable than we are, and that changes how a business gets picked.
Picture the next customer who looks you up. It reads every word on your page in a second, compares you with nine competitors, and never gets tired or bored. It is also, the research suggests, strangely easy to sway.
AI agents, software that searches, compares, and books or buys on someone’s behalf, are moving from demos into daily life. Adobe counted a 693.4% rise in traffic from AI tools to US retail sites over the holiday shopping season, compared with the year before, and it expects that traffic to more than double again over the coming holidays. The base is still small. The direction is not.
So how do these new customers decide? A wave of new studies has started to put AI shoppers through controlled experiments, and the results matter to anyone who sells anything.
A shopper with exaggerated instincts
Researchers at MIT and Dartmouth built a realistic online shop and sent 17 AI models shopping in it, changing one thing at a time: the price, the star rating, the order of the products, a line of marketing copy.
When one of two comparable products had the higher rating, 14 of the 17 models picked it 30 to 80 percentage points more often. People shown the same choices moved by about 5 points. A single line such as “This product is a best seller!” or “Buy 1 Get 1 Free” shifted the agents’ choices by 10 to 60 points, even with ratings and prices matched.
The researchers call the agents “strongly biased choosers”. They respond to the same cues we do, only much harder. For a business, that cuts both ways. Your ratings and the plain facts on your page count for more with a bot than with a person, and so do your competitors’.
The first answer wins
Microsoft Research went further and built a whole simulated market, where AI customers looked for Mexican restaurants and contractors and AI businesses competed to serve them. Every model tested showed the same habit: it took the first offer. First proposals were chosen 60% to 100% of the time, and the business that answered first was picked 10 to 30 times as often as the one that answered third. Speed, not quality, decided who won.
More choice did not fix it. For some models, showing the customer’s agent far more options made its picks much worse, not better. The encouraging part: the most capable models mostly shrugged off the manipulation the researchers planted, from fake credentials to psychological pressure, while older and smaller models fell for it.
One fake page is enough
AI shopping assistants lean on the web, and the web can be polluted. Researchers at the Chinese University of Hong Kong and EPFL rewrote a few of the pages an AI recommender retrieves, turning a real product into a fake one, and counted how often the AI recommended the fake.
Every one of the 12 models they tested was vulnerable. A single polluted page got the fake recommended up to 27% of the time. Polluting the top three pages raised that to 73.8%. Dining was the most easily fooled category for two thirds of the models, and in 74% of searches, a page anyone could write, such as a forum post or a blog, sat among the top three sources. Thinking harder did not help. Models that reasoned step by step often invented social proof to justify the fake.
When the assistant has a sponsor
Then there is the question of whose side the agent is on. Researchers at Princeton and the University of Washington tested 23 AI models in situations where a sponsor’s product cost more but was otherwise the same. All but five recommended the pricier sponsored option more than half the time. One recommended a sponsored product almost twice as expensive 83% of the time. Another pushed sponsored options into a purchase already under way 94% of the time. How often they did it depended on how wealthy the user seemed.
As AI assistants add advertising, being recommended may become partly something you pay for, the way search results did.
Sellers are already writing for machines
Some sellers have noticed, whether they meant to or not. Economists at Chicago Booth analyzed more than six million Etsy listings and found that after ChatGPT’s release, listings carried significantly more of the information AI agents use to decide what to show, 0.143 bits out of a maximum possible 0.355. A study with people found little to no change in what humans got from the same listings. Sellers are rewriting for machines, without making anything worse for people.
How to be the one the bot picks
- Answer first. In Microsoft’s market, the first offer won 60% to 100% of the time. Online booking, instant quotes and live availability let an agent get a yes from you in seconds.
- Treat reviews as a ranking signal. Agents weigh star ratings far more heavily than people do. Keep genuine reviews coming, and show ratings, prices and key terms as text on your pages.
- Say it plainly, and say all of it. What you sell, what it costs, where you work, when you are open, what is in stock. Agents shortlist on facts they can read. Make true claims such as awards explicit, and never fake them.
- Watch what others say about you. One polluted page can mislead an AI. Ask the main assistants about your business regularly, and report fake pages and reviews.
- Expect sponsored answers. As assistants add ads, decide deliberately whether to buy in, and keep an eye on how you are described when you do not.
- Start small, move early. AI referrals are still a small channel, growing fast. Track them on their own, and move first if customers research what you sell before they buy.
Agents will not replace your customers, but they will increasingly stand between you and them. The businesses that are easy for a machine to read, trust and reach will be the ones it recommends. That is what our website design and AI search optimization teams build. If you want a clear plan, book a strategy call.
The research
Every number in this story comes from one of these sources. Working papers, reports and industry data are marked, because other researchers have not reviewed them.
- AI shoppers over the holidays By Adobe Digital Insights Read the study (opens in a new tab)
- The forecast for the coming holidays By Adobe Digital Insights Read the study (opens in a new tab)
- How AI shoppers react to ratings, prices and nudges By Manuel Cherep, Chengtian Ma and colleagues Read the study (opens in a new tab)
- A simulated market of AI buyers and sellers By Gagan Bansal, Wenyue Hua and colleagues Read the study (opens in a new tab)
- Chatbots, sponsors and conflicts of interest By Addison J. Wu, Ryan Liu and colleagues Read the study (opens in a new tab)
- How one fake page misleads AI recommenders By Minghao Luo and Liang Chen Read the study (opens in a new tab)
- Sellers are already writing for machines By Giulio Frey and Kawin Ethayarajh Read the study (opens in a new tab)