How AI Agents Actually Shop

AI agents can weigh more options than a person, and that part of the story is true. In this baseline, a typical trip considered about ten brands. Then it narrowed hard, usually reading only a couple of sources before settling on a single recommendation.

Getting into the first set is a visibility game. The narrowing from ten to one is a different game, and it runs on signals a brand can change: product details, review signals, comparison language, and pages an agent can load without friction.

It is easy to assume AI answers come from memory. In about three of four trips, the agent went out and read live pages before recommending. About one in four leaned mainly on what the model already knew about brands.

Reachable site pages, marketplace listings, and trusted third-party sources become part of the recommendation surface. For commerce teams, the state of the page today is not just background reputation. It is evidence the agent can load and compare right now.

Two things slowed agents down with the same result: access limits and unclear page paths. In 64% of trips, the agent encountered an access limit somewhere in the path, and in almost one in five the page never resolved cleanly.

The agent usually kept going. It read another source, shifted to a review page, or moved to a marketplace surface. The brand was not necessarily gone, but the brand-owned page stopped doing the job.

For higher-consideration and larger purchases, independent expert reviews were up to 30% of the primary cited sources in recommendations. For faster, everyday buys, that share stayed close to zero and agents leaned more on brand pages, marketplaces, and product details.

In considered categories, trusted reviews can put a brand in the set before the agent evaluates the brand site. In everyday categories, the marketplace listing and product surface carry more of the work.

Across every category, features and specifications were the biggest factor in the agent's reasoning in roughly 70% of trips. The next factor changed by category: price, stock, or review weight.

Do not chase a universal formula. Give the agent the core product facts first, then make the category-specific proof easy to find: price, availability, review strength, warranty, fit, compatibility, or the factor that settles the buyer's question. The wrong second message is easy for an agent to ignore.