AI decides which products shoppers see—then filters them

AI product – Online shopping used to start with a product search. Now, many shoppers begin with a goal—and AI responds with a short list of recommendations. But brands aren’t judged the way they are in traditional search. Products must first pass an eligibility filter, the

A shopper doesn’t type “best shampoo” anymore. She starts with the problem: a sensitive scalp after a long week, under a certain price. With a GenAI tool like ChatGPT. Claude. or Gemini. the process can feel like a conversation—follow-up questions about what the child likes to eat. texture preferences. and dietary restrictions—or the specifics of what “sensitive” means for a product category.

The effect on commerce is immediate and unsettling for the old playbook. Research is being compressed into dialogue. and discovery is shifting upstream: from “capturing demand based on product searches” to “creating demand by recommending products shoppers might not even realize they need.” Nearly half—45%—of consumers globally now shop with AI. The challenge for brands is matching what AI needs to reliably recommend them.

Two hurdles decide whether a product shows up at all. Many assume AI simply reads keywords from a prompt and ranks products whose descriptions include those terms. In this model of shopping, that isn’t how recommendations work. It’s a recommendation system built on semantics, constraints, and authority—and products must clear two stages. If a brand fails the first hurdle, the second never matters.

Stage 1 is entry into the consideration set. Before ranking happens. an AI model determines which products belong “in the conversation.” If the shopper asks for the best shampoo for a sensitive scalp under a certain price. the system doesn’t begin with every brand and then sort. It identifies the relevant category neighborhood, applies shopper constraints and filters for attribute-level fit. If a shampoo brand hasn’t clearly positioned its products in the scalp-care category—or hasn’t structured its data to address sensitivity. ingredients. and price range—it can be filtered out before any brand authority is evaluated.

Stage 2 is rising in the ranking. Once products are eligible, AI decides which ones deserve to top the list. That matters because AI typically recommends only the top three to eight products in its ranking. In this stage. models evaluate trustworthiness indicators such as third-party testing and certifications. consistent product data across brand sites. retailer feeds. marketplaces. credible reviews. and media mentions. One peer-reviewed study found that structured, AI-ready content can achieve 40% higher visibility in GenAI responses. Trust signals matter—but only for products that have already passed stage 1.

There’s a wrinkle that can make large brands uncomfortable. Large brands can be at a disadvantage because their representation is broad and diffuse across many product types and categories. AI models don’t look for a halo. They look for the “right SKU with the right clearly documented attributes.”

Under the hood. the way AI researches products helps explain why this is so hard to game with simple keyword tactics. When a shopper enters a prompt. the model breaks the request into many smaller queries in a fan-out process and then launches those across the web and structured data sources. The AI synthesizes what it finds, identifies gaps, and often runs another round of queries to validate claims. Those fan-out queries often won’t resemble the original prompt. The model translates user intent into the attribute signals and credibility checks it needs.

That’s why focusing only on prompt optimization doesn’t work. A shopper can express a need dozens of ways—“warm jacket” or “insulated coat for commuting”—but the intent collapses into the same underlying requirements. Trying to anticipate every shopper’s phrasing becomes a dead end. Instead. brands need to consistently structure product data and trust signals across the web so AI’s fan-out queries can find and verify them. When product data is missing, inconsistent, or unverifiable, it turns into a trust deficit for AI models.

What brands should do now is shift from asking. in effect. “Did I appear for this prompt?” to asking “Why was I filtered out. and what would move me into eligibility?” The path is about providing clear signals for what a product is. who it’s for. and what constraints it satisfies. It also requires structured evidence that AI can find and verify—ingredients, certifications, and use cases.

A concrete example shows the gap. For a brand that wants to be recommended for a camping trip. describing a product as a “20. 000 BTU stove” isn’t enough on its own. The data needs to convey that it works for car camping, serves two to four people, and boils water quickly. The same standard applies across categories like beauty, food, home, and wellness.

A consistent digital footprint is also central. AI models look beyond owned product pages. scanning retail sites. reviews. media. forums. and other third-party sources to validate both whether a product belongs in a recommendation and whether its claims hold up elsewhere. Reddit is highlighted as a clear example: in Novi’s analysis. it was ChatGPT’s most-cited source for product recommendation research. The point isn’t that AI rewards upvotes. It looks for clear, useful evidence around use cases, comparisons, and category language.

Put together. the takeaway is blunt: as AI increasingly mediates consumer choices. brands that build clean. verified. well-structured data can have an advantage—and those that can’t map products to shoppers’ constraints or back up claims risk being filtered out before they ever reach the ranking. The stakes aren’t theoretical. With 45% of consumers globally now shopping with AI, the demand shift isn’t coming later. It’s already changing how products earn their place.

Kimberly Shenk is cofounder and CEO of Novi.

AI shopping GenAI recommendations Novi product data infrastructure ecommerce discovery trust signals structured content consumer behavior Stage 1 consideration set Stage 2 ranking