Why Sending Your Entire Catalog to AI Platforms Is the Wrong Move

In this week’s episode of Voices of Search, we spoke with Katie Morrow, Director of Managed Services at Products Up, about why most e-commerce teams are approaching AI commerce readiness completely backwards—and what to do instead.

Products Up processes 2.2 trillion products per month for some of the world’s largest retailers and brands, including Sephora, HP, Puma, and L’Oreal, optimizing and normalizing product data across every channel from search and social to emerging agentic commerce platforms. Katie brings 15 years of experience in product data and feed management, starting with bid management at a channel intelligence agency, then working as a shopping specialist at Google during its acquisition of Channel Intelligence, before joining Products Up more than a decade ago.

Today, Katie broke down why “send everything” is the wrong default for AI platforms, how to segment your catalog strategically for AI discovery, and what marketers should actually be measuring when traffic is no longer the primary signal.

Key Takeaways From This Episode:

• Sending your full product catalog to AI platforms creates noise that makes impact impossible to measure
• Start with a small, strategic subset of SKUs—your top performers or a single category—and expand from there
• Catalog segmentation should be driven by your KPIs, not by what’s easiest to export
• Product variants like size and color don’t always need to be exposed on every channel—match variant depth to intent
• Rich, SKU-level content around availability, pricing, and fulfillment options is increasingly what agents use to make recommendations
• Bad product data exposed to AI agents creates a bad historical presence that’s difficult to undo
• The household is now your audience, not the individual—marketers need to think about audiences of five, not one

AI Agents Are Reshaping the Discovery Phase

The conversation started with a foundational shift in how consumers find products. AI agents aren’t just a new search interface—they’re taking over the discovery phase entirely, making decisions about what shoppers see before those shoppers ever express explicit intent.

“It’s a whole new discovery phase,” Katie explained. “Maybe those that didn’t know the questions to ask are being given the keys to understanding how they’re going to make decisions in the future.”

For marketers, this creates a challenge that’s fundamentally different from optimizing for keyword-based search. Rather than ensuring a product page ranks for a specific query, brands now have to influence something far more opaque—an algorithmic layer that decides, on the shopper’s behalf, what products are even worth surfacing.

That shift extends into the household. Katie pointed out that the sphere of influence has always existed at the household level, but AI changes the scale of that influence dramatically:

• A household member researching e-bikes affects what ads appear across shared devices and platforms
• A ten-year-old asking a store when a toy will be restocked is exhibiting the same behavior that AI agents are beginning to automate
• Every device in the home is historically collecting data that bleeds across the household’s shared digital footprint

“Now I have an audience of five, not an audience of one,” Katie said. For brands and their product data strategies, this means optimizing not just for a single buyer persona, but for the range of people who may influence or complete a purchase within a single household.

The “Send Everything” Problem

One of the clearest themes across the conversation was the instinct—still prevalent across e-commerce teams—to expose the entire product catalog to every channel, including emerging AI platforms.

Katie was direct about why this is the wrong approach. “A lot of our 2.2 trillion products are still getting it wrong, and they are exposing everything. Every fuzzy sock, every toothpick.”

The problem isn’t just that low-priority products dilute catalog quality. It’s that sending everything makes it effectively impossible to measure impact. If a brand pushes 50,000 SKUs into an agentic commerce platform and doesn’t see a meaningful result, there’s no way to isolate what worked, what didn’t, or why.

Katie’s recommendation is to start small on purpose: “Start with your top 50 products or your top category and expose those. Then track from there.” rather than thinking of this method as a limitation, brands need to start looking at it as strategic discipline. 

Treating early AI commerce deployment as a controlled experiment gives teams:

• A manageable set of products to monitor daily
• A clear control group to measure against
• A foundation to build on as results come in
• The ability to isolate what’s actually driving impact before scaling

She drew a direct parallel to the old logic of site maps in SEO: where site maps were once about exposing everything to Google crawlers, LLMs don’t work the same way. They aren’t crawling the entire web the way traditional search engines did, which means the prioritization exercise for commerce becomes far more consequential—and far more specific.

Catalog Segmentation Should Follow Your KPIs

If “send everything” is the wrong move, the obvious next question is how you decide what to send. Katie’s answer starts with a question most teams skip: what are your actual KPIs?

• Revenue-focused? Expose your top revenue-driving SKUs
• New customer acquisition? Expose only net new products
• Traffic growth? Expose only your highest-click SKUs
• Category expansion? Expose only the category you’re trying to grow

This sounds simple, but it requires having a clear, current understanding of catalog performance at the SKU and category level—including how often that performance shifts. Katie noted that for many brands, top-performing products change on a seven to twenty day cadence, and catalog segments for AI platforms need to reflect that same rhythm to stay relevant.

Variant-Level Data Doesn’t Always Need to Be Exposed

She also pushed back on the assumption that variant-level data always needs to be exposed. 

For social channels operating in a visual discovery mode, sending every size and color of a men’s polo shirt creates unnecessary complexity. A shopper on Facebook isn’t searching for a size medium blue polo—they’re discovering a style they like. 

The variant-level detail matters on Google, where purchase intent is explicit, and increasingly in agentic environments where a shopper might specifically need a 28-inch inseam and wants the agent to do the narrowing for them. That’s the difference between someone who’s window shopping and someone shopping with intent to buy. 

“I finally have somebody—a friend, an agent—to tell me how to find those short pants that I’ve gone to 20 retailers looking for,” Katie said. 

The takeaway: Match the depth of your product data to the intent model of the channel. Discovery channels want hero products. High-intent and agentic channels want rich, variant-level specificity.

Dynamic Catalogs for a Trend-Driven World

A static catalog strategy also won’t hold up in an environment where consumer demand can shift in days. Katie used the example of a viral toy—something that sat on shelves for months and then suddenly became impossible to find—to illustrate how quickly product relevance can change, and how that puts real pressure on marketers to stay ahead of their own catalog.

The implications vary by category, but the pattern is consistent:

• Sporting goods retailers need to track cultural moments like March Madness and surface relevant jerseys before game day
• Fashion retailers need to sync new arrivals to what’s appearing in moments like New York Fashion Week
• Anyone selling trend-sensitive products needs the operational infrastructure to update AI platform exposure quickly—not on a quarterly roadmap

“If you don’t have a tool that helps you make those decisions quickly, you’re going to miss a lot of opportunity this year and ongoing,” Katie warned. “This is just the beginning.”

The implication for teams is that AI commerce readiness isn’t a one-time implementation. It’s an ongoing operational discipline that requires the same attention to timing and relevance that any high-performing campaign would demand.

Rich SKU-Level Content Is What Agents Actually Use

As the conversation turned to what actually influences agent recommendations, Katie pointed to something many brands overlook: the granular, functional attributes attached to individual SKUs.

The signals that agents increasingly use to resolve a purchase decision include:

• Current pricing and whether a product is on sale
• Availability in-store, online, or at a nearby location
• Fulfillment options like curbside pickup, same-day delivery, or door dash eligibility
• Whether the product is in stock right now versus available on a future date

A shopper who tells an agent they need a grill by five o’clock and don’t want to leave their car is giving the agent very specific constraints. If a brand’s product data doesn’t include fulfillment details at the SKU level, that brand won’t be recommended—regardless of how competitive its pricing is.

“Make sure your structured data has that information,” Katie said. “Is it the lowest price? Is it on sale? Is it available curbside? Can I get it delivered?”

This is where brand differentiation in AI commerce becomes less about brand equity in the traditional sense and more about data completeness. When the same barbecue grill is sold by ten different hardware stores and the product specs are identical across all of them, the differentiating signals become price, availability, and fulfillment speed—all of which need to be in the feed, accurate, and current.

The Readiness Trap: Speed Versus Strategic Pauses

Arguably, the most important part of our conversation addressed the pressure that e-commerce and marketing teams are feeling from leadership to do something visible with AI, right now.

Katie acknowledged that pressure directly—and pushed back on the reflex to move fast without a plan. “Everybody wants to know that somebody’s doing something. So we feel that pressure from the top. Now we have to pause.”

The risk of moving fast with unstructured data isn’t just that you don’t see results. It’s that you build a bad historical presence with the agents that will be making future buying decisions on your customers’ behalf. If an agent encounters poorly structured, inaccurate, or incomplete product data during its early interactions with a brand’s catalog, that shapes how it will evaluate and recommend those products going forward.

“The worst thing we can do is expose bad product content to the agents,” Katie said, “because at that point you’re building a bad historic presence with those that are making future buying decisions on our behalf.”

This is a more consequential version of the old SEO principle that bad inputs produce bad outputs. In an agentic environment, the downside of bad data isn’t just a missed impression—it becomes a negative pattern embedded in the agent’s understanding of your brand.

What Products Up Is Building for Agentic Commerce

Katie outlined several specific areas where Products Up has focused its development in response to the agentic commerce shift.

The company was early to partner with agentic commerce platforms, helping to establish what readiness actually looks like from a product data standpoint. On the product side, they’ve built AI data services designed to expand a single SKU into multiple persona-based versions—for example, taking one unisex shoe and making it discoverable across several distinct contexts:

• A women’s running shoe
• A men’s basketball shoe
• A shoe suited for wet weather conditions
• A shoe appropriate for kids’ sports

This ensures that one product can surface across the full range of agent-driven queries without being artificially constrained by a single category label. They’re also building toward readiness for llm.txt, the emerging standard for signaling to LLM crawlers which pages and products a brand wants prioritized—a concept Katie compared directly to the site map of early SEO.

“We’re here to make sure those conversations are happening,” she said. “We talk about what the KPIs are because please don’t send all 300 million products. That’s just not going to be the way to do this.”

Getting AI Commerce Right Before the Window Closes

AI commerce readiness isn’t about moving fast or exposing everything you have. It’s about making deliberate decisions—about which products to expose, to which platforms, at what depth, and according to which KPIs—and then building the operational infrastructure to keep those decisions current as trends and demand shift.

Brands that treat AI platform deployment as a controlled experiment, starting small and expanding with data, will be in a far stronger position than those that rush to ship their full catalog and hope something sticks. The agents being trained on product data right now are building a picture of your brand that will influence recommendations long after the initial deployment. 

Ultimately, what you put in front of them matters—and so does what you leave out. As Katie put it: “The answers are always in the data.”

Voices of Search is a daily SEO and content marketing podcast hosted by Jordan Keone and Tyson Stockton. The show delivers actionable strategies and data-driven insights to help marketers navigate the ever-evolving world of search engine optimization and content marketing. New episodes air weekly, covering everything from technical SEO to AI discovery, featuring industry leaders and practitioners sharing real-world frameworks and proven tactics.

Subscribe to Voices of Search on Apple Podcasts, Spotify, or your favorite podcast platform. Follow Previsible on LinkedIn for updates and subscribe to the VOS YouTube channel for video episodes and clips. You can also visit the official VOS site to explore the full episode archive and submit your SEO questions for future episodes.