For the last 20 years, eCommerce has been shaped by algorithms.
Search algorithms. Recommendation algorithms. Advertising algorithms. They decide which products appear first, which brands get discovered and which offers are most likely to convert. Every one of them lives behind the same basic interface: a search bar, a set of filters and a ranked page of results.
Now we're being told that AI agents will transform shopping. Which raises an obvious question:
Is an agent actually that different from an algorithm?
It's a fair thing to ask. Both sit between a shopper and a catalog. Both make choices on the shopper's behalf. And both are only as good as what they know about the products they're choosing from. If every shift in commerce eventually comes down to where the shopper's attention moves (opens in a new tab), this one deserves a harder look than the hype cycle usually allows.
But the difference matters, and it shows up in where the thinking happens. What can an LLM-powered agent do that traditional search and filtering can't?
Ranking is not reasoning
Here's how discovery works on a traditional search-and-filter site:
100,000 products → search algorithm → 100 results → human investigates 10 → considers 3 → buys 1
The search algorithm ranks the world. The human still does most of the reasoning.
Agent-assisted commerce moves that work:
100,000 products → agent expands the request into multiple queries → retrieves across product data, reviews and web sources → re-ranks using shopper context → presents 3 choices with rationale → human validates and buys 1
Take a simple purchase: a 65-inch TV under $1,000. Site search does a great job ranking products using price, ratings, brand and specifications. Four filters and a sort order get you there. There's nothing an agent adds that a good filter set doesn't already handle.
Now consider a different request:
I need a stroller for two children that fits in my trunk, works on cobblestones, can be carried up three flights of stairs and is small enough for European trains.
No retailer has ever built a filter for that. Type it into a search bar and you'll get back results for "stroller."
An LLM-powered agent can interpret the objective, reason across multiple attributes and make trade-offs on the shopper's behalf. That isn't a hypothetical behavior waiting on some future model. Shoppers are already using AI this way: in Europe, 63% of consumers now use AI to compare brands, models, prices and reviews (opens in a new tab), and 55% use it to learn about a category before they buy. Comparison and evaluation are reasoning tasks. Consumers have already handed them over — and they're doing it before they ever reach a retailer's search bar.
When the best answer isn't the product you asked for
And eventually agents will go further.
Consider AlphaGo's famous Move 37 (opens in a new tab) against Lee Sedol. In 2016, DeepMind's AI played a five-game match against the world's strongest player of Go, a board game with more legal positions than there are atoms in the observable universe. On the 37th move of the second game, AlphaGo placed a stone where no professional would have — commentators assumed it was a mistake, and it went on to win the game.
Expert Go players initially struggled to understand the move because it sat far outside conventional human play.
AlphaGo wasn't trying to replicate how humans played Go. It was optimizing for the outcome: winning.
The most interesting shopping agents will do the same. You ask for the best air conditioner for your bedroom.
A good agent might conclude that you shouldn't buy an air conditioner at all. Perhaps an awning, whole-house fan or dehumidifier better solves the underlying problem.
Search algorithms optimize selection within the category you named. Agents can increasingly optimize outcomes, category included.
That's a meaningful shift for anyone whose commercial model assumes the shopper arrives already committed to a category.
A Move 37 requires a different data diet for eCommerce
But an agent can't suggest an alternative solution — or know that a stroller struggles on cobblestones — if it's only reading standard catalog specs.
Neither can reason about information it doesn't have.
So while the shopping experience moves from filters and rankings to conversations and reasoning, the underlying requirement stays remarkably consistent:
Clean, rich product data — both structured and unstructured.
The structured catalog tells you what a product is: its dimensions, materials, specifications, compatibility and price.
User-generated content (UGC) tells you what the product is like: whether the stroller is awkward on cobblestones, whether the sofa stains easily, whether the drill struggles with masonry, or whether a dress runs small.
Search and recommendation algorithms have historically relied heavily on the first, while using UGC mostly as signals such as ratings, review counts and sentiment.
Agents can reason across both. And they're already doing it. Research from Columbia Business School and Yale found that AI purchasing agents consistently favor products with both high ratings and high review volume (opens in a new tab) — while discounting sponsored tags. Reviews aren't social proof sitting below the fold anymore. They're evidence an agent uses to justify a recommendation. That reach extends beyond retailer-owned data. In a July 2026 snapshot of roughly 3.3 million citations in ChatGPT Shopping, Profound found that Reddit accounted for about one-third of citations — a reminder that the evidence around a product can live in third-party conversation, not just on the product page.
That reframes what a thin review corpus costs you. It's no longer just a conversion problem on your product detail page. It's a reason your product doesn't make the shortlist at all.
The definition of product data is getting bigger
Here's the uncomfortable part for most catalogs.
Adobe found that individual product pages across US retail sites score an average of 66% on AI content visibility (opens in a new tab) — meaning roughly a third of what's on a typical product page can't be read by the models doing the evaluating. The information exists. It's on the page. It just isn't reaching the system making the recommendation.
That gap is the real work of agentic commerce, and it's less glamorous than the interface conversation. Better product data improves decisions for both people and agents. What's changed is the technical standard: content that may have helped a page rank — even when keyword-heavy or trapped in JavaScript — may not be sufficiently structured or machine-readable for an agent to retrieve, compare and reason across.
The algorithm needs product data to rank.
The agent needs product data — and the collective experience of customers — to reason.
The interface to commerce is changing. The foundation is still product data — but its definition now spans machine-readable catalog content, customer reviews and other UGC, brand reputation, and the operational facts that make a recommendation actionable: price, availability and delivery promise.
So the question worth taking back to your team isn't whether you're ready for agents. It's simpler than that: can your catalog answer a question no search filter was ever built for?
If you can't answer that yet, that's where the work starts — and it's the work the leading eCommerce businesses have already begun (opens in a new tab).
