AI Is Changing Ecommerce, But Not How You Think

The conversation nobody is having

Ask someone in ecommerce what AI is good for and you'll hear roughly the same list: customer support automation, product recommendations, ad targeting, fraud detection. These are real applications with real ROI, and companies are right to pursue them.

But there's a more fundamental problem that rarely makes the list: people struggle to figure out what they want to buy. Not because the products don't exist. Not because the prices are wrong. Because the cognitive overhead of modern product discovery is genuinely high, and almost nothing in the current stack is designed to reduce it.

That is the more interesting opportunity. And it requires thinking differently about what AI is actually for.

Search, recommendation, and decision support are not the same thing

The ecommerce industry tends to conflate three separate activities: search, recommendation, and decision support. They feel related because they all involve getting a user to a product. But they serve different user needs, and conflating them leads to misallocated investment.

Search is intent retrieval. The user knows what they want, or at least knows enough to describe it. They type "black running shoes size 10 waterproof" and expect results that match those constraints. A good search engine surfaces accurate results fast. The intelligence is in the retrieval, not the interaction.

Recommendation is pattern extrapolation. The system observes what users like, infers preferences, and surfaces products the user didn't know to ask for. Collaborative filtering, embedding similarity, content-based matching — all of this is fundamentally about predicting preferences from behavioral signals. It works well when users have established patterns and when preferences are relatively stable. It works poorly for new users, novel categories, or anything that requires understanding context beyond click history.

Decision support is something different entirely. The user has a goal, often one they can't fully articulate, and they need help figuring out what will actually achieve it. Not "I want a jacket" but "I'm going to a wedding in Austin in June and I want to look sharp without being overdressed and I tend to run warm." Decision support requires understanding the user's situation, reasoning about constraints, and narrowing a large option space into a small one. It's closer to what a knowledgeable friend does than what a search engine does.

Most AI investment in ecommerce targets the first two. Decision support gets far less attention, even though it's arguably where users have the highest unmet need.

Why product discovery is still overwhelming

The dominant model for product discovery is: show the user a grid, let them filter, trust that they'll find what they want. This works reasonably well when users know exactly what they're looking for. It breaks down everywhere else.

The problem isn't a shortage of products. If anything, catalog depth has made things worse. A mid-size retailer might carry 50,000 SKUs. A marketplace might carry millions. The filtering tools available to users are blunt instruments: category, price range, brand, size. These tools help with elimination but not with selection. The user still has to do the cognitive work of evaluating options, and that work compounds as catalog size grows.

Users respond to this in predictable ways. They abandon sessions without purchasing. They buy the first thing that seems good enough, then return it. They rely on external validation — influencers, Reddit threads, friends — to reduce the perceived risk of a bad decision. They develop loyalty to brands not because the brand is always the best fit but because familiarity reduces the decision cost.

None of this is a search quality problem or a recommendation quality problem. It's a decision support problem. The user isn't finding it hard to locate products. They're finding it hard to evaluate them against their own goals.

Why conversational interfaces alone aren't enough

The obvious response is: add a chatbot. Let users describe what they want in plain language and let an LLM figure it out. This is the right instinct, but the execution often stops too early.

Generic chat interfaces layered on top of existing search and recommendation systems don't solve the underlying problem. They change the input modality without improving what happens to the input. A user can describe their situation in natural language, but if that description gets translated into a keyword query and routed through the same retrieval system that was failing them before, the output quality doesn't meaningfully improve.

There's also the context problem. A single-turn interaction doesn't capture the kind of information that makes decision support genuinely useful. "Looking for a gift for my dad" is not enough. How old is he? What does he actually do with his time? What's the budget? Has he mentioned anything he wants? How technical is he? A conversational system that asks follow-up questions and accumulates context across a session can get to useful recommendations. A system that treats every message as independent cannot.

This is not a model capability problem. The models are capable enough. It's an architecture problem. Most chat implementations in ecommerce are built as thin wrappers, not as reasoning engines with state.

What actually makes a conversational system useful

Two inputs matter more than almost anything else: the quality of product data and the richness of user context. Both are underinvested in.

Product data in most catalogs is built for search, not for reasoning. Titles are keyword-stuffed. Descriptions are supplier copy-paste. Attributes are inconsistent across categories. A language model trying to answer "is this good for someone who gets cold easily and works long hours?" cannot do much with a product description that says "Premium insulated jacket. Multiple colors available." Structured, semantically rich product data — materials, intended use cases, fit characteristics, what the product is not good for — enables a reasoning layer to actually match products to situations rather than just matching keywords.

User context is the other side of the equation. Session-level context (what has the user said, what have they clicked, what have they rejected) and account-level context (past purchases, stated preferences, price sensitivity) can substantially improve the quality of recommendations. Not because the model is doing anything exotic, but because it has more signal to reason from.

When these inputs are rich, conversational decision support stops being a novelty and starts being genuinely useful. The interaction can feel less like querying a database and more like talking to someone who knows the catalog well and is paying attention to what you've told them.

The structural shift this requires

Building toward decision support rather than just better search and recommendations requires some organizational reckoning. Product teams have spent years optimizing for discovery metrics: sessions, clicks, conversion rate. Decision support might lower session length and reduce discovery-phase browsing, which looks bad in standard dashboards even when it's actually better for the user.

It also requires owning the quality of your product data in a way most companies don't. If your catalog is the raw output of supplier feeds with minimal enrichment, you can't build reasoning on top of it. The investment in structured, consistent, semantically rich product data is unsexy and slow, but it's load-bearing for anything more sophisticated than keyword retrieval.

And it requires rethinking what you're measuring. A user who finds the right product in three conversational turns and converts is a better outcome than a user who browses 40 products and bounces. The metrics need to reflect that.

Conclusion

The headline applications of AI in ecommerce — support automation, recommendation engines, ad personalization — are real and worth building. But they're also well-understood, increasingly commoditized, and not where the most meaningful unmet user need sits.

The harder, more interesting problem is decision support: helping users figure out what they actually want, given everything they're dealing with. It requires better product data, richer user context, and conversational systems designed to accumulate both across an interaction rather than treat each turn as isolated.

This isn't about making shopping more "AI-powered" in a marketing sense. It's about doing something that most ecommerce experiences still fail to do: helping people make good decisions without exhausting them in the process. That's a tractable problem. The tools exist. The gap is in how seriously the industry has chosen to pursue it.